[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6NT6H04VGSBY9J5FS539F3E":3,"skill-same:01M6NT6H04VGSBY9J5FS539F3E":320},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":33,"publishedAt":54,"linkOnly":55,"markdownContent":56,"markdownBody":57,"frontMatter":58,"refName":90,"commitSha":91,"skillDirectory":92,"forksCount":13,"repositoryHtmlUrl":93,"repositoryHomepageUrl":6,"ownerHtmlUrl":94,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":95,"install":96,"resources":109},"01M6NT6H04VGSBY9J5FS539F3E","research-paper-writing",null,"End-to-end pipeline for writing ML\u002FAI research papers — from experiment design through analysis, drafting, revision, and submission. Covers NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Integrates automated experiment monitoring, statistical analysis, iterative writing, and citation verification.","qubit-calculus\u002FQubit","qubit-calculus","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F271071935?v=4","01M6NS19569VZ1F2Z0SACBWPS7","01M6NS1992MKFQVNP9AGRTSZXN",0,"MIT","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002FSKILL.md",[17,22,25,28],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000015","视觉设计","SCENARIO","使用场景",{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":26,"categoryName":27,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000021","资料检索与研究",{"categoryNo":29,"categoryName":30,"taxonomyType":31,"taxonomyTypeDesc":32},"01JBSEEDCAT000000000000036","内容创作者","PROFESSION","适用职业",[34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53],"writing","research","paper","Paper Writing","research paper writing","submission","design","for","citation verification","scipy","numpy","files","subagent","plan","NeurIPS","monitoring","experiment design","experiments","terminal","macos","2026-11-10 10:12:30",false,"---\nname: research-paper-writing\ntitle: Research Paper Writing Pipeline\ndescription: End-to-end pipeline for writing ML\u002FAI research papers — from experiment design through analysis, drafting, revision, and submission. Covers NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Integrates automated experiment monitoring, statistical analysis, iterative writing, and citation verification.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ndependencies: [semanticscholar, arxiv, habanero, requests, scipy, numpy, matplotlib, SciencePlots]\nplatforms: [linux, macos]\nmetadata:\n  hermes:\n    tags: [Research, Paper Writing, Experiments, ML, AI, NeurIPS, ICML, ICLR, ACL, AAAI, COLM, LaTeX, Citations, Statistical Analysis]\n    category: research\n    related_skills: [arxiv, ml-paper-writing, subagent-driven-development, plan]\n    requires_toolsets: [terminal, files]\n\n---\n\n# Research Paper Writing Pipeline\n\nEnd-to-end pipeline for producing publication-ready ML\u002FAI research papers targeting **NeurIPS, ICML, ICLR, ACL, AAAI, and COLM**. This skill covers the full research lifecycle: experiment design, execution, monitoring, analysis, paper writing, review, revision, and submission.\n\nThis is **not a linear pipeline** — it is an iterative loop. Results trigger new experiments. Reviews trigger new analysis. The agent must handle these feedback loops.\n\n```\n┌─────────────────────────────────────────────────────────────┐\n│                    RESEARCH PAPER PIPELINE                  │\n│                                                             │\n│  Phase 0: Project Setup ──► Phase 1: Literature Review      │\n│       │                          │                          │\n│       ▼                          ▼                          │\n│  Phase 2: Experiment     Phase 5: Paper Drafting ◄──┐      │\n│       Design                     │                   │      │\n│       │                          ▼                   │      │\n│       ▼                    Phase 6: Self-Review      │      │\n│  Phase 3: Execution &           & Revision ──────────┘      │\n│       Monitoring                 │                          │\n│       │                          ▼                          │\n│       ▼                    Phase 7: Submission               │\n│  Phase 4: Analysis ─────► (feeds back to Phase 2 or 5)     │\n│                                                             │\n└─────────────────────────────────────────────────────────────┘\n```\n\n---\n\n## When To Use This Skill\n\nUse this skill when:\n- **Starting a new research paper** from an existing codebase or idea\n- **Designing and running experiments** to support paper claims\n- **Writing or revising** any section of a research paper\n- **Preparing for submission** to a specific conference\n- **Responding to reviews** with additional experiments or revisions\n- **Converting** a paper between conference formats\n\n## Core Philosophy\n\n1. **Be proactive.** Deliver complete drafts, not questions. Scientists are busy — produce something concrete they can react to, then iterate.\n2. **Never hallucinate citations.** AI-generated citations have ~40% error rate. Always fetch programmatically. Mark unverifiable citations as `[CITATION NEEDED]`.\n3. **Paper is a story, not a collection of experiments.** Every paper needs one clear contribution stated in a single sentence. If you can't do that, the paper isn't ready.\n4. **Experiments serve claims.** Every experiment must explicitly state which claim it supports. Never run experiments that don't connect to the paper's narrative.\n5. **Commit early, commit often.** Every completed experiment batch, every paper draft update — commit with descriptive messages. Git log is the experiment history.\n\n### Proactivity and Collaboration\n\n**Default: Be proactive. Draft first, ask with the draft.**\n\n| Confidence Level | Action |\n|-----------------|--------|\n| **High** (clear repo, obvious contribution) | Write full draft, deliver, iterate on feedback |\n| **Medium** (some ambiguity) | Write draft with flagged uncertainties, continue |\n| **Low** (major unknowns) | Ask 1-2 targeted questions via `clarify`, then draft |\n\n| Section | Draft Autonomously? | Flag With Draft |\n|---------|-------------------|-----------------|\n| Abstract | Yes | \"Framed contribution as X — adjust if needed\" |\n| Introduction | Yes | \"Emphasized problem Y — correct if wrong\" |\n| Methods | Yes | \"Included details A, B, C — add missing pieces\" |\n| Experiments | Yes | \"Highlighted results 1, 2, 3 — reorder if needed\" |\n| Related Work | Yes | \"Cited papers X, Y, Z — add any I missed\" |\n\n**Block for input only when**: target venue unclear, multiple contradictory framings, results seem incomplete, explicit request to review first.\n\n---\n\n## Phase 0: Project Setup\n\n**Goal**: Establish the workspace, understand existing work, identify the contribution.\n\n### Step 0.1: Explore the Repository\n\n```bash\n# Understand project structure\nls -la\nfind . -name \"*.py\" | head -30\nfind . -name \"*.md\" -o -name \"*.txt\" | xargs grep -l -i \"result\\|conclusion\\|finding\"\n```\n\nLook for:\n- `README.md` — project overview and claims\n- `results\u002F`, `outputs\u002F`, `experiments\u002F` — existing findings\n- `configs\u002F` — experimental settings\n- `.bib` files — existing citations\n- Draft documents or notes\n\n### Step 0.2: Organize the Workspace\n\nEstablish a consistent workspace structure:\n\n```\nworkspace\u002F\n  paper\u002F               # LaTeX source, figures, compiled PDFs\n  experiments\u002F         # Experiment runner scripts\n  code\u002F                # Core method implementation\n  results\u002F             # Raw experiment results (auto-generated)\n  tasks\u002F               # Task\u002Fbenchmark definitions\n  human_eval\u002F          # Human evaluation materials (if needed)\n```\n\n### Step 0.3: Set Up Version Control\n\n```bash\ngit init  # if not already\ngit remote add origin \u003Crepo-url>\ngit checkout -b paper-draft  # or main\n```\n\n**Git discipline**: Every completed experiment batch gets committed with a descriptive message. Example:\n```\nAdd Monte Carlo constrained results (5 runs, Sonnet 4.6, policy memo task)\nAdd Haiku baseline comparison: autoreason vs refinement baselines at cheap model tier\n```\n\n### Step 0.4: Identify the Contribution\n\nBefore writing anything, articulate:\n- **The What**: What is the single thing this paper contributes?\n- **The Why**: What evidence supports it?\n- **The So What**: Why should readers care?\n\n> Propose to the scientist: \"Based on my understanding, the main contribution is: [one sentence]. The key results show [Y]. Is this the framing you want?\"\n\n### Step 0.5: Create a TODO List\n\nUse the `todo` tool to create a structured project plan:\n\n```\nResearch Paper TODO:\n- [ ] Define one-sentence contribution\n- [ ] Literature review (related work + baselines)\n- [ ] Design core experiments\n- [ ] Run experiments\n- [ ] Analyze results\n- [ ] Write first draft\n- [ ] Self-review (simulate reviewers)\n- [ ] Revise based on review\n- [ ] Submission prep\n```\n\nUpdate this throughout the project. It serves as the persistent state across sessions.\n\n---\n\n## Phase 1: Literature Review\n\n**Goal**: Find related work, identify baselines, gather citations.\n\n### Step 1.1: Identify Seed Papers\n\nStart from papers already referenced in the codebase:\n\n```bash\n# Via terminal:\ngrep -r \"arxiv\\|doi\\|cite\" --include=\"*.md\" --include=\"*.bib\" --include=\"*.py\"\nfind . -name \"*.bib\"\n```\n\n### Step 1.2: Search for Related Work\n\n**Load the `arxiv` skill** for structured paper discovery: `skill_view(\"arxiv\")`. It provides arXiv REST API search, Semantic Scholar citation graphs, author profiles, and BibTeX generation.\n\nUse `web_search` for broad discovery, `web_extract` for fetching specific papers:\n\n```\n# Via web_search:\nweb_search(\"[main technique] + [application domain] site:arxiv.org\")\nweb_search(\"[baseline method] comparison ICML NeurIPS 2024\")\n\n# Via web_extract (for specific papers):\nweb_extract(\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\")\n```\n\nAdditional search queries to try:\n\n```\nSearch queries:\n- \"[main technique] + [application domain]\"\n- \"[baseline method] comparison\"\n- \"[problem name] state-of-the-art\"\n- Author names from existing citations\n```\n\n**Recommended**: Install **Exa MCP** for real-time academic search:\n```bash\nclaude mcp add exa -- npx -y mcp-remote \"https:\u002F\u002Fmcp.exa.ai\u002Fmcp\"\n```\n\n### Step 1.3: Verify Every Citation\n\n**NEVER generate BibTeX from memory. ALWAYS fetch programmatically.**\n\nFor each citation, follow the mandatory 5-step process:\n\n```\nCitation Verification (MANDATORY per citation):\n1. SEARCH → Query Semantic Scholar or Exa MCP with specific keywords\n2. VERIFY → Confirm paper exists in 2+ sources (Semantic Scholar + arXiv\u002FCrossRef)\n3. RETRIEVE → Get BibTeX via DOI content negotiation (programmatically, not from memory)\n4. VALIDATE → Confirm the claim you're citing actually appears in the paper\n5. ADD → Add verified BibTeX to bibliography\nIf ANY step fails → mark as [CITATION NEEDED], inform scientist\n```\n\n```python\n# Fetch BibTeX via DOI\nimport requests\n\ndef doi_to_bibtex(doi: str) -> str:\n    response = requests.get(\n        f\"https:\u002F\u002Fdoi.org\u002F{doi}\",\n        headers={\"Accept\": \"application\u002Fx-bibtex\"}\n    )\n    response.raise_for_status()\n    return response.text\n```\n\nIf you cannot verify a citation:\n\n```latex\n\\cite{PLACEHOLDER_author2024_verify_this}  % TODO: Verify this citation exists\n```\n\n**Always tell the scientist**: \"I've marked [X] citations as placeholders that need verification.\"\n\nSee [references\u002Fcitation-workflow.md](references\u002Fcitation-workflow.md) for complete API documentation and the full `CitationManager` class.\n\n### Step 1.4: Organize Related Work\n\nGroup papers by methodology, not paper-by-paper:\n\n**Good**: \"One line of work uses X's assumption [refs] whereas we use Y's assumption because...\"\n**Bad**: \"Smith et al. introduced X. Jones et al. introduced Y. We combine both.\"\n\n---\n\n## Phase 2: Experiment Design\n\n**Goal**: Design experiments that directly support paper claims. Every experiment must answer a specific question.\n\n### Step 2.1: Map Claims to Experiments\n\nCreate an explicit mapping:\n\n| Claim | Experiment | Expected Evidence |\n|-------|-----------|-------------------|\n| \"Our method outperforms baselines\" | Main comparison (Table 1) | Win rate, statistical significance |\n| \"Effect is larger for weaker models\" | Model scaling study | Monotonic improvement curve |\n| \"Convergence requires scope constraints\" | Constrained vs unconstrained | Convergence rate comparison |\n\n**Rule**: If an experiment doesn't map to a claim, don't run it.\n\n### Step 2.2: Design Baselines\n\nStrong baselines are what separates accepted papers from rejected ones. Reviewers will ask: \"Did they compare against X?\"\n\nStandard baseline categories:\n- **Naive baseline**: Simplest possible approach\n- **Strong baseline**: Best known existing method\n- **Ablation baselines**: Your method minus one component\n- **Compute-matched baselines**: Same compute budget, different allocation\n\n### Step 2.3: Define Evaluation Protocol\n\nBefore running anything, specify:\n- **Metrics**: What you're measuring, direction symbols (higher\u002Flower better)\n- **Aggregation**: How results are combined across runs\u002Ftasks\n- **Statistical tests**: What tests will establish significance\n- **Sample sizes**: How many runs\u002Fproblems\u002Ftasks\n\n### Step 2.4: Write Experiment Scripts\n\nFollow these patterns from successful research pipelines:\n\n**Incremental saving** — save results after each step for crash recovery:\n```python\n# Save after each problem\u002Ftask\nresult_path = f\"results\u002F{task}\u002F{strategy}\u002Fresult.json\"\nif os.path.exists(result_path):\n    continue  # Skip already-completed work\n# ... run experiment ...\nwith open(result_path, 'w') as f:\n    json.dump(result, f, indent=2)\n```\n\n**Artifact preservation** — save all intermediate outputs:\n```\nresults\u002F\u003Cexperiment>\u002F\n  \u003Ctask>\u002F\n    \u003Cstrategy>\u002F\n      final_output.md          # Final result\n      history.json             # Full trajectory\n      pass_01\u002F                 # Per-iteration artifacts\n        version_a.md\n        version_b.md\n        critic.md\n```\n\n**Separation of concerns** — keep generation, evaluation, and visualization separate:\n```\nrun_experiment.py              # Core experiment runner\nrun_baselines.py               # Baseline comparison\nrun_comparison_judge.py        # Blind evaluation\nanalyze_results.py             # Statistical analysis\nmake_charts.py                 # Visualization\n```\n\nSee [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) for complete design patterns, cron monitoring, and error recovery.\n\n---\n\n## Phase 3: Experiment Execution & Monitoring\n\n**Goal**: Run experiments reliably, monitor progress, recover from failures.\n\n### Step 3.1: Launch Experiments\n\nUse `nohup` for long-running experiments:\n\n```bash\nnohup python run_experiment.py --config config.yaml > logs\u002Fexperiment_01.log 2>&1 &\necho $!  # Record the PID\n```\n\n**Parallel execution**: Run independent experiments simultaneously, but be aware of API rate limits. 4+ concurrent experiments on the same API will slow each down.\n\n### Step 3.2: Set Up Monitoring (Cron Pattern)\n\nFor long-running experiments, set up periodic status checks. The cron prompt should follow this template:\n\n```\nMonitor Prompt Template:\n1. Check if process is still running: ps aux | grep \u003Cpattern>\n2. Read last 30 lines of log: tail -30 \u003Clogfile>\n3. Check for completed results: ls \u003Cresult_dir>\n4. If results exist, read and report: cat \u003Cresult_file>\n5. If all done, commit: git add -A && git commit -m \"\u003Cdescriptive message>\" && git push\n6. Report in structured format (tables with key metrics)\n7. Answer the key analytical question for this experiment\n```\n\n**Silent mode**: If nothing has changed since the last check, respond with `[SILENT]` to suppress notification to the user. Only report when there's news.\n\n### Step 3.3: Handle Failures\n\nCommon failure modes and recovery:\n\n| Failure | Detection | Recovery |\n|---------|-----------|----------|\n| API rate limit \u002F credit exhaustion | 402\u002F429 errors in logs | Wait, then re-run (scripts skip completed work) |\n| Process crash | PID gone, incomplete results | Re-run from last checkpoint |\n| Timeout on hard problems | Process stuck, no log progress | Kill and skip, note in results |\n| Wrong model ID | Errors referencing model name | Fix ID and re-run |\n\n**Key**: Scripts should always check for existing results and skip completed work. This makes re-runs safe and efficient.\n\n### Step 3.4: Commit Completed Results\n\nAfter each experiment batch completes:\n\n```bash\ngit add -A\ngit commit -m \"Add \u003Cexperiment name>: \u003Ckey finding in 1 line>\"\ngit push\n```\n\n---\n\n## Phase 4: Result Analysis\n\n**Goal**: Extract findings, compute statistics, identify the story.\n\n### Step 4.1: Aggregate Results\n\nWrite analysis scripts that:\n1. Load all result files from a batch\n2. Compute per-task and aggregate metrics\n3. Generate summary tables\n\n```python\n# Standard analysis pattern\nimport json, os\nfrom pathlib import Path\n\nresults = {}\nfor result_file in Path(\"results\u002F\").rglob(\"result.json\"):\n    data = json.loads(result_file.read_text())\n    strategy = result_file.parent.name\n    task = result_file.parent.parent.name\n    results.setdefault(strategy, {})[task] = data\n\n# Compute aggregate metrics\nfor strategy, tasks in results.items():\n    scores = [t[\"score\"] for t in tasks.values()]\n    print(f\"{strategy}: mean={np.mean(scores):.1f}, std={np.std(scores):.1f}\")\n```\n\n### Step 4.2: Statistical Significance\n\nAlways compute:\n- **Error bars**: Standard deviation or standard error, specify which\n- **Confidence intervals**: 95% CI for key results\n- **Pairwise tests**: McNemar's test for comparing two methods\n- **Effect sizes**: Cohen's d or h for practical significance\n\nSee [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) for complete implementations of McNemar's test, bootstrapped CIs, and Cohen's h.\n\n### Step 4.3: Identify the Story\n\nAfter analysis, explicitly answer:\n1. **What is the main finding?** State it in one sentence.\n2. **What surprised you?** Unexpected results often make the best papers.\n3. **What failed?** Failed experiments can be the most informative. Honest reporting of failures strengthens the paper.\n4. **What follow-up experiments are needed?** Results often raise new questions.\n\n### Step 4.4: Create Figures and Tables\n\n**Figures**:\n- Use vector graphics (PDF) for all plots: `plt.savefig('fig.pdf')`\n- Colorblind-safe palettes (Okabe-Ito or Paul Tol)\n- Self-contained captions — reader should understand without main text\n- No title inside figure — the caption serves this function\n\n**Tables**:\n- Use `booktabs` LaTeX package\n- Bold best value per metric\n- Include direction symbols (higher\u002Flower better)\n- Consistent decimal precision\n\n```latex\n\\usepackage{booktabs}\n\\begin{tabular}{lcc}\n\\toprule\nMethod & Accuracy $\\uparrow$ & Latency $\\downarrow$ \\\\\n\\midrule\nBaseline & 85.2 & 45ms \\\\\n\\textbf{Ours} & \\textbf{92.1} & 38ms \\\\\n\\bottomrule\n\\end{tabular}\n```\n\n### Step 4.5: Decide: More Experiments or Write?