[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6MGF4RSBMMGE9C6JYW5BPKX":3,"skill-same:01M6MGF4RSBMMGE9C6JYW5BPKX":81},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":30,"publishedAt":51,"linkOnly":52,"markdownContent":53,"markdownBody":54,"frontMatter":55,"refName":61,"commitSha":62,"skillDirectory":63,"forksCount":13,"repositoryHtmlUrl":64,"repositoryHomepageUrl":6,"ownerHtmlUrl":65,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":66,"install":67,"resources":80},"01M6MGF4RSBMMGE9C6JYW5BPKX","memory-engineering",null,"Use when designing, reviewing, or paying for an agent memory system — adding memory to an agent, choosing between long-context \u002F RAG \u002F graph \u002F agentic memory, auditing what a CLAUDE.md or memory directory actually holds, deciding what to keep and what to expire, or when a memory store keeps growing and nobody has said what leaves it. Prices the write path, picks which cost to pay, classifies records as facts \u002F skills \u002F logs, and refuses a design that has no forgetting policy.","JiayiV\u002Fquestion_mixing","JiayiV","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F54466889?v=4","01M6JFJYYMRZ3QRGP9339JNQ2K","01M6JFJZ35KB0QVT50NAHPJA7G",0,"MIT","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fpublic-02-claude-skills-main-gemini-skills-memory-engineering-skill-637c90fb\u002FSKILL.md",[17,22,25],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000015","视觉设计","SCENARIO","使用场景",{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":26,"categoryName":27,"taxonomyType":28,"taxonomyTypeDesc":29},"01JBSEEDCAT000000000000032","设计师","PROFESSION","适用职业",[31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50],"memory","engineering","memory engineering","SkillOpt","design spec","prices","which","pay","sleep","consolidation","lens","agent memory system","memory system","design","agent-harness","question","loop","wiki","logs","agentic","2026-11-10 10:12:30",false,"---\nname: memory-engineering\ndescription: Use when designing, reviewing, or paying for an agent memory system — adding memory to an agent, choosing between long-context \u002F RAG \u002F graph \u002F agentic memory, auditing what a CLAUDE.md or memory directory actually holds, deciding what to keep and what to expire, or when a memory store keeps growing and nobody has said what leaves it. Prices the write path, picks which cost to pay, classifies records as facts \u002F skills \u002F logs, and refuses a design that has no forgetting policy.\nargument-hint: \"[optional: path to a memory dir, design spec JSON, or a question]\"\nlicense: MIT\nmetadata:\n  version: 1.0.0\n  build_pattern: \"Four-lens synthesis (Stanford \u002F Microsoft \u002F Anthropic \u002F Nvidia) + 4 deterministic stdlib scripts, with a blocking forgetting gate\"\n  distinct_from: \"llm-wiki (maintains one specific markdown vault; this audits and prices any memory system); skillopt-sleep (runs a nightly consolidation loop; this decides whether that loop's output is worth keeping); agent-harness (bounds a task loop; this bounds a store)\"\n---\n\n# Memory Engineering — engineer the forgetting, not just the remembering\n\n> **Portability:** 4 stdlib scripts, no APIs\u002FLLM calls\u002Fnetwork. They measure and gate; you decide.\n\n## What this does\n\nAnyone can give an agent memory: vector store, pipe in the history, retrieve\ntop-k. That works until the history outgrows the context window, the write path\ncosts more than every query it serves, and the store fills with stale state\nnobody removes. Memory is not a bucket — it is a system with a metabolism.\n\n**The shift:** a storer optimizes what a system remembers; a memory engineer\noptimizes what it forgets. The problem was never that an agent forgets — it is\nthat it never forgets *on purpose*.\n\n## The four lenses\n\n| Lens | Question | The finding that hurts |\n|---|---|---|\n| **Stanford** | What does remembering cost? | Construction energy exceeds total query energy across 300 queries. The tuned half is the smaller half. |\n| **Microsoft** | What is worth keeping? | More raw memory can make an agent *worse*. Keep facts and skills; drop the events. |\n| **Anthropic** | Who controls what it keeps? | A wrong memory does not fail once — it persists into every future session that reads it. |\n| **Nvidia** | Where does it hit hardware? | It is all KV cache in HBM. Construction is prefill-heavy and stalls the query a user is waiting on. |\n\n## Workflow\n\n```bash\n# 1 - Price it first. Never quote a quality number without a cost number.