[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6MFGPK1J340JYVZ0T1ZR58Y":3,"skill-same:01M6MFGPK1J340JYVZ0T1ZR58Y":111},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":34,"publishedAt":55,"linkOnly":56,"markdownContent":57,"markdownBody":58,"frontMatter":59,"refName":90,"commitSha":91,"skillDirectory":92,"forksCount":93,"repositoryHtmlUrl":94,"repositoryHomepageUrl":6,"ownerHtmlUrl":95,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":96,"install":97,"resources":110},"01M6MFGPK1J340JYVZ0T1ZR58Y","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.","ibragimov-oasis\u002Fvibe-coder","ibragimov-oasis","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F240665573?v=4","01M6DF78Y45R8228FHAZ7G6JXM","IBRAGIMOV MUKHAMMADJON","01M6K5ZYR8TDPAFG4R6693E2A6",3,"MIT","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002FREPO_DOCS\u002F02-hermes-agent\u002Fskills\u002Fresearch\u002Fresearch-paper-writing\u002FSKILL.md",[18,23,26,29],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000015","视觉设计","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":27,"categoryName":28,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000021","资料检索与研究",{"categoryNo":30,"categoryName":31,"taxonomyType":32,"taxonomyTypeDesc":33},"01JBSEEDCAT000000000000036","内容创作者","PROFESSION","适用职业",[35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54],"writing","paper","research paper writing","Paper Writing","research","experiments","experiment design","monitoring","NeurIPS","ICML","terminal","plan","subagent","files","macos","numpy","scipy","citation verification","design","for","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\n## ? Связи\n\n- [[000 - Map of Maps]] — REPO_DOCS\n- [[000 - Map of Maps]] — Map of Maps\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\n## ? Связи\n\n- [[000 - Map of Maps]] — REPO_DOCS\n- [[000 - Map of Maps]] — Map of Maps\n\n",{"name":5,"title":60,"author":61,"license":15,"version":62,"metadata":63,"platforms":82,"description":7,"dependencies":84},"Research Paper Writing Pipeline","Orchestra Research","1.0.0",{"hermes":64},{"tags":65,"category":39,"related_skills":77,"requires_toolsets":81},[66,38,67,68,69,43,44,70,71,72,73,74,75,76],"Research","Experiments","ML","AI","ICLR","ACL","AAAI","COLM","LaTeX","Citations","Statistical Analysis",[78,79,80,46],"arxiv","ml-paper-writing","subagent-driven-development",[45,48],[83,49],"linux",[85,78,86,87,51,50,88,89],"semanticscholar","habanero","requests","matplotlib","SciencePlots","main","bdd9f958d7bb1a0f19b8ad8da020b0bab0f53eb1","vc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002FREPO_DOCS\u002F02-hermes-agent\u002Fskills\u002Fresearch\u002Fresearch-paper-writing",1,"https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis","2026-06-24 16:31:18",{"source":8,"skillName":5,"commands":98,"prompt":108,"zipAvailable":109},[99,102,105],{"tool":100,"command":101},"npx","npx skills add ibragimov-oasis\u002Fvibe-coder --skill research-paper-writing",{"tool":103,"command":104},"bunx","bunx skills add ibragimov-oasis\u002Fvibe-coder --skill research-paper-writing",{"tool":106,"command":107},"pnpm","pnpm dlx skills add ibragimov-oasis\u002Fvibe-coder --skill research-paper-writing","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002FREPO_DOCS\u002F02-hermes-agent\u002Fskills\u002Fresearch\u002Fresearch-paper-writing\u002FSKILL.md",true,[],[112,142,169,189,215,245],{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":122,"publishedAt":55},"01M6Q4Q22KKKVMJJYA411DQ4ER","ckm:ui-styling","Create beautiful, accessible user interfaces with shadcn\u002Fui components (built on Radix UI + Tailwind), Tailwind CSS utility-first styling, and canvas-based visual designs. Use when building user interfaces, implementing design systems, creating responsive layouts, adding accessible components (dialogs, dropdowns, forms, tables), customizing themes and colors, implementing dark mode, generating visual designs and posters, or establishing consistent styling patterns across applications.","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fui-design\u002Fui-rules\u002Fui-ux-pro-max\u002Fclaude\u002Fskills\u002Fui-styling\u002FSKILL.md",[118,119],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":120,"categoryName":121,"taxonomyType":32,"taxonomyTypeDesc":33},"01JBSEEDCAT000000000000032","设计师",[123,124,125,126,127,53,128,129,130,131,132,133,134,135,136,137,138,139,140,141],"ui styling","ui","styling","ckm:ui styling","ckm","css","canvas","components","component","create","design systems","Tailwind","across","themes","utility","mode","implementing","artifact\u002Fui-component","source\u002Fui-design",{"skillNo":143,"name":144,"nameZh":6,"description":145,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":146,"categories":147,"keywords":156,"publishedAt":55},"01M6Q4PQ2NX4NDVWTEHA3MQ2R0","ckm:design","Comprehensive design skill: brand identity, design tokens, UI styling, logo generation (55 styles, Gemini AI), corporate identity program (50 deliverables, CIP mockups), HTML presentations (Chart.js), banner design (22 styles, social\u002Fads\u002Fweb\u002Fprint), icon design (15 styles, SVG, Gemini 3.1 Pro), social photos (HTML→screenshot, multi-platform). Actions: design logo, create CIP, generate mockups, build slides, design banner, generate icon, create social photos, social media images, brand identity, design system. Platforms: Facebook, Twitter, LinkedIn, YouTube, Instagram, Pinterest, TikTok, Threads, Google