\n\n| Situation | Action |\n|-----------|--------|\n| Core claims supported, results significant | Move to Phase 5 (writing) |\n| Results inconclusive, need more data | Back to Phase 2 (design) |\n| Unexpected finding suggests new direction | Back to Phase 2 (design) |\n| Missing one ablation reviewers will ask for | Run it, then Phase 5 |\n| All experiments done but some failed | Note failures, move to Phase 5 |\n\n---\n\n## Iterative Refinement: Strategy Selection\n\nAny output in this pipeline — paper drafts, experiment scripts, analysis — can be iteratively refined. The autoreason research provides empirical evidence for when each refinement strategy works and when it fails. Use this section to choose the right approach.\n\n### Quick Decision Table\n\n| Your Situation | Strategy | Why |\n|---------------|----------|-----|\n| Mid-tier model + constrained task | **Autoreason** | Sweet spot. Generation-evaluation gap is widest. Baselines actively destroy weak model outputs. |\n| Mid-tier model + open task | **Autoreason** with scope constraints added | Add fixed facts, structure, or deliverable to bound the improvement space. |\n| Frontier model + constrained task | **Autoreason** | Wins 2\u002F3 constrained tasks even at frontier. |\n| Frontier model + unconstrained task | **Critique-and-revise** or **single pass** | Autoreason comes last. Model self-evaluates well enough. |\n| Concrete technical task (system design) | **Critique-and-revise** | Direct find-and-fix loop is more efficient. |\n| Template-filling task (one correct structure) | **Single pass** or **conservative** | Minimal decision space. Iteration adds no value. |\n| Code with test cases | **Autoreason (code variant)** | Structured analysis of *why* it failed before fixing. Recovery rate 62% vs 43%. |\n| Very weak model (Llama 8B class) | **Single pass** | Model too weak for diverse candidates. Invest in generation quality. |\n\n### The Generation-Evaluation Gap\n\n**Core insight**: Autoreason's value depends on the gap between a model's generation capability and its self-evaluation capability.\n\n```\nModel Tier        │ Generation │ Self-Eval │ Gap    │ Autoreason Value\n──────────────────┼────────────┼───────────┼────────┼─────────────────\nWeak (Llama 8B)   │ Poor       │ Poor      │ Small  │ None — can't generate diverse candidates\nMid (Haiku 3.5)   │ Decent     │ Poor      │ LARGE  │ MAXIMUM — 42\u002F42 perfect Borda\nMid (Gemini Flash)│ Decent     │ Moderate  │ Large  │ High — wins 2\u002F3\nStrong (Sonnet 4) │ Good       │ Decent    │ Medium │ Moderate — wins 3\u002F5\nFrontier (S4.6)   │ Excellent  │ Good      │ Small  │ Only with constraints\n```\n\nThis gap is structural, not temporary. As costs drop, today's frontier becomes tomorrow's mid-tier. The sweet spot moves but never disappears.\n\n### Autoreason Loop (Summary)\n\nEach pass produces three candidates from fresh, isolated agents:\n\n1. **Critic** → finds problems in incumbent A (no fixes)\n2. **Author B** → revises A based on critique\n3. **Synthesizer** → merges A and B (randomized labels)\n4. **Judge Panel** → 3 blind CoT judges rank A, B, AB via Borda count\n5. **Convergence** → A wins k=2 consecutive passes → done\n\n**Key parameters:**\n- k=2 convergence (k=1 premature, k=3 too expensive, no quality gain)\n- CoT judges always (3x faster convergence)\n- Temperature 0.8 authors, 0.3 judges\n- Conservative tiebreak: incumbent wins ties\n- Every role is a fresh agent with no shared context\n\n### Applying to Paper Drafts\n\nWhen refining the paper itself through autoreason:\n- **Provide ground truth to the critic**: actual experimental data, result JSONs, statistical outputs. Without this, models hallucinate fabricated ablation studies and fake confidence intervals.\n- **Use 3 working judges minimum**: A broken judge parser doesn't add noise — it prevents equilibrium entirely.\n- **Scope constrain the revision**: \"Address these specific weaknesses\" not \"improve the paper.\"\n\n### Failure Modes\n\n| Failure | Detection | Fix |\n|---------|-----------|-----|\n| No convergence (A never wins) | A wins \u003C15% over 20+ passes | Add scope constraints to the task |\n| Synthesis drift | Word counts grow unboundedly | Constrain structure and deliverable |\n| Degradation below single pass | Baselines score higher than iterated output | Switch to single pass; model may be too weak |\n| Overfitting (code) | High public-test pass, low private-test pass | Use structured analysis, not just test feedback |\n| Broken judges | Parsing failures reduce panel below 3 | Fix parser before continuing |\n\nSee [references\u002Fautoreason-methodology.md](references\u002Fautoreason-methodology.md) for complete prompts, Borda scoring details, model selection guide, scope constraint design patterns, and compute budget reference.\n\n---\n\n## Phase 5: Paper Drafting\n\n**Goal**: Write a complete, publication-ready paper.\n\n### The Narrative Principle\n\n**The single most critical insight**: Your paper is not a collection of experiments — it's a story with one clear contribution supported by evidence.\n\nEvery successful ML paper centers on what Neel Nanda calls \"the narrative\": a short, rigorous, evidence-based technical story with a takeaway readers care about.\n\n**Three Pillars (must be crystal clear by end of introduction):**\n\n| Pillar | Description | Test |\n|--------|-------------|------|\n| **The What** | 1-3 specific novel claims | Can you state them in one sentence? |\n| **The Why** | Rigorous empirical evidence | Do experiments distinguish your hypothesis from alternatives? |\n| **The So What** | Why readers should care | Does this connect to a recognized community problem? |\n\n**If you cannot state your contribution in one sentence, you don't yet have a paper.**\n\n### Time Allocation\n\nSpend approximately **equal time** on each of:\n1. The abstract\n2. The introduction\n3. The figures\n4. Everything else combined\n\n**Why?** Most reviewers form judgments before reaching your methods. Readers encounter your paper as: title → abstract → introduction → figures → maybe the rest.\n\n### Writing Workflow\n\n```\nPaper Writing Checklist:\n- [ ] Step 1: Define the one-sentence contribution\n- [ ] Step 2: Draft Figure 1 (core idea or most compelling result)\n- [ ] Step 3: Draft abstract (5-sentence formula)\n- [ ] Step 4: Draft introduction (1-1.5 pages max)\n- [ ] Step 5: Draft methods\n- [ ] Step 6: Draft experiments & results\n- [ ] Step 7: Draft related work\n- [ ] Step 8: Draft conclusion & discussion\n- [ ] Step 9: Draft limitations (REQUIRED by all venues)\n- [ ] Step 10: Plan appendix (proofs, extra experiments, details)\n- [ ] Step 11: Complete paper checklist\n- [ ] Step 12: Final review\n```\n\n### Step 5.0: Title\n\nThe title is the single most-read element of the paper. It determines whether anyone clicks through to the abstract.\n\n**Good titles**:\n- State the contribution or finding: \"Autoreason: When Iterative LLM Refinement Works and Why It Fails\"\n- Highlight a surprising result: \"Scaling Data-Constrained Language Models\" (implies you can)\n- Name the method + what it does: \"DPO: Direct Preference Optimization of Language Models\"\n\n**Bad titles**:\n- Too generic: \"An Approach to Improving Language Model Outputs\"\n- Too long: anything over ~15 words\n- Jargon-only: \"Asymptotic Convergence of Iterative Stochastic Policy Refinement\" (who is this for?)\n\n**Rules**:\n- Include your method name if you have one (for citability)\n- Include 1-2 keywords reviewers will search for\n- Avoid colons unless both halves carry meaning\n- Test: would a reviewer know the domain and contribution from the title alone?\n\n### Step 5.1: Abstract (5-Sentence Formula)\n\nFrom Sebastian Farquhar (DeepMind):\n\n```\n1. What you achieved: \"We introduce...\", \"We prove...\", \"We demonstrate...\"\n2. Why this is hard and important\n3. How you do it (with specialist keywords for discoverability)\n4. What evidence you have\n5. Your most remarkable number\u002Fresult\n```\n\n**Delete** generic openings like \"Large language models have achieved remarkable success...\"\n\n### Step 5.2: Figure 1\n\nFigure 1 is the second thing most readers look at (after abstract). Draft it before writing the introduction — it forces you to clarify the core idea.\n\n| Figure 1 Type | When to Use | Example |\n|---------------|-------------|---------|\n| **Method diagram** | New architecture or pipeline | TikZ flowchart showing your system |\n| **Results teaser** | One compelling result tells the whole story | Bar chart: \"Ours vs baselines\" with clear gap |\n| **Problem illustration** | The problem is unintuitive | Before\u002Fafter showing failure mode you fix |\n| **Conceptual diagram** | Abstract contribution needs visual grounding | 2x2 matrix of method properties |\n\n**Rules**: Figure 1 must be understandable without reading any text. The caption alone should communicate the core idea. Use color purposefully — don't just decorate.\n\n### Step 5.3: Introduction (1-1.5 pages max)\n\nMust include:\n- Clear problem statement\n- Brief approach overview\n- 2-4 bullet contribution list (max 1-2 lines each in two-column format)\n- Methods should start by page 2-3\n\n### Step 5.3: Methods\n\nEnable reimplementation:\n- Conceptual outline or pseudocode\n- All hyperparameters listed\n- Architectural details sufficient for reproduction\n- Present final design decisions; ablations go in experiments\n\n### Step 5.4: Experiments & Results\n\nFor each experiment, explicitly state:\n- **What claim it supports**\n- How it connects to main contribution\n- What to observe: \"the blue line shows X, which demonstrates Y\"\n\nRequirements:\n- Error bars with methodology (std dev vs std error)\n- Hyperparameter search ranges\n- Compute infrastructure (GPU type, total hours)\n- Seed-setting methods\n\n### Step 5.5: Related Work\n\nOrganize methodologically, not paper-by-paper. Cite generously — reviewers likely authored relevant papers.\n\n### Step 5.6: Limitations (REQUIRED)\n\nAll major conferences require this. Honesty helps:\n- Reviewers are instructed not to penalize honest limitation acknowledgment\n- Pre-empt criticisms by identifying weaknesses first\n- Explain why limitations don't undermine core claims\n\n### Step 5.7: Conclusion & Discussion\n\n**Conclusion** (required, 0.5-1 page):\n- Restate the contribution in one sentence (different wording from abstract)\n- Summarize key findings (2-3 sentences, not a list)\n- Implications: what does this mean for the field?\n- Future work: 2-3 concrete next steps (not vague \"we leave X for future work\")\n\n**Discussion** (optional, sometimes combined with conclusion):\n- Broader implications beyond immediate results\n- Connections to other subfields\n- Honest assessment of when the method does and doesn't work\n- Practical deployment considerations\n\n**Do NOT** introduce new results or claims in the conclusion.\n\n### Step 5.8: Appendix Strategy\n\nAppendices are unlimited at all major venues and are essential for reproducibility. Structure:\n\n| Appendix Section | What Goes Here |\n|-----------------|---------------|\n| **Proofs & Derivations** | Full proofs too long for main text. Main text can state theorems with \"proof in Appendix A.\" |\n| **Additional Experiments** | Ablations, scaling curves, per-dataset breakdowns, hyperparameter sensitivity |\n| **Implementation Details** | Full hyperparameter tables, training details, hardware specs, random seeds |\n| **Dataset Documentation** | Data collection process, annotation guidelines, licensing, preprocessing |\n| **Prompts & Templates** | Exact prompts used (for LLM-based methods), evaluation templates |\n| **Human Evaluation** | Annotation interface screenshots, instructions given to annotators, IRB details |\n| **Additional Figures** | Per-task breakdowns, trajectory visualizations, failure case examples |\n\n**Rules**:\n- The main paper must be self-contained — reviewers are not required to read appendices\n- Never put critical evidence only in the appendix\n- Cross-reference: \"Full results in Table 5 (Appendix B)\" not just \"see appendix\"\n- Use `\\appendix` command, then `\\section{A: Proofs}` etc.\n\n### Page Budget Management\n\nWhen over the page limit:\n\n| Cut Strategy | Saves | Risk |\n|-------------|-------|------|\n| Move proofs to appendix | 0.5-2 pages | Low — standard practice |\n| Condense related work | 0.5-1 page | Medium — may miss key citations |\n| Combine tables with subfigures | 0.25-0.5 page | Low — often improves readability |\n| Use `\\vspace{-Xpt}` sparingly | 0.1-0.3 page | Low if subtle, high if obvious |\n| Remove qualitative examples | 0.5-1 page | Medium — reviewers like examples |\n| Reduce figure sizes | 0.25-0.5 page | High — figures must remain readable |\n\n**Do NOT**: reduce font size, change margins, remove required sections (limitations, broader impact), or use `\\small`\u002F`\\footnotesize` for main text.\n\n### Writing Style\n\n**Sentence-level clarity (Gopen & Swan's 7 Principles):**\n\n| Principle | Rule |\n|-----------|------|\n| Subject-verb proximity | Keep subject and verb close |\n| Stress position | Place emphasis at sentence ends |\n| Topic position | Put context first, new info after |\n| Old before new | Familiar info → unfamiliar info |\n| One unit, one function | Each paragraph makes one point |\n| Action in verb | Use verbs, not nominalizations |\n| Context before new | Set stage before presenting |\n\n**Word choice (Lipton, Steinhardt):**\n- Be specific: \"accuracy\" not \"performance\"\n- Eliminate hedging: drop \"may\" unless genuinely uncertain\n- Consistent terminology throughout\n- Avoid incremental vocabulary: \"develop\", not \"combine\"\n\n**Full writing guide with examples**: See [references\u002Fwriting-guide.md](references\u002Fwriting-guide.md)\n\n### Using LaTeX Templates\n\n**Always copy the entire template directory first, then write within it.**\n\n```\nTemplate Setup Checklist:\n- [ ] Step 1: Copy entire template directory to new project\n- [ ] Step 2: Verify template compiles as-is (before any changes)\n- [ ] Step 3: Read the template's example content to understand structure\n- [ ] Step 4: Replace example content section by section\n- [ ] Step 5: Use template macros (check preamble for \\newcommand definitions)\n- [ ] Step 6: Clean up template artifacts only at the end\n```\n\n**Step 1: Copy the Full Template**\n\n```bash\ncp -r templates\u002Fneurips2025\u002F ~\u002Fpapers\u002Fmy-paper\u002F\ncd ~\u002Fpapers\u002Fmy-paper\u002F\nls -la  # Should see: main.tex, neurips.sty, Makefile, etc.\n```\n\nCopy the ENTIRE directory, not just the .tex file. Templates include style files (.sty), bibliography styles (.bst), example content, and Makefiles.\n\n**Step 2: Verify Template Compiles First**\n\nBefore making ANY changes:\n```bash\nlatexmk -pdf main.tex\n# Or manual: pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex\n```\n\nIf the unmodified template doesn't compile, fix that first (usually missing TeX packages — install via `tlmgr install \u003Cpackage>`).\n\n**Step 3: Keep Template Content as Reference**\n\nDon't immediately delete example content. Comment it out and use as formatting reference:\n```latex\n% Template example (keep for reference):\n% \\begin{figure}[t]\n%   \\centering\n%   \\includegraphics[width=0.8\\linewidth]{example-image}\n%   \\caption{Template shows caption style}\n% \\end{figure}\n\n% Your actual figure:\n\\begin{figure}[t]\n  \\centering\n  \\includegraphics[width=0.8\\linewidth]{your-figure.pdf}\n  \\caption{Your caption following the same style.}\n\\end{figure}\n```\n\n**Step 4: Replace Content Section by Section**\n\nWork through systematically: title\u002Fauthors → abstract → introduction → methods → experiments → related work → conclusion → references → appendix. Compile after each section.\n\n**Step 5: Use Template Macros**\n\n```latex\n\\newcommand{\\method}{YourMethodName}  % Consistent method naming\n\\newcommand{\\eg}{e.g.,\\xspace}        % Proper abbreviations\n\\newcommand{\\ie}{i.e.,\\xspace}\n```\n\n### Template Pitfalls\n\n| Pitfall | Problem | Solution |\n|---------|---------|----------|\n| Copying only `.tex` file | Missing `.sty`, won't compile | Copy entire directory |\n| Modifying `.sty` files | Breaks conference formatting | Never edit style files |\n| Adding random packages | Conflicts, breaks template | Only add if necessary |\n| Deleting template content early | Lose formatting reference | Keep as comments until done |\n| Not compiling frequently | Errors accumulate | Compile after each section |\n| Raster PNGs for figures | Blurry in paper | Always use vector PDF via `savefig('fig.pdf')` |\n\n### Quick Template Reference\n\n| Conference | Main File | Style File | Page Limit |\n|------------|-----------|------------|------------|\n| NeurIPS 2025 | `main.tex` | `neurips.sty` | 9 pages |\n| ICML 2026 | `example_paper.tex` | `icml2026.sty` | 8 pages |\n| ICLR 2026 | `iclr2026_conference.tex` | `iclr2026_conference.sty` | 9 pages |\n| ACL 2025 | `acl_latex.tex` | `acl.sty` | 8 pages (long) |\n| AAAI 2026 | `aaai2026-unified-template.tex` | `aaai2026.sty` | 7 pages |\n| COLM 2025 | `colm2025_conference.tex` | `colm2025_conference.sty` | 9 pages |\n\n**Universal**: Double-blind, references don't count, appendices unlimited, LaTeX required.\n\nTemplates in `templates\u002F` directory. See [templates\u002FREADME.md](templates\u002FREADME.md) for compilation setup (VS Code, CLI, Overleaf, other IDEs).\n\n### Tables and Figures\n\n**Tables** — use `booktabs` for professional formatting:\n\n```latex\n\\usepackage{booktabs}\n\\begin{tabular}{lcc}\n\\toprule\nMethod & Accuracy $\\uparrow$ & Latency $\\downarrow$ \\\\\n\\midrule\nBaseline & 85.2 & 45ms \\\\\n\\textbf{Ours} & \\textbf{92.1} & 38ms \\\\\n\\bottomrule\n\\end{tabular}\n```\n\nRules:\n- Bold best value per metric\n- Include direction symbols ($\\uparrow$ higher better, $\\downarrow$ lower better)\n- Right-align numerical columns\n- Consistent decimal precision\n\n**Figures**:\n- **Vector graphics** (PDF, EPS) for all plots and diagrams — `plt.savefig('fig.pdf')`\n- **Raster** (PNG 600 DPI) only for photographs\n- **Colorblind-safe palettes** (Okabe-Ito or Paul Tol)\n- Verify **grayscale readability** (8% of men have color vision deficiency)\n- **No title inside figure** — the caption serves this function\n- **Self-contained captions** — reader should understand without main text\n\n### Conference Resubmission\n\nFor converting between venues, see Phase 7 (Submission Preparation) — it covers the full conversion workflow, page-change table, and post-rejection guidance.