\npython scripts\u002Fmemory_cost_profiler.py --print-sample-spec > workload.json\npython scripts\u002Fmemory_cost_profiler.py --spec workload.json\n# 2 - Pick which cost to pay. No \"best\" verdict; on a tie it asks, exit 2.\npython scripts\u002Fmemory_architecture_picker.py --constraints workload.json\n# 3 - Audit what the store actually holds (skip if greenfield).\npython scripts\u002Fmemory_density_auditor.py --dir ~\u002F.claude\u002Fmemory\n# 4 - Gate on forgetting. Exit 4 is a stop, not a suggestion.\npython scripts\u002Fforgetting_policy_linter.py --policy design.json\n# 5 - No command. Prove each pass by hand before scheduling it.\n```\n\nStep 1 reports the construction\u002Fquery split, **cost per correct answer**, and\namortization — if construction dominates, cut construction tokens *before*\ntouching retrieval. Step 2 names the cost the winning family makes you pay.\nStep 3 classifies records FACT \u002F SKILL \u002F LOG \u002F PROSE (`LOG-HEAVY` = archiving\nevents; `PROSE-HEAVY` = docs, not memory).\n\nStep 4 is the gate: **F1** (explicit forgetting rule) and **F4** (contradictions\nsurfaced, never auto-merged) are blocking. Retrofitting forgetting onto two\nyears of records is a migration nobody does; auto-merging disagreeing memories\ndestroys the evidence the conflict existed.\n\nStep 5 has no script — prove each pass by hand, then automate. Run it once\nagainst real history and ask whether it changed a decision. If not, scheduling\nit only makes noise. Ship order: `forgetting_policy_design.md` §7.\n\n## Hard rules\n\n1. **Never quote accuracy without cost per correct answer.**\n2. **Never return a \"best\" memory system** — name the cost the choice makes you pay.\n3. **Never auto-merge contradictions.** The system surfaces; the human decides.\n4. **Never call a design done without a forgetting rule.** No evaluated system provides one by default.\n5. **Never schedule a pass not yet run by hand.**\n6. **Report findings as findings.** A non-zero exit is a result to surface, not an error to swallow.\n7. **Attribute every number** with its confidence level. Vendor customer figures are testimonials, not benchmarks.\n\n## Scripts\n\n| Script | Role | Exit codes |\n|---|---|---|\n| `scripts\u002Fmemory_cost_profiler.py` | Construction vs query split, cost per correct answer, amortization, co-location warning | 0 · 2 finding · 3 bad input |\n| `scripts\u002Fmemory_architecture_picker.py` | Scores 4 families, disqualifies, names the cost, refuses to pick on a tie | 0 · 2 ambiguous · 3 bad input · 4 none viable |\n| `scripts\u002Fmemory_density_auditor.py` | FACT\u002FSKILL\u002FLOG\u002FPROSE, duplicates, staleness, density (`--dir` or `--jsonl`) | 0 dense · 2 finding · 3 bad input |\n| `scripts\u002Fforgetting_policy_linter.py` | The gate: 8 checks, F1 and F4 blocking | 0 PASS · 2 CONDITIONAL · 4 FAIL |\n\nAll support `--output json` and `--sample` (no input file needed).\n\n## References and assets\n\n- [`references\u002Fmemory_cost_canon.md`](references\u002Fmemory_cost_canon.md) — construction dominance, energy per correct answer, the four families, ten recommendations (7 sources)\n- [`references\u002Fwhat_to_keep.md`](references\u002Fwhat_to_keep.md) — PlugMem and MEMENTO: facts over logs, density over volume (7 sources)\n- [`references\u002Fmemory_control_and_governance.md`](references\u002Fmemory_control_and_governance.md) — memory as files, scope\u002Faudit\u002Frollback, poisoning, reading vendor numbers (7 sources)\n- [`references\u002Fforgetting_policy_design.md`](references\u002Fforgetting_policy_design.md) — forgetting mechanisms, contradiction discipline, KV cache, ship order (7 sources)\n- [`assets\u002Fmemory_engineer_worksheet.md`](assets\u002Fmemory_engineer_worksheet.md) — seven forcing questions with recommended answers + citations; walk one at a time\n- [`assets\u002Fmemory_design_spec.example.json`](assets\u002Fmemory_design_spec.example.json) — one file covering every script's input\n- [`assets\u002Fforgetting_policy_template.md`](assets\u002Fforgetting_policy_template.md) — fillable policy covering F1–F8\n\n## Provenance\n\nFraming from *\"How to be a Memory Engineer\"* by [@N01ennn](https:\u002F\u002Fx.com\u002FN01ennn\u002Fstatus\u002F2083971749079581120); every\nnumber is cited to a primary source instead, and two paraphrases are corrected — `memory_cost_canon.md` §2, `memory_control_and_governance.md` §4.