Ads.","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fui-design\u002Fui-rules\u002Fui-ux-pro-max\u002Fclaude\u002Fskills\u002Fdesign\u002FSKILL.md",[148,151,152,155],{"categoryNo":149,"categoryName":150,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000003","演示文稿",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":153,"categoryName":154,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":120,"categoryName":121,"taxonomyType":32,"taxonomyTypeDesc":33},[53,144,127,157,131,123,158,132,159,160,161,162,163,164,165,166,167,168,140,141],"google","presentations","generate","screenshot","actions","Media","logo","ads","youtube","tiktok","artifact","chart",{"skillNo":170,"name":171,"nameZh":6,"description":172,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":173,"categories":174,"keywords":177,"publishedAt":55},"01M6Q4PJE27GEA3M0CECEBCZKS","ckm:design-system","Token architecture, component specifications, and slide generation. Three-layer tokens (primitive→semantic→component), CSS variables, spacing\u002Ftypography scales, component specs, strategic slide creation. Use for design tokens, systematic design, brand-compliant presentations.","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fui-design\u002Fui-rules\u002Fui-ux-pro-max\u002Fclaude\u002Fskills\u002Fdesign-system\u002FSKILL.md",[175,176],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":120,"categoryName":121,"taxonomyType":32,"taxonomyTypeDesc":33},[53,178,179,180,144,127,128,131,181,158,167,182,183,184,185,186,187,188,140,141],"design system","system","ckm:design system","architecture","token","creation","systematic","css variables","Variables","spacing","layer",{"skillNo":190,"name":191,"nameZh":6,"description":192,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":193,"categories":194,"keywords":200,"publishedAt":55},"01M6Q4P90JXFYMH2S634QM6SQJ","ckm:banner-design","Design banners for social media, ads, website heroes, creative assets, and print. Multiple art direction options with AI-generated visuals. Actions: design, create, generate banner. Platforms: Facebook, Twitter\u002FX, LinkedIn, YouTube, Instagram, Google Display, website hero, print. Styles: minimalist, gradient, bold typography, photo-based, illustrated, geometric, retro, glassmorphism, 3D, neon, duotone, editorial, collage. Uses ui-ux-pro-max, frontend-design, ai-artist, ai-multimodal skills.","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fui-design\u002Fui-rules\u002Fui-ux-pro-max\u002Fclaude\u002Fskills\u002Fbanner-design\u002FSKILL.md",[195,198,199],{"categoryNo":196,"categoryName":197,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000009","前端开发",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":120,"categoryName":121,"taxonomyType":32,"taxonomyTypeDesc":33},[53,201,202,203,127,204,161,162,205,206,207,208,209,210,211,212,140,141,213,214],"banner design","ckm:banner design","banner","frontend","3D","generated","Art","social","dimensions","neon","direction","display","frontend-design","hero",{"skillNo":216,"name":217,"nameZh":6,"description":218,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":219,"categories":220,"keywords":227,"publishedAt":55},"01M6Q2RZXBJ9TBE413GXCYQNDB","seo","Comprehensive SEO analysis for any website or business type. Performs full site audits, single-page deep analysis, technical SEO checks (crawlability, indexability, Core Web Vitals with INP), schema markup detection\u002Fvalidation\u002Fgeneration, content quality assessment (E-E-A-T framework per Dec 2025 update extending to all competitive queries), image optimization, sitemap analysis, and Generative Engine Optimization (GEO) for AI Overviews, ChatGPT, and Perplexity citations. Analyzes AI crawler accessibility (GPTBot, ClaudeBot, PerplexityBot), llms.txt compliance, brand mention signals, and passage-level citability. Industry detection for SaaS, e-commerce, local business, publishers, agencies. Triggers on: \"SEO\", \"audit\", \"schema\", \"Core Web Vitals\", \"sitemap\", \"E-E-A-T\", \"AI Overviews\", \"GEO\", \"technical SEO\", \"content quality\", \"page speed\", \"structured data\".","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fskills\u002Fskills-seo\u002Fseo\u002Fclaude-seo\u002Fskills\u002Fseo\u002FSKILL.md",[221,224],{"categoryNo":222,"categoryName":223,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000019","营销推广",{"categoryNo":225,"categoryName":226,"taxonomyType":32,"taxonomyTypeDesc":33},"01JBSEEDCAT000000000000035","市场营销",[217,228,229,230,231,75,232,167,233,234,235,236,237,238,239,240,241,242,243,244],"image","llms","assessment","deep analysis","analysis","local","deep","crawlability","seo analysis","2025","commerce","URL","domain\u002Fseo","source\u002Fskills-seo","vitals","generative","publishers",{"skillNo":246,"name":247,"nameZh":6,"description":248,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":249,"categories":250,"keywords":253,"publishedAt":55},"01M6Q2RVE7VGFCFQKBK468NHVY","seo-technical","Technical SEO audit across 9 categories: crawlability, indexability, security, URL structure, mobile, Core Web Vitals, structured data, JavaScript rendering, and IndexNow protocol. Use when user says \"technical SEO\", \"crawl issues\", \"robots.txt\", \"Core Web Vitals\", \"site speed\", or \"security headers\".","https:\u002F\u002Fgithub.com\u002Fibragimov-oasis\u002Fvibe-coder\u002Fblob\u002Fmain\u002Fvc-kit\u002Fcontent\u002FCOMBINED\u002FCOMBINED\u002Fskills\u002Fskills-seo\u002Fseo\u002Fclaude-seo\u002Fskills\u002Fseo-technical\u002FSKILL.md",[251,252],{"categoryNo":222,"categoryName":223,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":225,"categoryName":226,"taxonomyType":32,"taxonomyTypeDesc":33},[217,254,255,256,257,258,259,260,261,167,135,235,262,263,239,264,265,240,241,242],"technical","seo technical","issues","audit","data","mobile","crawl","structure","rendering","technical seo audit","headers","categories"]