\n\n### Professional LaTeX Preamble\n\nAdd these packages to any paper for professional quality. They are compatible with all major conference style files:\n\n```latex\n% --- Professional Packages (add after conference style file) ---\n\n% Typography\n\\usepackage{microtype}              % Microtypographic improvements (protrusion, expansion)\n                                     % Makes text noticeably more polished — always include\n\n% Tables\n\\usepackage{booktabs}               % Professional table rules (\\toprule, \\midrule, \\bottomrule)\n\\usepackage{siunitx}                % Consistent number formatting, decimal alignment\n                                     % Usage: \\num{12345} → 12,345; \\SI{3.5}{GHz} → 3.5 GHz\n                                     % Table alignment: S column type for decimal-aligned numbers\n\n% Figures\n\\usepackage{graphicx}               % Include graphics (\\includegraphics)\n\\usepackage{subcaption}             % Subfigures with (a), (b), (c) labels\n                                     % Usage: \\begin{subfigure}{0.48\\textwidth} ... \\end{subfigure}\n\n% Diagrams and Algorithms\n\\usepackage{tikz}                   % Programmable vector diagrams\n\\usetikzlibrary{arrows.meta, positioning, shapes.geometric, calc, fit, backgrounds}\n\\usepackage[ruled,vlined]{algorithm2e}  % Professional pseudocode\n                                     % Alternative: \\usepackage{algorithmicx} if template bundles it\n\n% Cross-references\n\\usepackage{cleveref}               % Smart references: \\cref{fig:x} → \"Figure 1\"\n                                     % MUST be loaded AFTER hyperref\n                                     % Handles: figures, tables, sections, equations, algorithms\n\n% Math (usually included by conference .sty, but verify)\n\\usepackage{amsmath,amssymb}        % AMS math environments and symbols\n\\usepackage{mathtools}              % Extends amsmath (dcases, coloneqq, etc.)\n\n% Colors (for figures and diagrams)\n\\usepackage{xcolor}                 % Color management\n% Okabe-Ito colorblind-safe palette:\n\\definecolor{okblue}{HTML}{0072B2}\n\\definecolor{okorange}{HTML}{E69F00}\n\\definecolor{okgreen}{HTML}{009E73}\n\\definecolor{okred}{HTML}{D55E00}\n\\definecolor{okpurple}{HTML}{CC79A7}\n\\definecolor{okcyan}{HTML}{56B4E9}\n\\definecolor{okyellow}{HTML}{F0E442}\n```\n\n**Notes:**\n- `microtype` is the single highest-impact package for visual quality. It adjusts character spacing at a sub-pixel level. Always include it.\n- `siunitx` handles decimal alignment in tables via the `S` column type — eliminates manual spacing.\n- `cleveref` must be loaded **after** `hyperref`. Most conference .sty files load hyperref, so put cleveref last.\n- Check if the conference template already loads any of these (especially `algorithm`, `amsmath`, `graphicx`). Don't double-load.\n\n### siunitx Table Alignment\n\n`siunitx` makes number-heavy tables significantly more readable:\n\n```latex\n\\begin{tabular}{l S[table-format=2.1] S[table-format=2.1] S[table-format=2.1]}\n\\toprule\nMethod & {Accuracy $\\uparrow$} & {F1 $\\uparrow$} & {Latency (ms) $\\downarrow$} \\\\\n\\midrule\nBaseline         & 85.2  & 83.7  & 45.3 \\\\\nAblation (no X)  & 87.1  & 85.4  & 42.1 \\\\\n\\textbf{Ours}    & \\textbf{92.1} & \\textbf{90.8} & \\textbf{38.7} \\\\\n\\bottomrule\n\\end{tabular}\n```\n\nThe `S` column type auto-aligns on the decimal point. Headers in `{}` escape the alignment.\n\n### Subfigures\n\nStandard pattern for side-by-side figures:\n\n```latex\n\\begin{figure}[t]\n  \\centering\n  \\begin{subfigure}[b]{0.48\\textwidth}\n    \\centering\n    \\includegraphics[width=\\textwidth]{fig_results_a.pdf}\n    \\caption{Results on Dataset A.}\n    \\label{fig:results-a}\n  \\end{subfigure}\n  \\hfill\n  \\begin{subfigure}[b]{0.48\\textwidth}\n    \\centering\n    \\includegraphics[width=\\textwidth]{fig_results_b.pdf}\n    \\caption{Results on Dataset B.}\n    \\label{fig:results-b}\n  \\end{subfigure}\n  \\caption{Comparison of our method across two datasets. (a) shows the scaling\n  behavior and (b) shows the ablation results. Both use 5 random seeds.}\n  \\label{fig:results}\n\\end{figure}\n```\n\nUse `\\cref{fig:results}` → \"Figure 1\", `\\cref{fig:results-a}` → \"Figure 1a\".\n\n### Pseudocode with algorithm2e\n\n```latex\n\\begin{algorithm}[t]\n\\caption{Iterative Refinement with Judge Panel}\n\\label{alg:method}\n\\KwIn{Task $T$, model $M$, judges $J_1 \\ldots J_n$, convergence threshold $k$}\n\\KwOut{Final output $A^*$}\n$A \\gets M(T)$ \\tcp*{Initial generation}\n$\\text{streak} \\gets 0$\\;\n\\While{$\\text{streak} \u003C k$}{\n  $C \\gets \\text{Critic}(A, T)$ \\tcp*{Identify weaknesses}\n  $B \\gets M(T, C)$ \\tcp*{Revised version addressing critique}\n  $AB \\gets \\text{Synthesize}(A, B)$ \\tcp*{Merge best elements}\n  \\ForEach{judge $J_i$}{\n    $\\text{rank}_i \\gets J_i(\\text{shuffle}(A, B, AB))$ \\tcp*{Blind ranking}\n  }\n  $\\text{winner} \\gets \\text{BordaCount}(\\text{ranks})$\\;\n  \\eIf{$\\text{winner} = A$}{\n    $\\text{streak} \\gets \\text{streak} + 1$\\;\n  }{\n    $A \\gets \\text{winner}$; $\\text{streak} \\gets 0$\\;\n  }\n}\n\\Return{$A$}\\;\n\\end{algorithm}\n```\n\n### TikZ Diagram Patterns\n\nTikZ is the standard for method diagrams in ML papers. Common patterns:\n\n**Pipeline\u002FFlow Diagram** (most common in ML papers):\n\n```latex\n\\begin{figure}[t]\n\\centering\n\\begin{tikzpicture}[\n  node distance=1.8cm,\n  box\u002F.style={rectangle, draw, rounded corners, minimum height=1cm, \n              minimum width=2cm, align=center, font=\\small},\n  arrow\u002F.style={-{Stealth[length=3mm]}, thick},\n]\n  \\node[box, fill=okcyan!20] (input) {Input\\\\$x$};\n  \\node[box, fill=okblue!20, right of=input] (encoder) {Encoder\\\\$f_\\theta$};\n  \\node[box, fill=okgreen!20, right of=encoder] (latent) {Latent\\\\$z$};\n  \\node[box, fill=okorange!20, right of=latent] (decoder) {Decoder\\\\$g_\\phi$};\n  \\node[box, fill=okred!20, right of=decoder] (output) {Output\\\\$\\hat{x}$};\n  \n  \\draw[arrow] (input) -- (encoder);\n  \\draw[arrow] (encoder) -- (latent);\n  \\draw[arrow] (latent) -- (decoder);\n  \\draw[arrow] (decoder) -- (output);\n\\end{tikzpicture}\n\\caption{Architecture overview. The encoder maps input $x$ to latent \nrepresentation $z$, which the decoder reconstructs.}\n\\label{fig:architecture}\n\\end{figure}\n```\n\n**Comparison\u002FMatrix Diagram** (for showing method variants):\n\n```latex\n\\begin{tikzpicture}[\n  cell\u002F.style={rectangle, draw, minimum width=2.5cm, minimum height=1cm, \n               align=center, font=\\small},\n  header\u002F.style={cell, fill=gray!20, font=\\small\\bfseries},\n]\n  % Headers\n  \\node[header] at (0, 0) {Method};\n  \\node[header] at (3, 0) {Converges?};\n  \\node[header] at (6, 0) {Quality?};\n  % Rows\n  \\node[cell] at (0, -1) {Single Pass};\n  \\node[cell, fill=okgreen!15] at (3, -1) {N\u002FA};\n  \\node[cell, fill=okorange!15] at (6, -1) {Baseline};\n  \\node[cell] at (0, -2) {Critique+Revise};\n  \\node[cell, fill=okred!15] at (3, -2) {No};\n  \\node[cell, fill=okred!15] at (6, -2) {Degrades};\n  \\node[cell] at (0, -3) {Ours};\n  \\node[cell, fill=okgreen!15] at (3, -3) {Yes ($k$=2)};\n  \\node[cell, fill=okgreen!15] at (6, -3) {Improves};\n\\end{tikzpicture}\n```\n\n**Iterative Loop Diagram** (for methods with feedback):\n\n```latex\n\\begin{tikzpicture}[\n  node distance=2cm,\n  box\u002F.style={rectangle, draw, rounded corners, minimum height=0.8cm, \n              minimum width=1.8cm, align=center, font=\\small},\n  arrow\u002F.style={-{Stealth[length=3mm]}, thick},\n  label\u002F.style={font=\\scriptsize, midway, above},\n]\n  \\node[box, fill=okblue!20] (gen) {Generator};\n  \\node[box, fill=okred!20, right=2.5cm of gen] (critic) {Critic};\n  \\node[box, fill=okgreen!20, below=1.5cm of $(gen)!0.5!(critic)$] (judge) {Judge Panel};\n  \n  \\draw[arrow] (gen) -- node[label] {output $A$} (critic);\n  \\draw[arrow] (critic) -- node[label, right] {critique $C$} (judge);\n  \\draw[arrow] (judge) -| node[label, left, pos=0.3] {winner} (gen);\n\\end{tikzpicture}\n```\n\n### latexdiff for Revision Tracking\n\nEssential for rebuttals — generates a marked-up PDF showing changes between versions:\n\n```bash\n# Install\n# macOS: brew install latexdiff (or comes with TeX Live)\n# Linux: sudo apt install latexdiff\n\n# Generate diff\nlatexdiff paper_v1.tex paper_v2.tex > paper_diff.tex\npdflatex paper_diff.tex\n\n# For multi-file projects (with \\input{} or \\include{})\nlatexdiff --flatten paper_v1.tex paper_v2.tex > paper_diff.tex\n```\n\nThis produces a PDF with deletions in red strikethrough and additions in blue — standard format for rebuttal supplements.\n\n### SciencePlots for matplotlib\n\nInstall and use for publication-quality plots:\n\n```bash\npip install SciencePlots\n```\n\n```python\nimport matplotlib.pyplot as plt\nimport scienceplots  # registers styles\n\n# Use science style (IEEE-like, clean)\nwith plt.style.context(['science', 'no-latex']):\n    fig, ax = plt.subplots(figsize=(3.5, 2.5))  # Single-column width\n    ax.plot(x, y, label='Ours', color='#0072B2')\n    ax.plot(x, y2, label='Baseline', color='#D55E00', linestyle='--')\n    ax.set_xlabel('Training Steps')\n    ax.set_ylabel('Accuracy')\n    ax.legend()\n    fig.savefig('paper\u002Ffig_results.pdf', bbox_inches='tight')\n\n# Available styles: 'science', 'ieee', 'nature', 'science+ieee'\n# Add 'no-latex' if LaTeX is not installed on the machine generating plots\n```\n\n**Standard figure sizes** (two-column format):\n- Single column: `figsize=(3.5, 2.5)` — fits in one column\n- Double column: `figsize=(7.0, 3.0)` — spans both columns\n- Square: `figsize=(3.5, 3.5)` — for heatmaps, confusion matrices\n\n---\n\n## Phase 6: Self-Review & Revision\n\n**Goal**: Simulate the review process before submission. Catch weaknesses early.\n\n### Step 6.1: Simulate Reviews\n\nGenerate reviews from multiple perspectives using strong models (Opus 4, Sonnet 4.6, Gemini 2.5 Pro). Use the reviewer guidelines from the target venue.\n\n**Review prompt template:**\n\n```\nYou are an expert reviewer for [VENUE]. Review this paper according to the \nofficial reviewer guidelines. Evaluate:\n\n1. Quality (technical soundness, baselines, claims supported by evidence)\n2. Clarity (writing, notation consistency, reproducibility)\n3. Significance (impact, importance of the problem)\n4. Originality (novelty, new insights)\n\nProvide:\n- Summary (2-3 sentences)\n- Strengths (bullet list)\n- Weaknesses (bullet list, most critical first)\n- Questions for authors\n- Missing references\n- Score (1-6 on NeurIPS scale)\n- Confidence (1-5)\n```\n\n### Step 6.2: Prioritize Feedback\n\nAfter collecting reviews, categorize:\n\n| Priority | Action |\n|----------|--------|\n| **Critical** (technical flaw, missing baseline) | Must fix. May require new experiments → back to Phase 2 |\n| **High** (clarity issue, missing ablation) | Should fix in this revision |\n| **Medium** (minor writing issues, extra experiments) | Fix if time allows |\n| **Low** (style preferences, tangential suggestions) | Note for future work |\n\n### Step 6.3: Revision Cycle\n\nFor each critical\u002Fhigh issue:\n1. Identify the specific section(s) affected\n2. Draft the fix\n3. Verify the fix doesn't break other claims\n4. Update the paper\n5. Re-check against the reviewer's concern\n\n### Step 6.4: Rebuttal Writing\n\nWhen responding to actual reviews (post-submission), rebuttals are a distinct skill from revision:\n\n**Format**: Point-by-point. For each reviewer concern:\n```\n> R1-W1: \"The paper lacks comparison with Method X.\"\n\nWe thank the reviewer for this suggestion. We have added a comparison with \nMethod X in Table 3 (revised). Our method outperforms X by 3.2pp on [metric] \n(p\u003C0.05). We note that X requires 2x our compute budget.\n```\n\n**Rules**:\n- Address every concern — reviewers notice if you skip one\n- Lead with the strongest responses\n- Be concise and direct — reviewers read dozens of rebuttals\n- Include new results if you ran experiments during the rebuttal period\n- Never be defensive or dismissive, even of weak criticisms\n- Use `latexdiff` to generate a marked-up PDF showing changes (see Professional LaTeX Tooling section)\n- Thank reviewers for specific, actionable feedback (not generic praise)\n\n**What NOT to do**: \"We respectfully disagree\" without evidence. \"This is out of scope\" without explanation. Ignoring a weakness by only responding to strengths.\n\n### Step 6.5: Paper Evolution Tracking\n\nSave snapshots at key milestones:\n```\npaper\u002F\n  paper.tex                    # Current working version\n  paper_v1_first_draft.tex     # First complete draft\n  paper_v2_post_review.tex     # After simulated review\n  paper_v3_pre_submission.tex  # Final before submission\n  paper_v4_camera_ready.tex    # Post-acceptance final\n```\n\n---\n\n## Phase 7: Submission Preparation\n\n**Goal**: Final checks, formatting, and submission.\n\n### Step 7.1: Conference Checklist\n\nEvery venue has mandatory checklists. Complete them carefully — incomplete checklists can result in desk rejection.\n\nSee [references\u002Fchecklists.md](references\u002Fchecklists.md) for:\n- NeurIPS 16-item paper checklist\n- ICML broader impact + reproducibility\n- ICLR LLM disclosure policy\n- ACL mandatory limitations section\n- Universal pre-submission checklist\n\n### Step 7.2: Anonymization Checklist\n\nDouble-blind review means reviewers cannot know who wrote the paper. Check ALL of these:\n\n```\nAnonymization Checklist:\n- [ ] No author names or affiliations anywhere in the PDF\n- [ ] No acknowledgments section (add after acceptance)\n- [ ] Self-citations written in third person: \"Smith et al. [1] showed...\" not \"We previously showed [1]...\"\n- [ ] No GitHub\u002FGitLab URLs pointing to your personal repos\n- [ ] Use Anonymous GitHub (https:\u002F\u002Fanonymous.4open.science\u002F) for code links\n- [ ] No institutional logos or identifiers in figures\n- [ ] No file metadata containing author names (check PDF properties)\n- [ ] No \"our previous work\" or \"in our earlier paper\" phrasing\n- [ ] Dataset names don't reveal institution (rename if needed)\n- [ ] Supplementary materials don't contain identifying information\n```\n\n**Common mistakes**: Git commit messages visible in supplementary code, watermarked figures from institutional tools, acknowledgments left in from a previous draft, arXiv preprint posted before anonymity period.\n\n### Step 7.3: Formatting Verification\n\n```\nPre-Submission Format Check:\n- [ ] Page limit respected (excluding references and appendix)\n- [ ] All figures are vector (PDF) or high-res raster (600 DPI PNG)\n- [ ] All figures readable in grayscale\n- [ ] All tables use booktabs\n- [ ] References compile correctly (no \"?\" in citations)\n- [ ] No overfull hboxes in critical areas\n- [ ] Appendix clearly labeled and separated\n- [ ] Required sections present (limitations, broader impact, etc.)\n```\n\n### Step 7.3: Final Compilation\n\n```bash\n# Clean build\nrm -f *.aux *.bbl *.blg *.log *.out *.pdf\nlatexmk -pdf main.tex\n\n# Or manual\npdflatex main.tex\nbibtex main\npdflatex main.tex\npdflatex main.tex\n```\n\n### Step 7.4: Conference-Specific Requirements\n\n| Venue | Special Requirements |\n|-------|---------------------|\n| **NeurIPS** | Paper checklist in appendix, lay summary if accepted |\n| **ICML** | Broader Impact Statement (after conclusion, doesn't count toward limit) |\n| **ICLR** | LLM disclosure required, reciprocal reviewing agreement |\n| **ACL** | Mandatory Limitations section, Responsible NLP checklist |\n| **AAAI** | Strict style file — no modifications whatsoever |\n| **COLM** | Frame contribution for language model community |\n\n### Step 7.6: Conference Resubmission & Format Conversion\n\nWhen converting between venues, **never copy LaTeX preambles between templates**:\n\n```bash\n# 1. Start fresh with target template\ncp -r templates\u002Ficml2026\u002F new_submission\u002F\n\n# 2. Copy ONLY content sections (not preamble)\n#    - Abstract text, section content, figures, tables, bib entries\n\n# 3. Adjust for page limits\n# 4. Add venue-specific required sections\n# 5. Update references\n```\n\n| From → To | Page Change | Key Adjustments |\n|-----------|-------------|-----------------|\n| NeurIPS → ICML | 9 → 8 | Cut 1 page, add Broader Impact |\n| ICML → ICLR | 8 → 9 | Expand experiments, add LLM disclosure |\n| NeurIPS → ACL | 9 → 8 | Restructure for NLP conventions, add Limitations |\n| ICLR → AAAI | 9 → 7 | Significant cuts, strict style adherence |\n| Any → COLM | varies → 9 | Reframe for language model focus |\n\nWhen cutting pages: move proofs to appendix, condense related work, combine tables, use subfigures.\nWhen expanding: add ablations, expand limitations, include additional baselines, add qualitative examples.\n\n**After rejection**: Address reviewer concerns in the new version, but don't include a \"changes\" section or reference the previous submission (blind review).\n\n### Step 7.7: Camera-Ready Preparation (Post-Acceptance)\n\nAfter acceptance, prepare the camera-ready version:\n\n```\nCamera-Ready Checklist:\n- [ ] De-anonymize: add author names, affiliations, email addresses\n- [ ] Add Acknowledgments section (funding, compute grants, helpful reviewers)\n- [ ] Add public code\u002Fdata URL (real GitHub, not anonymous)\n- [ ] Address any mandatory revisions from meta-reviewer\n- [ ] Switch template to camera-ready mode (if applicable — e.g., AAAI \\anon → \\camera)\n- [ ] Add copyright notice if required by venue\n- [ ] Update any \"anonymous\" placeholders in text\n- [ ] Verify final PDF compiles cleanly\n- [ ] Check page limit for camera-ready (sometimes differs from submission)\n- [ ] Upload supplementary materials (code, data, appendix) to venue portal\n```\n\n---\n\n## Hermes Agent Integration\n\nThis skill is designed for the Hermes agent. It uses Hermes tools, delegation, scheduling, and memory for the full research lifecycle.\n\n### Related Skills\n\nCompose this skill with other Hermes skills for specific phases:\n\n| Skill | When to Use | How to Load |\n|-------|-------------|-------------|\n| **arxiv** | Phase 1 (Literature Review): searching arXiv, generating BibTeX, finding related papers via Semantic Scholar | `skill_view(\"arxiv\")` |\n| **subagent-driven-development** | Phase 5 (Drafting): parallel section writing with 2-stage review (spec compliance then quality) | `skill_view(\"subagent-driven-development\")` |\n| **plan** | Phase 0 (Setup): creating structured plans before execution. Writes to `.hermes\u002Fplans\u002F` | `skill_view(\"plan\")` |\n| **qmd** | Phase 1 (Literature): searching local knowledge bases (notes, transcripts, docs) via hybrid BM25+vector search | Install: `skill_manage(\"install\", \"qmd\")` |\n| **diagramming** | Phase 4-5: creating Excalidraw-based figures and architecture diagrams | `skill_view(\"diagramming\")` |\n| **data-science** | Phase 4 (Analysis): Jupyter live kernel for interactive analysis and visualization | `skill_view(\"data-science\")` |\n\n**This skill supersedes `ml-paper-writing`** — it contains all of ml-paper-writing's content plus the full experiment\u002Fanalysis pipeline and autoreason methodology.\n\n### Hermes Tools Reference\n\n| Tool | Usage in This Pipeline |\n|------|----------------------|\n| **`terminal`** | LaTeX compilation (`latexmk -pdf`), git operations, launching experiments (`nohup python run.py &`), process checks |\n| **`process`** | Background experiment management: `process(\"start\", ...)