\n","# Memory Engineering — engineer the forgetting, not just the remembering\n\n> **Portability:** 4 stdlib scripts, no APIs\u002FLLM calls\u002Fnetwork. They measure and gate; you decide.\n\n## What this does\n\nAnyone can give an agent memory: vector store, pipe in the history, retrieve\ntop-k. That works until the history outgrows the context window, the write path\ncosts more than every query it serves, and the store fills with stale state\nnobody removes. Memory is not a bucket — it is a system with a metabolism.\n\n**The shift:** a storer optimizes what a system remembers; a memory engineer\noptimizes what it forgets. The problem was never that an agent forgets — it is\nthat it never forgets *on purpose*.\n\n## The four lenses\n\n| Lens | Question | The finding that hurts |\n|---|---|---|\n| **Stanford** | What does remembering cost? | Construction energy exceeds total query energy across 300 queries. The tuned half is the smaller half. |\n| **Microsoft** | What is worth keeping? | More raw memory can make an agent *worse*. Keep facts and skills; drop the events. |\n| **Anthropic** | Who controls what it keeps? | A wrong memory does not fail once — it persists into every future session that reads it. |\n| **Nvidia** | Where does it hit hardware? | It is all KV cache in HBM. Construction is prefill-heavy and stalls the query a user is waiting on. |\n\n## Workflow\n\n```bash\n# 1 - Price it first. Never quote a quality number without a cost number.\npython scripts\u002Fmemory_cost_profiler.py --print-sample-spec > workload.json\npython scripts\u002Fmemory_cost_profiler.py --spec workload.json\n# 2 - Pick which cost to pay. No \"best\" verdict; on a tie it asks, exit 2.\npython scripts\u002Fmemory_architecture_picker.py --constraints workload.json\n# 3 - Audit what the store actually holds (skip if greenfield).\npython scripts\u002Fmemory_density_auditor.py --dir ~\u002F.claude\u002Fmemory\n# 4 - Gate on forgetting. Exit 4 is a stop, not a suggestion.\npython scripts\u002Fforgetting_policy_linter.py --policy design.json\n# 5 - No command. Prove each pass by hand before scheduling it.\n```\n\nStep 1 reports the construction\u002Fquery split, **cost per correct answer**, and\namortization — if construction dominates, cut construction tokens *before*\ntouching retrieval. Step 2 names the cost the winning family makes you pay.\nStep 3 classifies records FACT \u002F SKILL \u002F LOG \u002F PROSE (`LOG-HEAVY` = archiving\nevents; `PROSE-HEAVY` = docs, not memory).\n\nStep 4 is the gate: **F1** (explicit forgetting rule) and **F4** (contradictions\nsurfaced, never auto-merged) are blocking. Retrofitting forgetting onto two\nyears of records is a migration nobody does; auto-merging disagreeing memories\ndestroys the evidence the conflict existed.\n\nStep 5 has no script — prove each pass by hand, then automate. Run it once\nagainst real history and ask whether it changed a decision. If not, scheduling\nit only makes noise. Ship order: `forgetting_policy_design.md` §7.\n\n## Hard rules\n\n1. **Never quote accuracy without cost per correct answer.**\n2. **Never return a \"best\" memory system** — name the cost the choice makes you pay.\n3. **Never auto-merge contradictions.** The system surfaces; the human decides.\n4. **Never call a design done without a forgetting rule.** No evaluated system provides one by default.\n5. **Never schedule a pass not yet run by hand.**\n6. **Report findings as findings.** A non-zero exit is a result to surface, not an error to swallow.\n7. **Attribute every number** with its confidence level. Vendor customer figures are testimonials, not benchmarks.\n\n## Scripts\n\n| Script | Role | Exit codes |\n|---|---|---|\n| `scripts\u002Fmemory_cost_profiler.py` | Construction vs query split, cost per correct answer, amortization, co-location warning | 0 · 2 finding · 3 bad input |\n| `scripts\u002Fmemory_architecture_picker.py` | Scores 4 families, disqualifies, names the cost, refuses to pick on a tie | 0 · 2 ambiguous · 3 bad input · 4 none viable |\n| `scripts\u002Fmemory_density_auditor.py` | FACT\u002FSKILL\u002FLOG\u002FPROSE, duplicates, staleness, density (`--dir` or `--jsonl`) | 0 dense · 2 finding · 3 bad input |\n| `scripts\u002Fforgetting_policy_linter.py` | The gate: 8 checks, F1 and F4 blocking | 0 PASS · 2 CONDITIONAL · 4 FAIL |\n\nAll support `--output json` and `--sample` (no input file needed).