`, `process(\"poll\", pid)`, `process(\"log\", pid)`, `process(\"kill\", pid)` |\n| **`execute_code`** | Run Python for citation verification, statistical analysis, data aggregation. Has tool access via RPC. |\n| **`read_file`** \u002F **`write_file`** \u002F **`patch`** | Paper editing, experiment scripts, result files. Use `patch` for targeted edits to large .tex files. |\n| **`web_search`** | Literature discovery: `web_search(\"transformer attention mechanism 2024\")` |\n| **`web_extract`** | Fetch paper content, verify citations: `web_extract(\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\")` |\n| **`delegate_task`** | **Parallel section drafting** — spawn isolated subagents for each section. Also for concurrent citation verification. |\n| **`todo`** | Primary state tracker across sessions. Update after every phase transition. |\n| **`memory`** | Persist key decisions across sessions: contribution framing, venue choice, reviewer feedback. |\n| **`cronjob`** | Schedule experiment monitoring, deadline countdowns, automated arXiv checks. |\n| **`clarify`** | Ask the user targeted questions when blocked (venue choice, contribution framing). |\n| **`send_message`** | Notify user when experiments complete or drafts are ready, even if user isn't in chat. |\n\n### Tool Usage Patterns\n\n**Experiment monitoring** (most common):\n```\nterminal(\"ps aux | grep \u003Cpattern>\")\n→ terminal(\"tail -30 \u003Clogfile>\")\n→ terminal(\"ls results\u002F\")\n→ execute_code(\"analyze results JSON, compute metrics\")\n→ terminal(\"git add -A && git commit -m '\u003Cdescriptive message>' && git push\")\n→ send_message(\"Experiment complete: \u003Csummary>\")\n```\n\n**Parallel section drafting** (using delegation):\n```\ndelegate_task(\"Draft the Methods section based on these experiment scripts and configs. \n  Include: pseudocode, all hyperparameters, architectural details sufficient for \n  reproduction. Write in LaTeX using the neurips2025 template conventions.\")\n\ndelegate_task(\"Draft the Related Work section. Use web_search and web_extract to \n  find papers. Verify every citation via Semantic Scholar. Group by methodology.\")\n\ndelegate_task(\"Draft the Experiments section. Read all result files in results\u002F. \n  State which claim each experiment supports. Include error bars and significance.\")\n```\n\nEach delegate runs as a **fresh subagent** with no shared context — provide all necessary information in the prompt. Collect outputs and integrate.\n\n**Citation verification** (using execute_code):\n```python\n# In execute_code:\nfrom semanticscholar import SemanticScholar\nimport requests\n\nsch = SemanticScholar()\nresults = sch.search_paper(\"attention mechanism transformers\", limit=5)\nfor paper in results:\n    doi = paper.externalIds.get('DOI', 'N\u002FA')\n    if doi != 'N\u002FA':\n        bibtex = requests.get(f\"https:\u002F\u002Fdoi.org\u002F{doi}\", \n                              headers={\"Accept\": \"application\u002Fx-bibtex\"}).text\n        print(bibtex)\n```\n\n### State Management with `memory` and `todo`\n\n**`memory` tool** — persist key decisions (bounded: ~2200 chars for MEMORY.md):\n\n```\nmemory(\"add\", \"Paper: autoreason. Venue: NeurIPS 2025 (9 pages). \n  Contribution: structured refinement works when generation-evaluation gap is wide.\n  Key results: Haiku 42\u002F42, Sonnet 3\u002F5, S4.6 constrained 2\u002F3.\n  Status: Phase 5 — drafting Methods section.\")\n```\n\nUpdate memory after major decisions or phase transitions. This persists across sessions.\n\n**`todo` tool** — track granular progress:\n\n```\ntodo(\"add\", \"Design constrained task experiments for Sonnet 4.6\")\ntodo(\"add\", \"Run Haiku baseline comparison\")\ntodo(\"add\", \"Draft Methods section\")\ntodo(\"update\", id=3, status=\"in_progress\")\ntodo(\"update\", id=1, status=\"completed\")\n```\n\n**Session startup protocol:**\n```\n1. todo(\"list\")                           # Check current task list\n2. memory(\"read\")                         # Recall key decisions\n3. terminal(\"git log --oneline -10\")      # Check recent commits\n4. terminal(\"ps aux | grep python\")       # Check running experiments\n5. terminal(\"ls results\u002F | tail -20\")     # Check for new results\n6. Report status to user, ask for direction\n```\n\n### Cron Monitoring with `cronjob`\n\nUse the `cronjob` tool to schedule periodic experiment checks:\n\n```\ncronjob(\"create\", {\n  \"schedule\": \"*\u002F30 * * * *\",  # Every 30 minutes\n  \"prompt\": \"Check experiment status:\n    1. ps aux | grep run_experiment\n    2. tail -30 logs\u002Fexperiment_haiku.log\n    3. ls results\u002Fhaiku_baselines\u002F\n    4. If complete: read results, compute Borda scores, \n       git add -A && git commit -m 'Add Haiku results' && git push\n    5. Report: table of results, key finding, next step\n    6. If nothing changed: respond with [SILENT]\"\n})\n```\n\n**[SILENT] protocol**: When nothing has changed since the last check, respond with exactly `[SILENT]`. This suppresses notification delivery to the user. Only report when there are genuine changes worth knowing about.\n\n**Deadline tracking**:\n```\ncronjob(\"create\", {\n  \"schedule\": \"0 9 * * *\",  # Daily at 9am\n  \"prompt\": \"NeurIPS 2025 deadline: May 22. Today is {date}. \n    Days remaining: {compute}. \n    Check todo list — are we on track? \n    If \u003C7 days: warn user about remaining tasks.\"\n})\n```\n\n### Communication Patterns\n\n**When to notify the user** (via `send_message` or direct response):\n- Experiment batch completed (with results table)\n- Unexpected finding or failure requiring decision\n- Draft section ready for review\n- Deadline approaching with incomplete tasks\n\n**When NOT to notify:**\n- Experiment still running, no new results → `[SILENT]`\n- Routine monitoring with no changes → `[SILENT]`\n- Intermediate steps that don't need attention\n\n**Report format** — always include structured data:\n```\n## Experiment: \u003Cname>\nStatus: Complete \u002F Running \u002F Failed\n\n| Task | Method A | Method B | Method C |\n|------|---------|---------|---------|\n| Task 1 | 85.2 | 82.1 | **89.4** |\n\nKey finding: \u003Cone sentence>\nNext step: \u003Cwhat happens next>\n```\n\n### Decision Points Requiring Human Input\n\nUse `clarify` for targeted questions when genuinely blocked:\n\n| Decision | When to Ask |\n|----------|-------------|\n| Target venue | Before starting paper (affects page limits, framing) |\n| Contribution framing | When multiple valid framings exist |\n| Experiment priority | When TODO list has more experiments than time allows |\n| Submission readiness | Before final submission |\n\n**Do NOT ask about** (be proactive, make a choice, flag it):\n- Word choice, section ordering\n- Which specific results to highlight\n- Citation completeness (draft with what you find, note gaps)\n\n---\n\n## Reviewer Evaluation Criteria\n\nUnderstanding what reviewers look for helps focus effort:\n\n| Criterion | What They Check |\n|-----------|----------------|\n| **Quality** | Technical soundness, well-supported claims, fair baselines |\n| **Clarity** | Clear writing, reproducible by experts, consistent notation |\n| **Significance** | Community impact, advances understanding |\n| **Originality** | New insights (doesn't require new method) |\n\n**Scoring (NeurIPS 6-point scale):**\n- 6: Strong Accept — groundbreaking, flawless\n- 5: Accept — technically solid, high impact\n- 4: Borderline Accept — solid, limited evaluation\n- 3: Borderline Reject — weaknesses outweigh\n- 2: Reject — technical flaws\n- 1: Strong Reject — known results or ethics issues\n\nSee [references\u002Freviewer-guidelines.md](references\u002Freviewer-guidelines.md) for detailed guidelines, common concerns, and rebuttal strategies.\n\n---\n\n## Common Issues and Solutions\n\n| Issue | Solution |\n|-------|----------|\n| Abstract too generic | Delete first sentence if it could prepend any ML paper. Start with your specific contribution. |\n| Introduction exceeds 1.5 pages | Split background into Related Work. Front-load contribution bullets. |\n| Experiments lack explicit claims | Add: \"This experiment tests whether [specific claim]...\" before each one. |\n| Reviewers find paper hard to follow | Add signposting, use consistent terminology, make figure captions self-contained. |\n| Missing statistical significance | Add error bars, number of runs, statistical tests, confidence intervals. |\n| Scope creep in experiments | Every experiment must map to a specific claim. Cut experiments that don't. |\n| Paper rejected, need to resubmit | See Conference Resubmission in Phase 7. Address reviewer concerns without referencing reviews. |\n\n---\n\n## Reference Documents\n\n| Document | Contents |\n|----------|----------|\n| [references\u002Fwriting-guide.md](references\u002Fwriting-guide.md) | Gopen & Swan 7 principles, Perez micro-tips, Lipton word choice, Steinhardt precision, figure design |\n| [references\u002Fcitation-workflow.md](references\u002Fcitation-workflow.md) | Citation APIs, Python code, CitationManager class, BibTeX management |\n| [references\u002Fchecklists.md](references\u002Fchecklists.md) | NeurIPS 16-item, ICML, ICLR, ACL requirements, universal pre-submission checklist |\n| [references\u002Freviewer-guidelines.md](references\u002Freviewer-guidelines.md) | Evaluation criteria, scoring, common concerns, rebuttal template |\n| [references\u002Fsources.md](references\u002Fsources.md) | Complete bibliography of all writing guides, conference guidelines, APIs |\n| [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) | Experiment design patterns, evaluation protocols, monitoring, error recovery |\n| [references\u002Fautoreason-methodology.md](references\u002Fautoreason-methodology.md) | Autoreason loop, strategy selection, model guide, prompts, scope constraints, Borda scoring |\n\n### LaTeX Templates\n\nTemplates in `templates\u002F` for: **NeurIPS 2025**, **ICML 2026**, **ICLR 2026**, **ACL**, **AAAI 2026**, **COLM 2025**.\n\nSee [templates\u002FREADME.md](templates\u002FREADME.md) for compilation instructions.\n\n### Key External Sources\n\n**Writing Philosophy:**\n- [Neel Nanda: How to Write ML Papers](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FeJGptPbbFPZGLpjsp\u002Fhighly-opinionated-advice-on-how-to-write-ml-papers)\n- [Sebastian Farquhar: How to Write ML Papers](https:\u002F\u002Fsebastianfarquhar.com\u002Fon-research\u002F2024\u002F11\u002F04\u002Fhow_to_write_ml_papers\u002F)\n- [Gopen & Swan: Science of Scientific Writing](https:\u002F\u002Fcseweb.ucsd.edu\u002F~swanson\u002Fpapers\u002Fscience-of-writing.pdf)\n- [Lipton: Heuristics for Scientific Writing](https:\u002F\u002Fwww.approximatelycorrect.com\u002F2018\u002F01\u002F29\u002Fheuristics-technical-scientific-writing-machine-learning-perspective\u002F)\n- [Perez: Easy Paper Writing Tips](https:\u002F\u002Fethanperez.net\u002Feasy-paper-writing-tips\u002F)\n\n**APIs:** [Semantic Scholar](https:\u002F\u002Fapi.semanticscholar.org\u002Fapi-docs\u002F) | [CrossRef](https:\u002F\u002Fwww.crossref.org\u002Fdocumentation\u002Fretrieve-metadata\u002Frest-api\u002F) | [arXiv](https:\u002F\u002Finfo.arxiv.org\u002Fhelp\u002Fapi\u002Fbasics.html)\n\n**Venues:** [NeurIPS](https:\u002F\u002Fneurips.cc\u002FConferences\u002F2025\u002FPaperInformation\u002FStyleFiles) | [ICML](https:\u002F\u002Ficml.cc\u002FConferences\u002F2025\u002FAuthorInstructions) | [ICLR](https:\u002F\u002Ficlr.cc\u002FConferences\u002F2026\u002FAuthorGuide) | [ACL](https:\u002F\u002Fgithub.com\u002Facl-org\u002Facl-style-files)\n","# Research Paper Writing Pipeline\n\nEnd-to-end pipeline for producing publication-ready ML\u002FAI research papers targeting **NeurIPS, ICML, ICLR, ACL, AAAI, and COLM**. This skill covers the full research lifecycle: experiment design, execution, monitoring, analysis, paper writing, review, revision, and submission.\n\nThis is **not a linear pipeline** — it is an iterative loop. Results trigger new experiments. Reviews trigger new analysis. The agent must handle these feedback loops.\n\n```\n┌─────────────────────────────────────────────────────────────┐\n│                    RESEARCH PAPER PIPELINE                  │\n│                                                             │\n│  Phase 0: Project Setup ──► Phase 1: Literature Review      │\n│       │                          │                          │\n│       ▼                          ▼                          │\n│  Phase 2: Experiment     Phase 5: Paper Drafting ◄──┐      │\n│       Design                     │                   │      │\n│       │                          ▼                   │      │\n│       ▼                    Phase 6: Self-Review      │      │\n│  Phase 3: Execution &           & Revision ──────────┘      │\n│       Monitoring                 │                          │\n│       │                          ▼                          │\n│       ▼                    Phase 7: Submission               │\n│  Phase 4: Analysis ─────► (feeds back to Phase 2 or 5)     │\n│                                                             │\n└─────────────────────────────────────────────────────────────┘\n```\n\n---\n\n## When To Use This Skill\n\nUse this skill when:\n- **Starting a new research paper** from an existing codebase or idea\n- **Designing and running experiments** to support paper claims\n- **Writing or revising** any section of a research paper\n- **Preparing for submission** to a specific conference\n- **Responding to reviews** with additional experiments or revisions\n- **Converting** a paper between conference formats\n\n## Core Philosophy\n\n1. **Be proactive.** Deliver complete drafts, not questions. Scientists are busy — produce something concrete they can react to, then iterate.\n2. **Never hallucinate citations.** AI-generated citations have ~40% error rate. Always fetch programmatically. Mark unverifiable citations as `[CITATION NEEDED]`.\n3. **Paper is a story, not a collection of experiments.** Every paper needs one clear contribution stated in a single sentence. If you can't do that, the paper isn't ready.\n4. **Experiments serve claims.** Every experiment must explicitly state which claim it supports. Never run experiments that don't connect to the paper's narrative.\n5. **Commit early, commit often.** Every completed experiment batch, every paper draft update — commit with descriptive messages. Git log is the experiment history.\n\n### Proactivity and Collaboration\n\n**Default: Be proactive. Draft first, ask with the draft.**\n\n| Confidence Level | Action |\n|-----------------|--------|\n| **High** (clear repo, obvious contribution) | Write full draft, deliver, iterate on feedback |\n| **Medium** (some ambiguity) | Write draft with flagged uncertainties, continue |\n| **Low** (major unknowns) | Ask 1-2 targeted questions via `clarify`, then draft |\n\n| Section | Draft Autonomously? | Flag With Draft |\n|---------|-------------------|-----------------|\n| Abstract | Yes | \"Framed contribution as X — adjust if needed\" |\n| Introduction | Yes | \"Emphasized problem Y — correct if wrong\" |\n| Methods | Yes | \"Included details A, B, C — add missing pieces\" |\n| Experiments | Yes | \"Highlighted results 1, 2, 3 — reorder if needed\" |\n| Related Work | Yes | \"Cited papers X, Y, Z — add any I missed\" |\n\n**Block for input only when**: target venue unclear, multiple contradictory framings, results seem incomplete, explicit request to review first.\n\n---\n\n## Phase 0: Project Setup\n\n**Goal**: Establish the workspace, understand existing work, identify the contribution.\n\n### Step 0.1: Explore the Repository\n\n```bash\n# Understand project structure\nls -la\nfind . -name \"*.py\" | head -30\nfind . -name \"*.md\" -o -name \"*.txt\" | xargs grep -l -i \"result\\|conclusion\\|finding\"\n```\n\nLook for:\n- `README.md` — project overview and claims\n- `results\u002F`, `outputs\u002F`, `experiments\u002F` — existing findings\n- `configs\u002F` — experimental settings\n- `.bib` files — existing citations\n- Draft documents or notes\n\n### Step 0.2: Organize the Workspace\n\nEstablish a consistent workspace structure:\n\n```\nworkspace\u002F\n  paper\u002F               # LaTeX source, figures, compiled PDFs\n  experiments\u002F         # Experiment runner scripts\n  code\u002F                # Core method implementation\n  results\u002F             # Raw experiment results (auto-generated)\n  tasks\u002F               # Task\u002Fbenchmark definitions\n  human_eval\u002F          # Human evaluation materials (if needed)\n```\n\n### Step 0.3: Set Up Version Control\n\n```bash\ngit init  # if not already\ngit remote add origin \u003Crepo-url>\ngit checkout -b paper-draft  # or main\n```\n\n**Git discipline**: Every completed experiment batch gets committed with a descriptive message. Example:\n```\nAdd Monte Carlo constrained results (5 runs, Sonnet 4.6, policy memo task)\nAdd Haiku baseline comparison: autoreason vs refinement baselines at cheap model tier\n```\n\n### Step 0.4: Identify the Contribution\n\nBefore writing anything, articulate:\n- **The What**: What is the single thing this paper contributes?\n- **The Why**: What evidence supports it?\n- **The So What**: Why should readers care?\n\n> Propose to the scientist: \"Based on my understanding, the main contribution is: [one sentence]. The key results show [Y]. Is this the framing you want?\"\n\n### Step 0.5: Create a TODO List\n\nUse the `todo` tool to create a structured project plan:\n\n```\nResearch Paper TODO:\n- [ ] Define one-sentence contribution\n- [ ] Literature review (related work + baselines)\n- [ ] Design core experiments\n- [ ] Run experiments\n- [ ] Analyze results\n- [ ] Write first draft\n- [ ] Self-review (simulate reviewers)\n- [ ] Revise based on review\n- [ ] Submission prep\n```\n\nUpdate this throughout the project. It serves as the persistent state across sessions.\n\n---\n\n## Phase 1: Literature Review\n\n**Goal**: Find related work, identify baselines, gather citations.\n\n### Step 1.1: Identify Seed Papers\n\nStart from papers already referenced in the codebase:\n\n```bash\n# Via terminal:\ngrep -r \"arxiv\\|doi\\|cite\" --include=\"*.md\" --include=\"*.bib\" --include=\"*.py\"\nfind . -name \"*.bib\"\n```\n\n### Step 1.2: Search for Related Work\n\n**Load the `arxiv` skill** for structured paper discovery: `skill_view(\"arxiv\")`. It provides arXiv REST API search, Semantic Scholar citation graphs, author profiles, and BibTeX generation.\n\nUse `web_search` for broad discovery, `web_extract` for fetching specific papers:\n\n```\n# Via web_search:\nweb_search(\"[main technique] + [application domain] site:arxiv.org\")\nweb_search(\"[baseline method] comparison ICML NeurIPS 2024\")\n\n# Via web_extract (for specific papers):\nweb_extract(\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\")\n```\n\nAdditional search queries to try:\n\n```\nSearch queries:\n- \"[main technique] + [application domain]\"\n- \"[baseline method] comparison\"\n- \"[problem name] state-of-the-art\"\n- Author names from existing citations\n```\n\n**Recommended**: Install **Exa MCP** for real-time academic search:\n```bash\nclaude mcp add exa -- npx -y mcp-remote \"https:\u002F\u002Fmcp.exa.ai\u002Fmcp\"\n```\n\n### Step 1.3: Verify Every Citation\n\n**NEVER generate BibTeX from memory. ALWAYS fetch programmatically.**\n\nFor each citation, follow the mandatory 5-step process:\n\n```\nCitation Verification (MANDATORY per citation):\n1. SEARCH → Query Semantic Scholar or Exa MCP with specific keywords\n2. VERIFY → Confirm paper exists in 2+ sources (Semantic Scholar + arXiv\u002FCrossRef)\n3. RETRIEVE → Get BibTeX via DOI content negotiation (programmatically, not from memory)\n4. VALIDATE → Confirm the claim you're citing actually appears in the paper\n5. ADD → Add verified BibTeX to bibliography\nIf ANY step fails → mark as [CITATION NEEDED], inform scientist\n```\n\n```python\n# Fetch BibTeX via DOI\nimport requests\n\ndef doi_to_bibtex(doi: str) -> str:\n    response = requests.get(\n        f\"https:\u002F\u002Fdoi.org\u002F{doi}\",\n        headers={\"Accept\": \"application\u002Fx-bibtex\"}\n    )\n    response.raise_for_status()\n    return response.text\n```\n\nIf you cannot verify a citation:\n\n```latex\n\\cite{PLACEHOLDER_author2024_verify_this}  % TODO: Verify this citation exists\n```\n\n**Always tell the scientist**: \"I've marked [X] citations as placeholders that need verification.\"\n\nSee [references\u002Fcitation-workflow.md](references\u002Fcitation-workflow.md) for complete API documentation and the full `CitationManager` class.\n\n### Step 1.4: Organize Related Work\n\nGroup papers by methodology, not paper-by-paper:\n\n**Good**: \"One line of work uses X's assumption [refs] whereas we use Y's assumption because...\"\n**Bad**: \"Smith et al. introduced X. Jones et al. introduced Y. We combine both.\"\n\n---\n\n## Phase 2: Experiment Design\n\n**Goal**: Design experiments that directly support paper claims. Every experiment must answer a specific question.\n\n### Step 2.1: Map Claims to Experiments\n\nCreate an explicit mapping:\n\n| Claim | Experiment | Expected Evidence |\n|-------|-----------|-------------------|\n| \"Our method outperforms baselines\" | Main comparison (Table 1) | Win rate, statistical significance |\n| \"Effect is larger for weaker models\" | Model scaling study | Monotonic improvement curve |\n| \"Convergence requires scope constraints\" | Constrained vs unconstrained | Convergence rate comparison |\n\n**Rule**: If an experiment doesn't map to a claim, don't run it.\n\n### Step 2.2: Design Baselines\n\nStrong baselines are what separates accepted papers from rejected ones. Reviewers will ask: \"Did they compare against X?\"\n\nStandard baseline categories:\n- **Naive baseline**: Simplest possible approach\n- **Strong baseline**: Best known existing method\n- **Ablation baselines**: Your method minus one component\n- **Compute-matched baselines**: Same compute budget, different allocation\n\n### Step 2.3: Define Evaluation Protocol\n\nBefore running anything, specify:\n- **Metrics**: What you're measuring, direction symbols (higher\u002Flower better)\n- **Aggregation**: How results are combined across runs\u002Ftasks\n- **Statistical tests**: What tests will establish significance\n- **Sample sizes**: How many runs\u002Fproblems\u002Ftasks\n\n### Step 2.4: Write Experiment Scripts\n\nFollow these patterns from successful research pipelines:\n\n**Incremental saving** — save results after each step for crash recovery:\n```python\n# Save after each problem\u002Ftask\nresult_path = f\"results\u002F{task}\u002F{strategy}\u002Fresult.json\"\nif os.path.exists(result_path):\n    continue  # Skip already-completed work\n# ... run experiment ...\nwith open(result_path, 'w') as f:\n    json.dump(result, f, indent=2)\n```\n\n**Artifact preservation** — save all intermediate outputs:\n```\nresults\u002F\u003Cexperiment>\u002F\n  \u003Ctask>\u002F\n    \u003Cstrategy>\u002F\n      final_output.md          # Final result\n      history.json             # Full trajectory\n      pass_01\u002F                 # Per-iteration artifacts\n        version_a.md\n        version_b.md\n        critic.md\n```\n\n**Separation of concerns** — keep generation, evaluation, and visualization separate:\n```\nrun_experiment.py              # Core experiment runner\nrun_baselines.py               # Baseline comparison\nrun_comparison_judge.py        # Blind evaluation\nanalyze_results.py             # Statistical analysis\nmake_charts.py                 # Visualization\n```\n\nSee [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) for complete design patterns, cron monitoring, and error recovery.\n\n---\n\n## Phase 3: Experiment Execution & Monitoring\n\n**Goal**: Run experiments reliably, monitor progress, recover from failures.\n\n### Step 3.1: Launch Experiments\n\nUse `nohup` for long-running experiments:\n\n```bash\nnohup python run_experiment.py --config config.yaml > logs\u002Fexperiment_01.log 2>&1 &\necho $!  # Record the PID\n```\n\n**Parallel execution**: Run independent experiments simultaneously, but be aware of API rate limits. 4+ concurrent experiments on the same API will slow each down.\n\n### Step 3.2: Set Up Monitoring (Cron Pattern)\n\nFor long-running experiments, set up periodic status checks. The cron prompt should follow this template:\n\n```\nMonitor Prompt Template:\n1. Check if process is still running: ps aux | grep \u003Cpattern>\n2. Read last 30 lines of log: tail -30 \u003Clogfile>\n3. Check for completed results: ls \u003Cresult_dir>\n4. If results exist, read and report: cat \u003Cresult_file>\n5. If all done, commit: git add -A && git commit -m \"\u003Cdescriptive message>\" && git push\n6. Report in structured format (tables with key metrics)\n7. Answer the key analytical question for this experiment\n```\n\n**Silent mode**: If nothing has changed since the last check, respond with `[SILENT]` to suppress notification to the user. Only report when there's news.\n\n### Step 3.3: Handle Failures\n\nCommon failure modes and recovery:\n\n| Failure | Detection | Recovery |\n|---------|-----------|----------|\n| API rate limit \u002F credit exhaustion | 402\u002F429 errors in logs | Wait, then re-run (scripts skip completed work) |\n| Process crash | PID gone, incomplete results | Re-run from last checkpoint |\n| Timeout on hard problems | Process stuck, no log progress | Kill and skip, note in results |\n| Wrong model ID | Errors referencing model name | Fix ID and re-run |\n\n**Key**: Scripts should always check for existing results and skip completed work. This makes re-runs safe and efficient.\n\n### Step 3.4: Commit Completed Results\n\nAfter each experiment batch completes:\n\n```bash\ngit add -A\ngit commit -m \"Add \u003Cexperiment name>: \u003Ckey finding in 1 line>\"\ngit push\n```\n\n---\n\n## Phase 4: Result Analysis\n\n**Goal**: Extract findings, compute statistics, identify the story.\n\n### Step 4.1: Aggregate Results\n\nWrite analysis scripts that:\n1. Load all result files from a batch\n2. Compute per-task and aggregate metrics\n3. Generate summary tables\n\n```python\n# Standard analysis pattern\nimport json, os\nfrom pathlib import Path\n\nresults = {}\nfor result_file in Path(\"results\u002F\").rglob(\"result.json\"):\n    data = json.loads(result_file.read_text())\n    strategy = result_file.parent.name\n    task = result_file.parent.parent.name\n    results.setdefault(strategy, {})[task] = data\n\n# Compute aggregate metrics\nfor strategy, tasks in results.items():\n    scores = [t[\"score\"] for t in tasks.values()]\n    print(f\"{strategy}: mean={np.mean(scores):.1f}, std={np.std(scores):.1f}\")\n```\n\n### Step 4.2: Statistical Significance\n\nAlways compute:\n- **Error bars**: Standard deviation or standard error, specify which\n- **Confidence intervals**: 95% CI for key results\n- **Pairwise tests**: McNemar's test for comparing two methods\n- **Effect sizes**: Cohen's d or h for practical significance\n\nSee [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) for complete implementations of McNemar's test, bootstrapped CIs, and Cohen's h.\n\n### Step 4.3: Identify the Story\n\nAfter analysis, explicitly answer:\n1. **What is the main finding?** State it in one sentence.\n2. **What surprised you?** Unexpected results often make the best papers.\n3. **What failed?** Failed experiments can be the most informative. Honest reporting of failures strengthens the paper.\n4. **What follow-up experiments are needed?** Results often raise new questions.\n\n### Step 4.4: Create Figures and Tables\n\n**Figures**:\n- Use vector graphics (PDF) for all plots: `plt.savefig('fig.pdf')`\n- Colorblind-safe palettes (Okabe-Ito or Paul Tol)\n- Self-contained captions — reader should understand without main text\n- No title inside figure — the caption serves this function\n\n**Tables**:\n- Use `booktabs` LaTeX package\n- Bold best value per metric\n- Include direction symbols (higher\u002Flower better)\n- Consistent decimal precision\n\n```latex\n\\usepackage{booktabs}\n\\begin{tabular}{lcc}\n\\toprule\nMethod & Accuracy $\\uparrow$ & Latency $\\downarrow$ \\\\\n\\midrule\nBaseline & 85.2 & 45ms \\\\\n\\textbf{Ours} & \\textbf{92.1} & 38ms \\\\\n\\bottomrule\n\\end{tabular}\n```\n\n### Step 4.5: Decide: More Experiments or Write?\n\n| Situation | Action |\n|-----------|--------|\n| Core claims supported, results significant | Move to Phase 5 (writing) |\n| Results inconclusive, need more data | Back to Phase 2 (design) |\n| Unexpected finding suggests new direction | Back to Phase 2 (design) |\n| Missing one ablation reviewers will ask for | Run it, then Phase 5 |\n| All experiments done but some failed | Note failures, move to Phase 5 |\n\n---\n\n## Iterative Refinement: Strategy Selection\n\nAny output in this pipeline — paper drafts, experiment scripts, analysis — can be iteratively refined. The autoreason research provides empirical evidence for when each refinement strategy works and when it fails. Use this section to choose the right approach.\n\n### Quick Decision Table\n\n| Your Situation | Strategy | Why |\n|---------------|----------|-----|\n| Mid-tier model + constrained task | **Autoreason** | Sweet spot. Generation-evaluation gap is widest. Baselines actively destroy weak model outputs. |\n| Mid-tier model + open task | **Autoreason** with scope constraints added | Add fixed facts, structure, or deliverable to bound the improvement space. |\n| Frontier model + constrained task | **Autoreason** | Wins 2\u002F3 constrained tasks even at frontier. |\n| Frontier model + unconstrained task | **Critique-and-revise** or **single pass** | Autoreason comes last. Model self-evaluates well enough. |\n| Concrete technical task (system design) | **Critique-and-revise** | Direct find-and-fix loop is more efficient. |\n| Template-filling task (one correct structure) | **Single pass** or **conservative** | Minimal decision space. Iteration adds no value. |\n| Code with test cases | **Autoreason (code variant)** | Structured analysis of *why* it failed before fixing. Recovery rate 62% vs 43%. |\n| Very weak model (Llama 8B class) | **Single pass** | Model too weak for diverse candidates. Invest in generation quality. |\n\n### The Generation-Evaluation Gap\n\n**Core insight**: Autoreason's value depends on the gap between a model's generation capability and its self-evaluation capability.\n\n```\nModel Tier        │ Generation │ Self-Eval │ Gap    │ Autoreason Value\n──────────────────┼────────────┼───────────┼────────┼─────────────────\nWeak (Llama 8B)   │ Poor       │ Poor      │ Small  │ None — can't generate diverse candidates\nMid (Haiku 3.5)   │ Decent     │ Poor      │ LARGE  │ MAXIMUM — 42\u002F42 perfect Borda\nMid (Gemini Flash)│ Decent     │ Moderate  │ Large  │ High — wins 2\u002F3\nStrong (Sonnet 4) │ Good       │ Decent    │ Medium │ Moderate — wins 3\u002F5\nFrontier (S4.6)   │ Excellent  │ Good      │ Small  │ Only with constraints\n```\n\nThis gap is structural, not temporary. As costs drop, today's frontier becomes tomorrow's mid-tier. The sweet spot moves but never disappears.\n\n### Autoreason Loop (Summary)\n\nEach pass produces three candidates from fresh, isolated agents:\n\n1. **Critic** → finds problems in incumbent A (no fixes)\n2. **Author B** → revises A based on critique\n3. **Synthesizer** → merges A and B (randomized labels)\n4. **Judge Panel** → 3 blind CoT judges rank A, B, AB via Borda count\n5. **Convergence** → A wins k=2 consecutive passes → done\n\n**Key parameters:**\n- k=2 convergence (k=1 premature, k=3 too expensive, no quality gain)\n- CoT judges always (3x faster convergence)\n- Temperature 0.8 authors, 0.3 judges\n- Conservative tiebreak: incumbent wins ties\n- Every role is a fresh agent with no shared context\n\n### Applying to Paper Drafts\n\nWhen refining the paper itself through autoreason:\n- **Provide ground truth to the critic**: actual experimental data, result JSONs, statistical outputs. Without this, models hallucinate fabricated ablation studies and fake confidence intervals.\n- **Use 3 working judges minimum**: A broken judge parser doesn't add noise — it prevents equilibrium entirely.\n- **Scope constrain the revision**: \"Address these specific weaknesses\" not \"improve the paper.\"\n\n### Failure Modes\n\n| Failure | Detection | Fix |\n|---------|-----------|-----|\n| No convergence (A never wins) | A wins \u003C15% over 20+ passes | Add scope constraints to the task |\n| Synthesis drift | Word counts grow unboundedly | Constrain structure and deliverable |\n| Degradation below single pass | Baselines score higher than iterated output | Switch to single pass; model may be too weak |\n| Overfitting (code) | High public-test pass, low private-test pass | Use structured analysis, not just test feedback |\n| Broken judges | Parsing failures reduce panel below 3 | Fix parser before continuing |\n\nSee [references\u002Fautoreason-methodology.md](references\u002Fautoreason-methodology.md) for complete prompts, Borda scoring details, model selection guide, scope constraint design patterns, and compute budget reference.\n\n---\n\n## Phase 5: Paper Drafting\n\n**Goal**: Write a complete, publication-ready paper.\n\n### The Narrative Principle\n\n**The single most critical insight**: Your paper is not a collection of experiments — it's a story with one clear contribution supported by evidence.\n\nEvery successful ML paper centers on what Neel Nanda calls \"the narrative\": a short, rigorous, evidence-based technical story with a takeaway readers care about.\n\n**Three Pillars (must be crystal clear by end of introduction):**\n\n| Pillar | Description | Test |\n|--------|-------------|------|\n| **The What** | 1-3 specific novel claims | Can you state them in one sentence? |\n| **The Why** | Rigorous empirical evidence | Do experiments distinguish your hypothesis from alternatives? |\n| **The So What** | Why readers should care | Does this connect to a recognized community problem? |\n\n**If you cannot state your contribution in one sentence, you don't yet have a paper.**\n\n### Time Allocation\n\nSpend approximately **equal time** on each of:\n1. The abstract\n2. The introduction\n3. The figures\n4. Everything else combined\n\n**Why?** Most reviewers form judgments before reaching your methods. Readers encounter your paper as: title → abstract → introduction → figures → maybe the rest.\n\n### Writing Workflow\n\n```\nPaper Writing Checklist:\n- [ ] Step 1: Define the one-sentence contribution\n- [ ] Step 2: Draft Figure 1 (core idea or most compelling result)\n- [ ] Step 3: Draft abstract (5-sentence formula)\n- [ ] Step 4: Draft introduction (1-1.5 pages max)\n- [ ] Step 5: Draft methods\n- [ ] Step 6: Draft experiments & results\n- [ ] Step 7: Draft related work\n- [ ] Step 8: Draft conclusion & discussion\n- [ ] Step 9: Draft limitations (REQUIRED by all venues)\n- [ ] Step 10: Plan appendix (proofs, extra experiments, details)\n- [ ] Step 11: Complete paper checklist\n- [ ] Step 12: Final review\n```\n\n### Step 5.0: Title\n\nThe title is the single most-read element of the paper. It determines whether anyone clicks through to the abstract.\n\n**Good titles**:\n- State the contribution or finding: \"Autoreason: When Iterative LLM Refinement Works and Why It Fails\"\n- Highlight a surprising result: \"Scaling Data-Constrained Language Models\" (implies you can)\n- Name the method + what it does: \"DPO: Direct Preference Optimization of Language Models\"\n\n**Bad titles**:\n- Too generic: \"An Approach to Improving Language Model Outputs\"\n- Too long: anything over ~15 words\n- Jargon-only: \"Asymptotic Convergence of Iterative Stochastic Policy Refinement\" (who is this for?)\n\n**Rules**:\n- Include your method name if you have one (for citability)\n- Include 1-2 keywords reviewers will search for\n- Avoid colons unless both halves carry meaning\n- Test: would a reviewer know the domain and contribution from the title alone?\n\n### Step 5.1: Abstract (5-Sentence Formula)\n\nFrom Sebastian Farquhar (DeepMind):\n\n```\n1. What you achieved: \"We introduce...\", \"We prove...\", \"We demonstrate...\"\n2. Why this is hard and important\n3. How you do it (with specialist keywords for discoverability)\n4. What evidence you have\n5. Your most remarkable number\u002Fresult\n```\n\n**Delete** generic openings like \"Large language models have achieved remarkable success...\"\n\n### Step 5.2: Figure 1\n\nFigure 1 is the second thing most readers look at (after abstract). Draft it before writing the introduction — it forces you to clarify the core idea.\n\n| Figure 1 Type | When to Use | Example |\n|---------------|-------------|---------|\n| **Method diagram** | New architecture or pipeline | TikZ flowchart showing your system |\n| **Results teaser** | One compelling result tells the whole story | Bar chart: \"Ours vs baselines\" with clear gap |\n| **Problem illustration** | The problem is unintuitive | Before\u002Fafter showing failure mode you fix |\n| **Conceptual diagram** | Abstract contribution needs visual grounding | 2x2 matrix of method properties |\n\n**Rules**: Figure 1 must be understandable without reading any text. The caption alone should communicate the core idea. Use color purposefully — don't just decorate.\n\n### Step 5.3: Introduction (1-1.5 pages max)\n\nMust include:\n- Clear problem statement\n- Brief approach overview\n- 2-4 bullet contribution list (max 1-2 lines each in two-column format)\n- Methods should start by page 2-3\n\n### Step 5.3: Methods\n\nEnable reimplementation:\n- Conceptual outline or pseudocode\n- All hyperparameters listed\n- Architectural details sufficient for reproduction\n- Present final design decisions; ablations go in experiments\n\n### Step 5.4: Experiments & Results\n\nFor each experiment, explicitly state:\n- **What claim it supports**\n- How it connects to main contribution\n- What to observe: \"the blue line shows X, which demonstrates Y\"\n\nRequirements:\n- Error bars with methodology (std dev vs std error)\n- Hyperparameter search ranges\n- Compute infrastructure (GPU type, total hours)\n- Seed-setting methods\n\n### Step 5.5: Related Work\n\nOrganize methodologically, not paper-by-paper. Cite generously — reviewers likely authored relevant papers.\n\n### Step 5.6: Limitations (REQUIRED)\n\nAll major conferences require this. Honesty helps:\n- Reviewers are instructed not to penalize honest limitation acknowledgment\n- Pre-empt criticisms by identifying weaknesses first\n- Explain why limitations don't undermine core claims\n\n### Step 5.7: Conclusion & Discussion\n\n**Conclusion** (required, 0.5-1 page):\n- Restate the contribution in one sentence (different wording from abstract)\n- Summarize key findings (2-3 sentences, not a list)\n- Implications: what does this mean for the field?\n- Future work: 2-3 concrete next steps (not vague \"we leave X for future work\")\n\n**Discussion** (optional, sometimes combined with conclusion):\n- Broader implications beyond immediate results\n- Connections to other subfields\n- Honest assessment of when the method does and doesn't work\n- Practical deployment considerations\n\n**Do NOT** introduce new results or claims in the conclusion.\n\n### Step 5.8: Appendix Strategy\n\nAppendices are unlimited at all major venues and are essential for reproducibility. Structure:\n\n| Appendix Section | What Goes Here |\n|-----------------|---------------|\n| **Proofs & Derivations** | Full proofs too long for main text. Main text can state theorems with \"proof in Appendix A.\" |\n| **Additional Experiments** | Ablations, scaling curves, per-dataset breakdowns, hyperparameter sensitivity |\n| **Implementation Details** | Full hyperparameter tables, training details, hardware specs, random seeds |\n| **Dataset Documentation** | Data collection process, annotation guidelines, licensing, preprocessing |\n| **Prompts & Templates** | Exact prompts used (for LLM-based methods), evaluation templates |\n| **Human Evaluation** | Annotation interface screenshots, instructions given to annotators, IRB details |\n| **Additional Figures** | Per-task breakdowns, trajectory visualizations, failure case examples |\n\n**Rules**:\n- The main paper must be self-contained — reviewers are not required to read appendices\n- Never put critical evidence only in the appendix\n- Cross-reference: \"Full results in Table 5 (Appendix B)\" not just \"see appendix\"\n- Use `\\appendix` command, then `\\section{A: Proofs}` etc.\n\n### Page Budget Management\n\nWhen over the page limit:\n\n| Cut Strategy | Saves | Risk |\n|-------------|-------|------|\n| Move proofs to appendix | 0.5-2 pages | Low — standard practice |\n| Condense related work | 0.5-1 page | Medium — may miss key citations |\n| Combine tables with subfigures | 0.25-0.5 page | Low — often improves readability |\n| Use `\\vspace{-Xpt}` sparingly | 0.1-0.3 page | Low if subtle, high if obvious |\n| Remove qualitative examples | 0.5-1 page | Medium — reviewers like examples |\n| Reduce figure sizes | 0.25-0.5 page | High — figures must remain readable |\n\n**Do NOT**: reduce font size, change margins, remove required sections (limitations, broader impact), or use `\\small`\u002F`\\footnotesize` for main text.\n\n### Writing Style\n\n**Sentence-level clarity (Gopen & Swan's 7 Principles):**\n\n| Principle | Rule |\n|-----------|------|\n| Subject-verb proximity | Keep subject and verb close |\n| Stress position | Place emphasis at sentence ends |\n| Topic position | Put context first, new info after |\n| Old before new | Familiar info → unfamiliar info |\n| One unit, one function | Each paragraph makes one point |\n| Action in verb | Use verbs, not nominalizations |\n| Context before new | Set stage before presenting |\n\n**Word choice (Lipton, Steinhardt):**\n- Be specific: \"accuracy\" not \"performance\"\n- Eliminate hedging: drop \"may\" unless genuinely uncertain\n- Consistent terminology throughout\n- Avoid incremental vocabulary: \"develop\", not \"combine\"\n\n**Full writing guide with examples**: See [references\u002Fwriting-guide.md](references\u002Fwriting-guide.md)\n\n### Using LaTeX Templates\n\n**Always copy the entire template directory first, then write within it.**\n\n```\nTemplate Setup Checklist:\n- [ ] Step 1: Copy entire template directory to new project\n- [ ] Step 2: Verify template compiles as-is (before any changes)\n- [ ] Step 3: Read the template's example content to understand structure\n- [ ] Step 4: Replace example content section by section\n- [ ] Step 5: Use template macros (check preamble for \\newcommand definitions)\n- [ ] Step 6: Clean up template artifacts only at the end\n```\n\n**Step 1: Copy the Full Template**\n\n```bash\ncp -r templates\u002Fneurips2025\u002F ~\u002Fpapers\u002Fmy-paper\u002F\ncd ~\u002Fpapers\u002Fmy-paper\u002F\nls -la  # Should see: main.tex, neurips.sty, Makefile, etc.\n```\n\nCopy the ENTIRE directory, not just the .tex file. Templates include style files (.sty), bibliography styles (.bst), example content, and Makefiles.\n\n**Step 2: Verify Template Compiles First**\n\nBefore making ANY changes:\n```bash\nlatexmk -pdf main.tex\n# Or manual: pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex\n```\n\nIf the unmodified template doesn't compile, fix that first (usually missing TeX packages — install via `tlmgr install \u003Cpackage>`).\n\n**Step 3: Keep Template Content as Reference**\n\nDon't immediately delete example content. Comment it out and use as formatting reference:\n```latex\n% Template example (keep for reference):\n% \\begin{figure}[t]\n%   \\centering\n%   \\includegraphics[width=0.8\\linewidth]{example-image}\n%   \\caption{Template shows caption style}\n% \\end{figure}\n\n% Your actual figure:\n\\begin{figure}[t]\n  \\centering\n  \\includegraphics[width=0.8\\linewidth]{your-figure.pdf}\n  \\caption{Your caption following the same style.}\n\\end{figure}\n```\n\n**Step 4: Replace Content Section by Section**\n\nWork through systematically: title\u002Fauthors → abstract → introduction → methods → experiments → related work → conclusion → references → appendix. Compile after each section.\n\n**Step 5: Use Template Macros**\n\n```latex\n\\newcommand{\\method}{YourMethodName}  % Consistent method naming\n\\newcommand{\\eg}{e.g.,\\xspace}        % Proper abbreviations\n\\newcommand{\\ie}{i.e.,\\xspace}\n```\n\n### Template Pitfalls\n\n| Pitfall | Problem | Solution |\n|---------|---------|----------|\n| Copying only `.tex` file | Missing `.sty`, won't compile | Copy entire directory |\n| Modifying `.sty` files | Breaks conference formatting | Never edit style files |\n| Adding random packages | Conflicts, breaks template | Only add if necessary |\n| Deleting template content early | Lose formatting reference | Keep as comments until done |\n| Not compiling frequently | Errors accumulate | Compile after each section |\n| Raster PNGs for figures | Blurry in paper | Always use vector PDF via `savefig('fig.pdf')` |\n\n### Quick Template Reference\n\n| Conference | Main File | Style File | Page Limit |\n|------------|-----------|------------|------------|\n| NeurIPS 2025 | `main.tex` | `neurips.sty` | 9 pages |\n| ICML 2026 | `example_paper.tex` | `icml2026.sty` | 8 pages |\n| ICLR 2026 | `iclr2026_conference.tex` | `iclr2026_conference.sty` | 9 pages |\n| ACL 2025 | `acl_latex.tex` | `acl.sty` | 8 pages (long) |\n| AAAI 2026 | `aaai2026-unified-template.tex` | `aaai2026.sty` | 7 pages |\n| COLM 2025 | `colm2025_conference.tex` | `colm2025_conference.sty` | 9 pages |\n\n**Universal**: Double-blind, references don't count, appendices unlimited, LaTeX required.\n\nTemplates in `templates\u002F` directory. See [templates\u002FREADME.md](templates\u002FREADME.md) for compilation setup (VS Code, CLI, Overleaf, other IDEs).\n\n### Tables and Figures\n\n**Tables** — use `booktabs` for professional formatting:\n\n```latex\n\\usepackage{booktabs}\n\\begin{tabular}{lcc}\n\\toprule\nMethod & Accuracy $\\uparrow$ & Latency $\\downarrow$ \\\\\n\\midrule\nBaseline & 85.2 & 45ms \\\\\n\\textbf{Ours} & \\textbf{92.1} & 38ms \\\\\n\\bottomrule\n\\end{tabular}\n```\n\nRules:\n- Bold best value per metric\n- Include direction symbols ($\\uparrow$ higher better, $\\downarrow$ lower better)\n- Right-align numerical columns\n- Consistent decimal precision\n\n**Figures**:\n- **Vector graphics** (PDF, EPS) for all plots and diagrams — `plt.savefig('fig.pdf')`\n- **Raster** (PNG 600 DPI) only for photographs\n- **Colorblind-safe palettes** (Okabe-Ito or Paul Tol)\n- Verify **grayscale readability** (8% of men have color vision deficiency)\n- **No title inside figure** — the caption serves this function\n- **Self-contained captions** — reader should understand without main text\n\n### Conference Resubmission\n\nFor converting between venues, see Phase 7 (Submission Preparation) — it covers the full conversion workflow, page-change table, and post-rejection guidance.\n\n### Professional LaTeX Preamble\n\nAdd these packages to any paper for professional quality. They are compatible with all major conference style files:\n\n```latex\n% --- Professional Packages (add after conference style file) ---\n\n% Typography\n\\usepackage{microtype}              % Microtypographic improvements (protrusion, expansion)\n                                     % Makes text noticeably more polished — always include\n\n% Tables\n\\usepackage{booktabs}               % Professional table rules (\\toprule, \\midrule, \\bottomrule)\n\\usepackage{siunitx}                % Consistent number formatting, decimal alignment\n                                     % Usage: \\num{12345} → 12,345; \\SI{3.5}{GHz} → 3.5 GHz\n                                     % Table alignment: S column type for decimal-aligned numbers\n\n% Figures\n\\usepackage{graphicx}               % Include graphics (\\includegraphics)\n\\usepackage{subcaption}             % Subfigures with (a), (b), (c) labels\n                                     % Usage: \\begin{subfigure}{0.48\\textwidth} ... \\end{subfigure}\n\n% Diagrams and Algorithms\n\\usepackage{tikz}                   % Programmable vector diagrams\n\\usetikzlibrary{arrows.meta, positioning, shapes.geometric, calc, fit, backgrounds}\n\\usepackage[ruled,vlined]{algorithm2e}  % Professional pseudocode\n                                     % Alternative: \\usepackage{algorithmicx} if template bundles it\n\n% Cross-references\n\\usepackage{cleveref}               % Smart references: \\cref{fig:x} → \"Figure 1\"\n                                     % MUST be loaded AFTER hyperref\n                                     % Handles: figures, tables, sections, equations, algorithms\n\n% Math (usually included by conference .sty, but verify)\n\\usepackage{amsmath,amssymb}        % AMS math environments and symbols\n\\usepackage{mathtools}              % Extends amsmath (dcases, coloneqq, etc.)\n\n% Colors (for figures and diagrams)\n\\usepackage{xcolor}                 % Color management\n% Okabe-Ito colorblind-safe palette:\n\\definecolor{okblue}{HTML}{0072B2}\n\\definecolor{okorange}{HTML}{E69F00}\n\\definecolor{okgreen}{HTML}{009E73}\n\\definecolor{okred}{HTML}{D55E00}\n\\definecolor{okpurple}{HTML}{CC79A7}\n\\definecolor{okcyan}{HTML}{56B4E9}\n\\definecolor{okyellow}{HTML}{F0E442}\n```\n\n**Notes:**\n- `microtype` is the single highest-impact package for visual quality. It adjusts character spacing at a sub-pixel level. Always include it.\n- `siunitx` handles decimal alignment in tables via the `S` column type — eliminates manual spacing.\n- `cleveref` must be loaded **after** `hyperref`. Most conference .sty files load hyperref, so put cleveref last.\n- Check if the conference template already loads any of these (especially `algorithm`, `amsmath`, `graphicx`). Don't double-load.\n\n### siunitx Table Alignment\n\n`siunitx` makes number-heavy tables significantly more readable:\n\n```latex\n\\begin{tabular}{l S[table-format=2.1] S[table-format=2.1] S[table-format=2.1]}\n\\toprule\nMethod & {Accuracy $\\uparrow$} & {F1 $\\uparrow$} & {Latency (ms) $\\downarrow$} \\\\\n\\midrule\nBaseline         & 85.2  & 83.7  & 45.3 \\\\\nAblation (no X)  & 87.1  & 85.4  & 42.1 \\\\\n\\textbf{Ours}    & \\textbf{92.1} & \\textbf{90.8} & \\textbf{38.7} \\\\\n\\bottomrule\n\\end{tabular}\n```\n\nThe `S` column type auto-aligns on the decimal point. Headers in `{}` escape the alignment.\n\n### Subfigures\n\nStandard pattern for side-by-side figures:\n\n```latex\n\\begin{figure}[t]\n  \\centering\n  \\begin{subfigure}[b]{0.48\\textwidth}\n    \\centering\n    \\includegraphics[width=\\textwidth]{fig_results_a.pdf}\n    \\caption{Results on Dataset A.}\n    \\label{fig:results-a}\n  \\end{subfigure}\n  \\hfill\n  \\begin{subfigure}[b]{0.48\\textwidth}\n    \\centering\n    \\includegraphics[width=\\textwidth]{fig_results_b.pdf}\n    \\caption{Results on Dataset B.}\n    \\label{fig:results-b}\n  \\end{subfigure}\n  \\caption{Comparison of our method across two datasets. (a) shows the scaling\n  behavior and (b) shows the ablation results. Both use 5 random seeds.}\n  \\label{fig:results}\n\\end{figure}\n```\n\nUse `\\cref{fig:results}` → \"Figure 1\", `\\cref{fig:results-a}` → \"Figure 1a\".\n\n### Pseudocode with algorithm2e\n\n```latex\n\\begin{algorithm}[t]\n\\caption{Iterative Refinement with Judge Panel}\n\\label{alg:method}\n\\KwIn{Task $T$, model $M$, judges $J_1 \\ldots J_n$, convergence threshold $k$}\n\\KwOut{Final output $A^*$}\n$A \\gets M(T)$ \\tcp*{Initial generation}\n$\\text{streak} \\gets 0$\\;\n\\While{$\\text{streak} \u003C k$}{\n  $C \\gets \\text{Critic}(A, T)$ \\tcp*{Identify weaknesses}\n  $B \\gets M(T, C)$ \\tcp*{Revised version addressing critique}\n  $AB \\gets \\text{Synthesize}(A, B)$ \\tcp*{Merge best elements}\n  \\ForEach{judge $J_i$}{\n    $\\text{rank}_i \\gets J_i(\\text{shuffle}(A, B, AB))$ \\tcp*{Blind ranking}\n  }\n  $\\text{winner} \\gets \\text{BordaCount}(\\text{ranks})$\\;\n  \\eIf{$\\text{winner} = A$}{\n    $\\text{streak} \\gets \\text{streak} + 1$\\;\n  }{\n    $A \\gets \\text{winner}$; $\\text{streak} \\gets 0$\\;\n  }\n}\n\\Return{$A$}\\;\n\\end{algorithm}\n```\n\n### TikZ Diagram Patterns\n\nTikZ is the standard for method diagrams in ML papers. Common patterns:\n\n**Pipeline\u002FFlow Diagram** (most common in ML papers):\n\n```latex\n\\begin{figure}[t]\n\\centering\n\\begin{tikzpicture}[\n  node distance=1.8cm,\n  box\u002F.style={rectangle, draw, rounded corners, minimum height=1cm, \n              minimum width=2cm, align=center, font=\\small},\n  arrow\u002F.style={-{Stealth[length=3mm]}, thick},\n]\n  \\node[box, fill=okcyan!20] (input) {Input\\\\$x$};\n  \\node[box, fill=okblue!20, right of=input] (encoder) {Encoder\\\\$f_\\theta$};\n  \\node[box, fill=okgreen!20, right of=encoder] (latent) {Latent\\\\$z$};\n  \\node[box, fill=okorange!20, right of=latent] (decoder) {Decoder\\\\$g_\\phi$};\n  \\node[box, fill=okred!20, right of=decoder] (output) {Output\\\\$\\hat{x}$};\n  \n  \\draw[arrow] (input) -- (encoder);\n  \\draw[arrow] (encoder) -- (latent);\n  \\draw[arrow] (latent) -- (decoder);\n  \\draw[arrow] (decoder) -- (output);\n\\end{tikzpicture}\n\\caption{Architecture overview. The encoder maps input $x$ to latent \nrepresentation $z$, which the decoder reconstructs.}\n\\label{fig:architecture}\n\\end{figure}\n```\n\n**Comparison\u002FMatrix Diagram** (for showing method variants):\n\n```latex\n\\begin{tikzpicture}[\n  cell\u002F.style={rectangle, draw, minimum width=2.5cm, minimum height=1cm, \n               align=center, font=\\small},\n  header\u002F.style={cell, fill=gray!20, font=\\small\\bfseries},\n]\n  % Headers\n  \\node[header] at (0, 0) {Method};\n  \\node[header] at (3, 0) {Converges?};\n  \\node[header] at (6, 0) {Quality?};\n  % Rows\n  \\node[cell] at (0, -1) {Single Pass};\n  \\node[cell, fill=okgreen!15] at (3, -1) {N\u002FA};\n  \\node[cell, fill=okorange!15] at (6, -1) {Baseline};\n  \\node[cell] at (0, -2) {Critique+Revise};\n  \\node[cell, fill=okred!15] at (3, -2) {No};\n  \\node[cell, fill=okred!15] at (6, -2) {Degrades};\n  \\node[cell] at (0, -3) {Ours};\n  \\node[cell, fill=okgreen!15] at (3, -3) {Yes ($k$=2)};\n  \\node[cell, fill=okgreen!15] at (6, -3) {Improves};\n\\end{tikzpicture}\n```\n\n**Iterative Loop Diagram** (for methods with feedback):\n\n```latex\n\\begin{tikzpicture}[\n  node distance=2cm,\n  box\u002F.style={rectangle, draw, rounded corners, minimum height=0.8cm, \n              minimum width=1.8cm, align=center, font=\\small},\n  arrow\u002F.style={-{Stealth[length=3mm]}, thick},\n  label\u002F.style={font=\\scriptsize, midway, above},\n]\n  \\node[box, fill=okblue!20] (gen) {Generator};\n  \\node[box, fill=okred!20, right=2.5cm of gen] (critic) {Critic};\n  \\node[box, fill=okgreen!20, below=1.5cm of $(gen)!0.5!(critic)$] (judge) {Judge Panel};\n  \n  \\draw[arrow] (gen) -- node[label] {output $A$} (critic);\n  \\draw[arrow] (critic) -- node[label, right] {critique $C$} (judge);\n  \\draw[arrow] (judge) -| node[label, left, pos=0.3] {winner} (gen);\n\\end{tikzpicture}\n```\n\n### latexdiff for Revision Tracking\n\nEssential for rebuttals — generates a marked-up PDF showing changes between versions:\n\n```bash\n# Install\n# macOS: brew install latexdiff (or comes with TeX Live)\n# Linux: sudo apt install latexdiff\n\n# Generate diff\nlatexdiff paper_v1.tex paper_v2.tex > paper_diff.tex\npdflatex paper_diff.tex\n\n# For multi-file projects (with \\input{} or \\include{})\nlatexdiff --flatten paper_v1.tex paper_v2.tex > paper_diff.tex\n```\n\nThis produces a PDF with deletions in red strikethrough and additions in blue — standard format for rebuttal supplements.\n\n### SciencePlots for matplotlib\n\nInstall and use for publication-quality plots:\n\n```bash\npip install SciencePlots\n```\n\n```python\nimport matplotlib.pyplot as plt\nimport scienceplots  # registers styles\n\n# Use science style (IEEE-like, clean)\nwith plt.style.context(['science', 'no-latex']):\n    fig, ax = plt.subplots(figsize=(3.5, 2.5))  # Single-column width\n    ax.plot(x, y, label='Ours', color='#0072B2')\n    ax.plot(x, y2, label='Baseline', color='#D55E00', linestyle='--')\n    ax.set_xlabel('Training Steps')\n    ax.set_ylabel('Accuracy')\n    ax.legend()\n    fig.savefig('paper\u002Ffig_results.pdf', bbox_inches='tight')\n\n# Available styles: 'science', 'ieee', 'nature', 'science+ieee'\n# Add 'no-latex' if LaTeX is not installed on the machine generating plots\n```\n\n**Standard figure sizes** (two-column format):\n- Single column: `figsize=(3.5, 2.5)` — fits in one column\n- Double column: `figsize=(7.0, 3.0)` — spans both columns\n- Square: `figsize=(3.5, 3.5)` — for heatmaps, confusion matrices\n\n---\n\n## Phase 6: Self-Review & Revision\n\n**Goal**: Simulate the review process before submission. Catch weaknesses early.\n\n### Step 6.1: Simulate Reviews\n\nGenerate reviews from multiple perspectives using strong models (Opus 4, Sonnet 4.6, Gemini 2.5 Pro). Use the reviewer guidelines from the target venue.\n\n**Review prompt template:**\n\n```\nYou are an expert reviewer for [VENUE]. Review this paper according to the \nofficial reviewer guidelines. Evaluate:\n\n1. Quality (technical soundness, baselines, claims supported by evidence)\n2. Clarity (writing, notation consistency, reproducibility)\n3. Significance (impact, importance of the problem)\n4. Originality (novelty, new insights)\n\nProvide:\n- Summary (2-3 sentences)\n- Strengths (bullet list)\n- Weaknesses (bullet list, most critical first)\n- Questions for authors\n- Missing references\n- Score (1-6 on NeurIPS scale)\n- Confidence (1-5)\n```\n\n### Step 6.2: Prioritize Feedback\n\nAfter collecting reviews, categorize:\n\n| Priority | Action |\n|----------|--------|\n| **Critical** (technical flaw, missing baseline) | Must fix. May require new experiments → back to Phase 2 |\n| **High** (clarity issue, missing ablation) | Should fix in this revision |\n| **Medium** (minor writing issues, extra experiments) | Fix if time allows |\n| **Low** (style preferences, tangential suggestions) | Note for future work |\n\n### Step 6.3: Revision Cycle\n\nFor each critical\u002Fhigh issue:\n1. Identify the specific section(s) affected\n2. Draft the fix\n3. Verify the fix doesn't break other claims\n4. Update the paper\n5. Re-check against the reviewer's concern\n\n### Step 6.4: Rebuttal Writing\n\nWhen responding to actual reviews (post-submission), rebuttals are a distinct skill from revision:\n\n**Format**: Point-by-point. For each reviewer concern:\n```\n> R1-W1: \"The paper lacks comparison with Method X.\"\n\nWe thank the reviewer for this suggestion. We have added a comparison with \nMethod X in Table 3 (revised). Our method outperforms X by 3.2pp on [metric] \n(p\u003C0.05). We note that X requires 2x our compute budget.\n```\n\n**Rules**:\n- Address every concern — reviewers notice if you skip one\n- Lead with the strongest responses\n- Be concise and direct — reviewers read dozens of rebuttals\n- Include new results if you ran experiments during the rebuttal period\n- Never be defensive or dismissive, even of weak criticisms\n- Use `latexdiff` to generate a marked-up PDF showing changes (see Professional LaTeX Tooling section)\n- Thank reviewers for specific, actionable feedback (not generic praise)\n\n**What NOT to do**: \"We respectfully disagree\" without evidence. \"This is out of scope\" without explanation. Ignoring a weakness by only responding to strengths.\n\n### Step 6.5: Paper Evolution Tracking\n\nSave snapshots at key milestones:\n```\npaper\u002F\n  paper.tex                    # Current working version\n  paper_v1_first_draft.tex     # First complete draft\n  paper_v2_post_review.tex     # After simulated review\n  paper_v3_pre_submission.tex  # Final before submission\n  paper_v4_camera_ready.tex    # Post-acceptance final\n```\n\n---\n\n## Phase 7: Submission Preparation\n\n**Goal**: Final checks, formatting, and submission.\n\n### Step 7.1: Conference Checklist\n\nEvery venue has mandatory checklists. Complete them carefully — incomplete checklists can result in desk rejection.\n\nSee [references\u002Fchecklists.md](references\u002Fchecklists.md) for:\n- NeurIPS 16-item paper checklist\n- ICML broader impact + reproducibility\n- ICLR LLM disclosure policy\n- ACL mandatory limitations section\n- Universal pre-submission checklist\n\n### Step 7.2: Anonymization Checklist\n\nDouble-blind review means reviewers cannot know who wrote the paper. Check ALL of these:\n\n```\nAnonymization Checklist:\n- [ ] No author names or affiliations anywhere in the PDF\n- [ ] No acknowledgments section (add after acceptance)\n- [ ] Self-citations written in third person: \"Smith et al. [1] showed...\" not \"We previously showed [1]...\"\n- [ ] No GitHub\u002FGitLab URLs pointing to your personal repos\n- [ ] Use Anonymous GitHub (https:\u002F\u002Fanonymous.4open.science\u002F) for code links\n- [ ] No institutional logos or identifiers in figures\n- [ ] No file metadata containing author names (check PDF properties)\n- [ ] No \"our previous work\" or \"in our earlier paper\" phrasing\n- [ ] Dataset names don't reveal institution (rename if needed)\n- [ ] Supplementary materials don't contain identifying information\n```\n\n**Common mistakes**: Git commit messages visible in supplementary code, watermarked figures from institutional tools, acknowledgments left in from a previous draft, arXiv preprint posted before anonymity period.\n\n### Step 7.3: Formatting Verification\n\n```\nPre-Submission Format Check:\n- [ ] Page limit respected (excluding references and appendix)\n- [ ] All figures are vector (PDF) or high-res raster (600 DPI PNG)\n- [ ] All figures readable in grayscale\n- [ ] All tables use booktabs\n- [ ] References compile correctly (no \"?\" in citations)\n- [ ] No overfull hboxes in critical areas\n- [ ] Appendix clearly labeled and separated\n- [ ] Required sections present (limitations, broader impact, etc.)\n```\n\n### Step 7.3: Final Compilation\n\n```bash\n# Clean build\nrm -f *.aux *.bbl *.blg *.log *.out *.pdf\nlatexmk -pdf main.tex\n\n# Or manual\npdflatex main.tex\nbibtex main\npdflatex main.tex\npdflatex main.tex\n```\n\n### Step 7.4: Conference-Specific Requirements\n\n| Venue | Special Requirements |\n|-------|---------------------|\n| **NeurIPS** | Paper checklist in appendix, lay summary if accepted |\n| **ICML** | Broader Impact Statement (after conclusion, doesn't count toward limit) |\n| **ICLR** | LLM disclosure required, reciprocal reviewing agreement |\n| **ACL** | Mandatory Limitations section, Responsible NLP checklist |\n| **AAAI** | Strict style file — no modifications whatsoever |\n| **COLM** | Frame contribution for language model community |\n\n### Step 7.6: Conference Resubmission & Format Conversion\n\nWhen converting between venues, **never copy LaTeX preambles between templates**:\n\n```bash\n# 1. Start fresh with target template\ncp -r templates\u002Ficml2026\u002F new_submission\u002F\n\n# 2. Copy ONLY content sections (not preamble)\n#    - Abstract text, section content, figures, tables, bib entries\n\n# 3. Adjust for page limits\n# 4. Add venue-specific required sections\n# 5. Update references\n```\n\n| From → To | Page Change | Key Adjustments |\n|-----------|-------------|-----------------|\n| NeurIPS → ICML | 9 → 8 | Cut 1 page, add Broader Impact |\n| ICML → ICLR | 8 → 9 | Expand experiments, add LLM disclosure |\n| NeurIPS → ACL | 9 → 8 | Restructure for NLP conventions, add Limitations |\n| ICLR → AAAI | 9 → 7 | Significant cuts, strict style adherence |\n| Any → COLM | varies → 9 | Reframe for language model focus |\n\nWhen cutting pages: move proofs to appendix, condense related work, combine tables, use subfigures.\nWhen expanding: add ablations, expand limitations, include additional baselines, add qualitative examples.\n\n**After rejection**: Address reviewer concerns in the new version, but don't include a \"changes\" section or reference the previous submission (blind review).\n\n### Step 7.7: Camera-Ready Preparation (Post-Acceptance)\n\nAfter acceptance, prepare the camera-ready version:\n\n```\nCamera-Ready Checklist:\n- [ ] De-anonymize: add author names, affiliations, email addresses\n- [ ] Add Acknowledgments section (funding, compute grants, helpful reviewers)\n- [ ] Add public code\u002Fdata URL (real GitHub, not anonymous)\n- [ ] Address any mandatory revisions from meta-reviewer\n- [ ] Switch template to camera-ready mode (if applicable — e.g., AAAI \\anon → \\camera)\n- [ ] Add copyright notice if required by venue\n- [ ] Update any \"anonymous\" placeholders in text\n- [ ] Verify final PDF compiles cleanly\n- [ ] Check page limit for camera-ready (sometimes differs from submission)\n- [ ] Upload supplementary materials (code, data, appendix) to venue portal\n```\n\n---\n\n## Hermes Agent Integration\n\nThis skill is designed for the Hermes agent. It uses Hermes tools, delegation, scheduling, and memory for the full research lifecycle.\n\n### Related Skills\n\nCompose this skill with other Hermes skills for specific phases:\n\n| Skill | When to Use | How to Load |\n|-------|-------------|-------------|\n| **arxiv** | Phase 1 (Literature Review): searching arXiv, generating BibTeX, finding related papers via Semantic Scholar | `skill_view(\"arxiv\")` |\n| **subagent-driven-development** | Phase 5 (Drafting): parallel section writing with 2-stage review (spec compliance then quality) | `skill_view(\"subagent-driven-development\")` |\n| **plan** | Phase 0 (Setup): creating structured plans before execution. Writes to `.hermes\u002Fplans\u002F` | `skill_view(\"plan\")` |\n| **qmd** | Phase 1 (Literature): searching local knowledge bases (notes, transcripts, docs) via hybrid BM25+vector search | Install: `skill_manage(\"install\", \"qmd\")` |\n| **diagramming** | Phase 4-5: creating Excalidraw-based figures and architecture diagrams | `skill_view(\"diagramming\")` |\n| **data-science** | Phase 4 (Analysis): Jupyter live kernel for interactive analysis and visualization | `skill_view(\"data-science\")` |\n\n**This skill supersedes `ml-paper-writing`** — it contains all of ml-paper-writing's content plus the full experiment\u002Fanalysis pipeline and autoreason methodology.\n\n### Hermes Tools Reference\n\n| Tool | Usage in This Pipeline |\n|------|----------------------|\n| **`terminal`** | LaTeX compilation (`latexmk -pdf`), git operations, launching experiments (`nohup python run.py &`), process checks |\n| **`process`** | Background experiment management: `process(\"start\", ...)`, `process(\"poll\", pid)`, `process(\"log\", pid)`, `process(\"kill\", pid)` |\n| **`execute_code`** | Run Python for citation verification, statistical analysis, data aggregation. Has tool access via RPC. |\n| **`read_file`** \u002F **`write_file`** \u002F **`patch`** | Paper editing, experiment scripts, result files. Use `patch` for targeted edits to large .tex files. |\n| **`web_search`** | Literature discovery: `web_search(\"transformer attention mechanism 2024\")` |\n| **`web_extract`** | Fetch paper content, verify citations: `web_extract(\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.17651\")` |\n| **`delegate_task`** | **Parallel section drafting** — spawn isolated subagents for each section. Also for concurrent citation verification. |\n| **`todo`** | Primary state tracker across sessions. Update after every phase transition. |\n| **`memory`** | Persist key decisions across sessions: contribution framing, venue choice, reviewer feedback. |\n| **`cronjob`** | Schedule experiment monitoring, deadline countdowns, automated arXiv checks. |\n| **`clarify`** | Ask the user targeted questions when blocked (venue choice, contribution framing). |\n| **`send_message`** | Notify user when experiments complete or drafts are ready, even if user isn't in chat. |\n\n### Tool Usage Patterns\n\n**Experiment monitoring** (most common):\n```\nterminal(\"ps aux | grep \u003Cpattern>\")\n→ terminal(\"tail -30 \u003Clogfile>\")\n→ terminal(\"ls results\u002F\")\n→ execute_code(\"analyze results JSON, compute metrics\")\n→ terminal(\"git add -A && git commit -m '\u003Cdescriptive message>' && git push\")\n→ send_message(\"Experiment complete: \u003Csummary>\")\n```\n\n**Parallel section drafting** (using delegation):\n```\ndelegate_task(\"Draft the Methods section based on these experiment scripts and configs. \n  Include: pseudocode, all hyperparameters, architectural details sufficient for \n  reproduction. Write in LaTeX using the neurips2025 template conventions.\")\n\ndelegate_task(\"Draft the Related Work section. Use web_search and web_extract to \n  find papers. Verify every citation via Semantic Scholar. Group by methodology.\")\n\ndelegate_task(\"Draft the Experiments section. Read all result files in results\u002F. \n  State which claim each experiment supports. Include error bars and significance.\")\n```\n\nEach delegate runs as a **fresh subagent** with no shared context — provide all necessary information in the prompt. Collect outputs and integrate.\n\n**Citation verification** (using execute_code):\n```python\n# In execute_code:\nfrom semanticscholar import SemanticScholar\nimport requests\n\nsch = SemanticScholar()\nresults = sch.search_paper(\"attention mechanism transformers\", limit=5)\nfor paper in results:\n    doi = paper.externalIds.get('DOI', 'N\u002FA')\n    if doi != 'N\u002FA':\n        bibtex = requests.get(f\"https:\u002F\u002Fdoi.org\u002F{doi}\", \n                              headers={\"Accept\": \"application\u002Fx-bibtex\"}).text\n        print(bibtex)\n```\n\n### State Management with `memory` and `todo`\n\n**`memory` tool** — persist key decisions (bounded: ~2200 chars for MEMORY.md):\n\n```\nmemory(\"add\", \"Paper: autoreason. Venue: NeurIPS 2025 (9 pages). \n  Contribution: structured refinement works when generation-evaluation gap is wide.\n  Key results: Haiku 42\u002F42, Sonnet 3\u002F5, S4.6 constrained 2\u002F3.\n  Status: Phase 5 — drafting Methods section.\")\n```\n\nUpdate memory after major decisions or phase transitions. This persists across sessions.\n\n**`todo` tool** — track granular progress:\n\n```\ntodo(\"add\", \"Design constrained task experiments for Sonnet 4.6\")\ntodo(\"add\", \"Run Haiku baseline comparison\")\ntodo(\"add\", \"Draft Methods section\")\ntodo(\"update\", id=3, status=\"in_progress\")\ntodo(\"update\", id=1, status=\"completed\")\n```\n\n**Session startup protocol:**\n```\n1. todo(\"list\")                           # Check current task list\n2. memory(\"read\")                         # Recall key decisions\n3. terminal(\"git log --oneline -10\")      # Check recent commits\n4. terminal(\"ps aux | grep python\")       # Check running experiments\n5. terminal(\"ls results\u002F | tail -20\")     # Check for new results\n6. Report status to user, ask for direction\n```\n\n### Cron Monitoring with `cronjob`\n\nUse the `cronjob` tool to schedule periodic experiment checks:\n\n```\ncronjob(\"create\", {\n  \"schedule\": \"*\u002F30 * * * *\",  # Every 30 minutes\n  \"prompt\": \"Check experiment status:\n    1. ps aux | grep run_experiment\n    2. tail -30 logs\u002Fexperiment_haiku.log\n    3. ls results\u002Fhaiku_baselines\u002F\n    4. If complete: read results, compute Borda scores, \n       git add -A && git commit -m 'Add Haiku results' && git push\n    5. Report: table of results, key finding, next step\n    6. If nothing changed: respond with [SILENT]\"\n})\n```\n\n**[SILENT] protocol**: When nothing has changed since the last check, respond with exactly `[SILENT]`. This suppresses notification delivery to the user. Only report when there are genuine changes worth knowing about.\n\n**Deadline tracking**:\n```\ncronjob(\"create\", {\n  \"schedule\": \"0 9 * * *\",  # Daily at 9am\n  \"prompt\": \"NeurIPS 2025 deadline: May 22. Today is {date}. \n    Days remaining: {compute}. \n    Check todo list — are we on track? \n    If \u003C7 days: warn user about remaining tasks.\"\n})\n```\n\n### Communication Patterns\n\n**When to notify the user** (via `send_message` or direct response):\n- Experiment batch completed (with results table)\n- Unexpected finding or failure requiring decision\n- Draft section ready for review\n- Deadline approaching with incomplete tasks\n\n**When NOT to notify:**\n- Experiment still running, no new results → `[SILENT]`\n- Routine monitoring with no changes → `[SILENT]`\n- Intermediate steps that don't need attention\n\n**Report format** — always include structured data:\n```\n## Experiment: \u003Cname>\nStatus: Complete \u002F Running \u002F Failed\n\n| Task | Method A | Method B | Method C |\n|------|---------|---------|---------|\n| Task 1 | 85.2 | 82.1 | **89.4** |\n\nKey finding: \u003Cone sentence>\nNext step: \u003Cwhat happens next>\n```\n\n### Decision Points Requiring Human Input\n\nUse `clarify` for targeted questions when genuinely blocked:\n\n| Decision | When to Ask |\n|----------|-------------|\n| Target venue | Before starting paper (affects page limits, framing) |\n| Contribution framing | When multiple valid framings exist |\n| Experiment priority | When TODO list has more experiments than time allows |\n| Submission readiness | Before final submission |\n\n**Do NOT ask about** (be proactive, make a choice, flag it):\n- Word choice, section ordering\n- Which specific results to highlight\n- Citation completeness (draft with what you find, note gaps)\n\n---\n\n## Reviewer Evaluation Criteria\n\nUnderstanding what reviewers look for helps focus effort:\n\n| Criterion | What They Check |\n|-----------|----------------|\n| **Quality** | Technical soundness, well-supported claims, fair baselines |\n| **Clarity** | Clear writing, reproducible by experts, consistent notation |\n| **Significance** | Community impact, advances understanding |\n| **Originality** | New insights (doesn't require new method) |\n\n**Scoring (NeurIPS 6-point scale):**\n- 6: Strong Accept — groundbreaking, flawless\n- 5: Accept — technically solid, high impact\n- 4: Borderline Accept — solid, limited evaluation\n- 3: Borderline Reject — weaknesses outweigh\n- 2: Reject — technical flaws\n- 1: Strong Reject — known results or ethics issues\n\nSee [references\u002Freviewer-guidelines.md](references\u002Freviewer-guidelines.md) for detailed guidelines, common concerns, and rebuttal strategies.\n\n---\n\n## Common Issues and Solutions\n\n| Issue | Solution |\n|-------|----------|\n| Abstract too generic | Delete first sentence if it could prepend any ML paper. Start with your specific contribution. |\n| Introduction exceeds 1.5 pages | Split background into Related Work. Front-load contribution bullets. |\n| Experiments lack explicit claims | Add: \"This experiment tests whether [specific claim]...\" before each one. |\n| Reviewers find paper hard to follow | Add signposting, use consistent terminology, make figure captions self-contained. |\n| Missing statistical significance | Add error bars, number of runs, statistical tests, confidence intervals. |\n| Scope creep in experiments | Every experiment must map to a specific claim. Cut experiments that don't. |\n| Paper rejected, need to resubmit | See Conference Resubmission in Phase 7. Address reviewer concerns without referencing reviews. |\n\n---\n\n## Reference Documents\n\n| Document | Contents |\n|----------|----------|\n| [references\u002Fwriting-guide.md](references\u002Fwriting-guide.md) | Gopen & Swan 7 principles, Perez micro-tips, Lipton word choice, Steinhardt precision, figure design |\n| [references\u002Fcitation-workflow.md](references\u002Fcitation-workflow.md) | Citation APIs, Python code, CitationManager class, BibTeX management |\n| [references\u002Fchecklists.md](references\u002Fchecklists.md) | NeurIPS 16-item, ICML, ICLR, ACL requirements, universal pre-submission checklist |\n| [references\u002Freviewer-guidelines.md](references\u002Freviewer-guidelines.md) | Evaluation criteria, scoring, common concerns, rebuttal template |\n| [references\u002Fsources.md](references\u002Fsources.md) | Complete bibliography of all writing guides, conference guidelines, APIs |\n| [references\u002Fexperiment-patterns.md](references\u002Fexperiment-patterns.md) | Experiment design patterns, evaluation protocols, monitoring, error recovery |\n| [references\u002Fautoreason-methodology.md](references\u002Fautoreason-methodology.md) | Autoreason loop, strategy selection, model guide, prompts, scope constraints, Borda scoring |\n\n### LaTeX Templates\n\nTemplates in `templates\u002F` for: **NeurIPS 2025**, **ICML 2026**, **ICLR 2026**, **ACL**, **AAAI 2026**, **COLM 2025**.\n\nSee [templates\u002FREADME.md](templates\u002FREADME.md) for compilation instructions.\n\n### Key External Sources\n\n**Writing Philosophy:**\n- [Neel Nanda: How to Write ML Papers](https:\u002F\u002Fwww.alignmentforum.org\u002Fposts\u002FeJGptPbbFPZGLpjsp\u002Fhighly-opinionated-advice-on-how-to-write-ml-papers)\n- [Sebastian Farquhar: How to Write ML Papers](https:\u002F\u002Fsebastianfarquhar.com\u002Fon-research\u002F2024\u002F11\u002F04\u002Fhow_to_write_ml_papers\u002F)\n- [Gopen & Swan: Science of Scientific Writing](https:\u002F\u002Fcseweb.ucsd.edu\u002F~swanson\u002Fpapers\u002Fscience-of-writing.pdf)\n- [Lipton: Heuristics for Scientific Writing](https:\u002F\u002Fwww.approximatelycorrect.com\u002F2018\u002F01\u002F29\u002Fheuristics-technical-scientific-writing-machine-learning-perspective\u002F)\n- [Perez: Easy Paper Writing Tips](https:\u002F\u002Fethanperez.net\u002Feasy-paper-writing-tips\u002F)\n\n**APIs:** [Semantic Scholar](https:\u002F\u002Fapi.semanticscholar.org\u002Fapi-docs\u002F) | [CrossRef](https:\u002F\u002Fwww.crossref.org\u002Fdocumentation\u002Fretrieve-metadata\u002Frest-api\u002F) | [arXiv](https:\u002F\u002Finfo.arxiv.org\u002Fhelp\u002Fapi\u002Fbasics.html)\n\n**Venues:** [NeurIPS](https:\u002F\u002Fneurips.cc\u002FConferences\u002F2025\u002FPaperInformation\u002FStyleFiles) | [ICML](https:\u002F\u002Ficml.cc\u002FConferences\u002F2025\u002FAuthorInstructions) | [ICLR](https:\u002F\u002Ficlr.cc\u002FConferences\u002F2026\u002FAuthorGuide) | [ACL](https:\u002F\u002Fgithub.com\u002Facl-org\u002Facl-style-files)\n",{"name":5,"title":59,"author":60,"license":14,"version":61,"metadata":62,"platforms":82,"description":7,"dependencies":84},"Research Paper Writing Pipeline","Orchestra Research","1.0.0",{"hermes":63},{"tags":64,"category":35,"related_skills":77,"requires_toolsets":81},[65,37,66,67,68,48,69,70,71,72,73,74,75,76],"Research","Experiments","ML","AI","ICML","ICLR","ACL","AAAI","COLM","LaTeX","Citations","Statistical Analysis",[78,79,80,47],"arxiv","ml-paper-writing","subagent-driven-development",[52,45],[83,53],"linux",[85,78,86,87,43,44,88,89],"semanticscholar","habanero","requests","matplotlib","SciencePlots","main","a05c0ccfff1cda2a5d12e9ee7d02b75350ac0557","packages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit","https:\u002F\u002Fgithub.com\u002Fqubit-calculus","2026-04-07 15:16:13",{"source":8,"skillName":5,"commands":97,"prompt":107,"zipAvailable":108},[98,101,104],{"tool":99,"command":100},"npx","npx skills add qubit-calculus\u002FQubit --skill research-paper-writing",{"tool":102,"command":103},"bunx","bunx skills add qubit-calculus\u002FQubit --skill research-paper-writing",{"tool":105,"command":106},"pnpm","pnpm dlx skills add qubit-calculus\u002FQubit --skill research-paper-writing","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002FSKILL.md",true,[110,116,120,124,128,132,136,140,146,150,154,158,162,166,170,174,178,182,186,190,194,198,202,206,212,216,220,224,228,232,236,239,243,246,250,254,258,261,264,268,272,276,280,284,288,292,296,300,304,308,312,316],{"path":111,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":114,"sourceHtmlUrl":115},"references\u002Fautoreason-methodology.md","REFERENCE","参考文档",19146,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fautoreason-methodology.md",{"path":117,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":118,"sourceHtmlUrl":119},"references\u002Fchecklists.md",13622,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fchecklists.md",{"path":121,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":122,"sourceHtmlUrl":123},"references\u002Fcitation-workflow.md",15433,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fcitation-workflow.md",{"path":125,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":126,"sourceHtmlUrl":127},"references\u002Fexperiment-patterns.md",25742,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fexperiment-patterns.md",{"path":129,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":130,"sourceHtmlUrl":131},"references\u002Freviewer-guidelines.md",13738,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Freviewer-guidelines.md",{"path":133,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":134,"sourceHtmlUrl":135},"references\u002Fsources.md",7310,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fsources.md",{"path":137,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":138,"sourceHtmlUrl":139},"references\u002Fwriting-guide.md",16292,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Freferences\u002Fwriting-guide.md",{"path":141,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":144,"sourceHtmlUrl":145},"templates\u002Faaai2026\u002Faaai2026-unified-supp.tex","OTHER","其他",4548,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002Faaai2026-unified-supp.tex",{"path":147,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":148,"sourceHtmlUrl":149},"templates\u002Faaai2026\u002Faaai2026-unified-template.tex",63140,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002Faaai2026-unified-template.tex",{"path":151,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":152,"sourceHtmlUrl":153},"templates\u002Faaai2026\u002Faaai2026.bib",4766,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002Faaai2026.bib",{"path":155,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":156,"sourceHtmlUrl":157},"templates\u002Faaai2026\u002Faaai2026.bst",30207,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002Faaai2026.bst",{"path":159,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":160,"sourceHtmlUrl":161},"templates\u002Faaai2026\u002Faaai2026.sty",11802,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002Faaai2026.sty",{"path":163,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":164,"sourceHtmlUrl":165},"templates\u002Faaai2026\u002FREADME.md",17987,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Faaai2026\u002FREADME.md",{"path":167,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":168,"sourceHtmlUrl":169},"templates\u002Facl\u002Facl_latex.tex",14533,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Facl_latex.tex",{"path":171,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":172,"sourceHtmlUrl":173},"templates\u002Facl\u002Facl_lualatex.tex",3050,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Facl_lualatex.tex",{"path":175,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":176,"sourceHtmlUrl":177},"templates\u002Facl\u002Facl_natbib.bst",45186,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Facl_natbib.bst",{"path":179,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":180,"sourceHtmlUrl":181},"templates\u002Facl\u002Facl.sty",11615,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Facl.sty",{"path":183,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":184,"sourceHtmlUrl":185},"templates\u002Facl\u002Fanthology.bib.txt",1169,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Fanthology.bib.txt",{"path":187,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":188,"sourceHtmlUrl":189},"templates\u002Facl\u002Fcustom.bib",2071,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Fcustom.bib",{"path":191,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":192,"sourceHtmlUrl":193},"templates\u002Facl\u002Fformatting.md",17923,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002Fformatting.md",{"path":195,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":196,"sourceHtmlUrl":197},"templates\u002Facl\u002FREADME.md",2126,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Facl\u002FREADME.md",{"path":199,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":200,"sourceHtmlUrl":201},"templates\u002Fcolm2025\u002Fcolm2025_conference.bib",496,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fcolm2025_conference.bib",{"path":203,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":204,"sourceHtmlUrl":205},"templates\u002Fcolm2025\u002Fcolm2025_conference.bst",26973,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fcolm2025_conference.bst",{"path":207,"resourceType":208,"resourceTypeDesc":209,"sizeBytes":210,"sourceHtmlUrl":211},"templates\u002Fcolm2025\u002Fcolm2025_conference.pdf","ASSET","资产文件",122635,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fcolm2025_conference.pdf",{"path":213,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":214,"sourceHtmlUrl":215},"templates\u002Fcolm2025\u002Fcolm2025_conference.sty",7727,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fcolm2025_conference.sty",{"path":217,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":218,"sourceHtmlUrl":219},"templates\u002Fcolm2025\u002Fcolm2025_conference.tex",12830,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fcolm2025_conference.tex",{"path":221,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":222,"sourceHtmlUrl":223},"templates\u002Fcolm2025\u002Ffancyhdr.sty",20521,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Ffancyhdr.sty",{"path":225,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":226,"sourceHtmlUrl":227},"templates\u002Fcolm2025\u002Fmath_commands.tex",12284,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fmath_commands.tex",{"path":229,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":230,"sourceHtmlUrl":231},"templates\u002Fcolm2025\u002Fnatbib.sty",45154,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002Fnatbib.sty",{"path":233,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":234,"sourceHtmlUrl":235},"templates\u002Fcolm2025\u002FREADME.md",51,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fcolm2025\u002FREADME.md",{"path":237,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":222,"sourceHtmlUrl":238},"templates\u002Ficlr2026\u002Ffancyhdr.sty","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ffancyhdr.sty",{"path":240,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":241,"sourceHtmlUrl":242},"templates\u002Ficlr2026\u002Ficlr2026_conference.bib",629,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ficlr2026_conference.bib",{"path":244,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":204,"sourceHtmlUrl":245},"templates\u002Ficlr2026\u002Ficlr2026_conference.bst","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ficlr2026_conference.bst",{"path":247,"resourceType":208,"resourceTypeDesc":209,"sizeBytes":248,"sourceHtmlUrl":249},"templates\u002Ficlr2026\u002Ficlr2026_conference.pdf",200508,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ficlr2026_conference.pdf",{"path":251,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":252,"sourceHtmlUrl":253},"templates\u002Ficlr2026\u002Ficlr2026_conference.sty",9025,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ficlr2026_conference.sty",{"path":255,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":256,"sourceHtmlUrl":257},"templates\u002Ficlr2026\u002Ficlr2026_conference.tex",16899,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Ficlr2026_conference.tex",{"path":259,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":226,"sourceHtmlUrl":260},"templates\u002Ficlr2026\u002Fmath_commands.tex","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Fmath_commands.tex",{"path":262,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":230,"sourceHtmlUrl":263},"templates\u002Ficlr2026\u002Fnatbib.sty","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficlr2026\u002Fnatbib.sty",{"path":265,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":266,"sourceHtmlUrl":267},"templates\u002Ficml2026\u002Falgorithm.sty",2223,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Falgorithm.sty",{"path":269,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":270,"sourceHtmlUrl":271},"templates\u002Ficml2026\u002Falgorithmic.sty",7414,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Falgorithmic.sty",{"path":273,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":274,"sourceHtmlUrl":275},"templates\u002Ficml2026\u002Fexample_paper.bib",2051,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Fexample_paper.bib",{"path":277,"resourceType":208,"resourceTypeDesc":209,"sizeBytes":278,"sourceHtmlUrl":279},"templates\u002Ficml2026\u002Fexample_paper.pdf",193509,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Fexample_paper.pdf",{"path":281,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":282,"sourceHtmlUrl":283},"templates\u002Ficml2026\u002Fexample_paper.tex",29714,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Fexample_paper.tex",{"path":285,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":286,"sourceHtmlUrl":287},"templates\u002Ficml2026\u002Ffancyhdr.sty",31715,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Ffancyhdr.sty",{"path":289,"resourceType":208,"resourceTypeDesc":209,"sizeBytes":290,"sourceHtmlUrl":291},"templates\u002Ficml2026\u002Ficml_numpapers.pdf",2823,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Ficml_numpapers.pdf",{"path":293,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":294,"sourceHtmlUrl":295},"templates\u002Ficml2026\u002Ficml2026.bst",27147,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Ficml2026.bst",{"path":297,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":298,"sourceHtmlUrl":299},"templates\u002Ficml2026\u002Ficml2026.sty",27344,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Ficml2026\u002Ficml2026.sty",{"path":301,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":302,"sourceHtmlUrl":303},"templates\u002Fneurips2025\u002Fextra_pkgs.tex",2837,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fneurips2025\u002Fextra_pkgs.tex",{"path":305,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":306,"sourceHtmlUrl":307},"templates\u002Fneurips2025\u002Fmain.tex",574,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fneurips2025\u002Fmain.tex",{"path":309,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":310,"sourceHtmlUrl":311},"templates\u002Fneurips2025\u002FMakefile",1054,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fneurips2025\u002FMakefile",{"path":313,"resourceType":142,"resourceTypeDesc":143,"sizeBytes":314,"sourceHtmlUrl":315},"templates\u002Fneurips2025\u002Fneurips.sty",11625,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002Fneurips2025\u002Fneurips.sty",{"path":317,"resourceType":112,"resourceTypeDesc":113,"sizeBytes":318,"sourceHtmlUrl":319},"templates\u002FREADME.md",6707,"https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fresearch\u002Fresearch-paper-writing\u002Ftemplates\u002FREADME.md",[321,350,378,395,417,439],{"skillNo":322,"name":323,"nameZh":6,"description":324,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":325,"categories":326,"keywords":333,"publishedAt":54},"01M6NT8MKBM6Q660B4D7SHXE3F","writing-plans","Use when you have a spec or requirements for a multi-step task. Creates comprehensive implementation plans with bite-sized tasks, exact file paths, and complete code examples.","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Fwriting-plans\u002FSKILL.md",[327,328,329,332],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":330,"categoryName":331,"taxonomyType":31,"taxonomyTypeDesc":32},"01JBSEEDCAT000000000000030","开发工程师",{"categoryNo":29,"categoryName":30,"taxonomyType":31,"taxonomyTypeDesc":32},[34,334,335,40,336,337,338,339,340,341,342,343,46,344,345,346,347,348,349,41],"writing plans","plans","review","workflow","documentation","tasks","development","code","planning","code-review","use","superpowers","complete","paths","test-driven-development","you",{"skillNo":351,"name":348,"nameZh":6,"description":352,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":353,"categories":354,"keywords":363,"publishedAt":54},"01M6NT8G9H510DV7VMY4X4Q9F3","Use when implementing any feature or bugfix, before writing implementation code. Enforces RED-GREEN-REFACTOR cycle with test-first approach.","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Ftest-driven-development\u002FSKILL.md",[355,358,361,362],{"categoryNo":356,"categoryName":357,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000007","代码审查",{"categoryNo":359,"categoryName":360,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000008","测试与质量",{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":29,"categoryName":30,"taxonomyType":31,"taxonomyTypeDesc":32},[364,365,340,366,367,34,368,341,369,46,345,370,371,372,373,335,374,375,376,377],"test","test driven development","driven","testing","debugging","refactor","red-green-refactor","cycle","green","feature","systematic","bugfix","implementing","systematic-debugging",{"skillNo":379,"name":377,"nameZh":6,"description":380,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":381,"categories":382,"keywords":384,"publishedAt":54},"01M6NT87T34SVN9P1S3PCBWY2D","Use when encountering any bug, test failure, or unexpected behavior. 4-phase root cause investigation — NO fixes without understanding the problem first.","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Fsystematic-debugging\u002FSKILL.md",[383],{"categoryNo":359,"categoryName":360,"taxonomyType":20,"taxonomyTypeDesc":21},[385,368,374,34,364,340,386,387,46,388,389,344,345,390,391,392,335,393,348,394],"systematic debugging","first","investigation","hermes","behavior","bug","root-cause","problem-solving","problem","fixes",{"skillNo":396,"name":80,"nameZh":6,"description":397,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":398,"categories":399,"keywords":409,"publishedAt":54},"01M6NT83GMFJRDY84BWJHQSVQE","Use when executing implementation plans with independent tasks. Dispatches fresh delegate_task per task with two-stage review (spec compliance then code quality).","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Fsubagent-driven-development\u002FSKILL.md",[400,401,404,405,406],{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":402,"categoryName":403,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":330,"categoryName":331,"taxonomyType":31,"taxonomyTypeDesc":32},{"categoryNo":29,"categoryName":30,"taxonomyType":31,"taxonomyTypeDesc":32},{"categoryNo":407,"categoryName":408,"taxonomyType":31,"taxonomyTypeDesc":32},"01JBSEEDCAT000000000000037","律师与法务",[340,410,46,366,34,336,337,411,339,341,412,343,413,414,344,345,415,335,416,348],"subagent driven development","code quality","delegation","parallel","delegate","stage","fresh",{"skillNo":418,"name":419,"nameZh":6,"description":420,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":421,"categories":422,"keywords":425,"publishedAt":54},"01M6NT7YMC7H5NK632ED23WMV6","requesting-code-review","Pre-commit verification pipeline — static security scan, baseline-aware quality gates, independent reviewer subagent, and auto-fix loop. Use after code changes and before committing, pushing, or opening a PR.","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Frequesting-code-review\u002FSKILL.md",[423,424],{"categoryNo":356,"categoryName":357,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":330,"categoryName":331,"taxonomyType":31,"taxonomyTypeDesc":32},[426,336,341,427,34,428,429,343,430,431,432,433,46,434,435,436,345,348,437,438],"code review","requesting code review","security","commit","verification","security scan","pipeline","loop","reviewer","fix","auto","static","auto-fix",{"skillNo":440,"name":47,"nameZh":6,"description":441,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":442,"categories":443,"keywords":444,"publishedAt":54},"01M6NT7AWK0VW1RFSJTAG92363","Plan mode for Hermes — inspect context, write a markdown plan into the active workspace's `.hermes\u002Fplans\u002F` directory, and do not execute the work.","https:\u002F\u002Fgithub.com\u002Fqubit-calculus\u002FQubit\u002Fblob\u002Fmain\u002Fpackages\u002Fskills\u002Fhermes\u002Fsoftware-development\u002Fplan\u002FSKILL.md",[],[47,34,445,337,340,342,446,46,447,448,449,450,451,452,335,453,454,455,41,456],"directory","markdown","workspace","work","write","implementation","do","active","plan mode","inspect","mode","plan-mode"]