\n\n## References and assets\n\n- [`references\u002Fmemory_cost_canon.md`](references\u002Fmemory_cost_canon.md) — construction dominance, energy per correct answer, the four families, ten recommendations (7 sources)\n- [`references\u002Fwhat_to_keep.md`](references\u002Fwhat_to_keep.md) — PlugMem and MEMENTO: facts over logs, density over volume (7 sources)\n- [`references\u002Fmemory_control_and_governance.md`](references\u002Fmemory_control_and_governance.md) — memory as files, scope\u002Faudit\u002Frollback, poisoning, reading vendor numbers (7 sources)\n- [`references\u002Fforgetting_policy_design.md`](references\u002Fforgetting_policy_design.md) — forgetting mechanisms, contradiction discipline, KV cache, ship order (7 sources)\n- [`assets\u002Fmemory_engineer_worksheet.md`](assets\u002Fmemory_engineer_worksheet.md) — seven forcing questions with recommended answers + citations; walk one at a time\n- [`assets\u002Fmemory_design_spec.example.json`](assets\u002Fmemory_design_spec.example.json) — one file covering every script's input\n- [`assets\u002Fforgetting_policy_template.md`](assets\u002Fforgetting_policy_template.md) — fillable policy covering F1–F8\n\n## Provenance\n\nFraming from *\"How to be a Memory Engineer\"* by [@N01ennn](https:\u002F\u002Fx.com\u002FN01ennn\u002Fstatus\u002F2083971749079581120); every\nnumber is cited to a primary source instead, and two paraphrases are corrected — `memory_cost_canon.md` §2, `memory_control_and_governance.md` §4.\n",{"name":5,"license":14,"metadata":56,"description":7,"argument-hint":60},{"version":57,"build_pattern":58,"distinct_from":59},"1.0.0","Four-lens synthesis (Stanford \u002F Microsoft \u002F Anthropic \u002F Nvidia) + 4 deterministic stdlib scripts, with a blocking forgetting gate","llm-wiki (maintains one specific markdown vault; this audits and prices any memory system); skillopt-sleep (runs a nightly consolidation loop; this decides whether that loop's output is worth keeping); agent-harness (bounds a task loop; this bounds a store)","[optional: path to a memory dir, design spec JSON, or a question]","main","068c2486ccc0f8d200a3db4d21ec137064eda3b4","skills\u002Fpublic-02-claude-skills-main-gemini-skills-memory-engineering-skill-637c90fb","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing","https:\u002F\u002Fgithub.com\u002FJiayiV","2026-08-25 16:30:52",{"source":8,"skillName":5,"commands":68,"prompt":78,"zipAvailable":79},[69,72,75],{"tool":70,"command":71},"npx","npx skills add JiayiV\u002Fquestion_mixing --skill memory-engineering",{"tool":73,"command":74},"bunx","bunx skills add JiayiV\u002Fquestion_mixing --skill memory-engineering",{"tool":76,"command":77},"pnpm","pnpm dlx skills add JiayiV\u002Fquestion_mixing --skill memory-engineering","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fpublic-02-claude-skills-main-gemini-skills-memory-engineering-skill-637c90fb\u002FSKILL.md",true,[],[82,108,136,165,193,226],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":51},"01M6NN9PJS9NHWPDG4QP93BP8F","webmcpify","Make a web app agent-ready — propose a WebMCP tool manifest, integrate, verify in a real browser, heal; unrelated code stays untouched. Use for \"webmcpify\", \"add WebMCP\", or \"expose app actions to AI agents\".","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fwebmcpify\u002FSKILL.md",[],[84,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107],"github","ai agents","web","notes","code","browser","https","actions","manifest","inventory","verify","tool","use","ready","unrelated","app","scope","for","webmcp",{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":112,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":51},"01M6NN6KZE5RBN6NGHN2WXV1RE","vardoger-analyze","Use when the user asks to personalize the GitHub Copilot CLI assistant, adapt Copilot to their style, use vardoger, or analyze their Copilot CLI conversation history. Reads the local session directory at `~\u002F.copilot\u002Fsession-state\u002F`, extracts recurring preferences and conventions, and writes a fenced personalization block into `~\u002F.copilot\u002Fcopilot-instructions.md`. Runs entirely on the user's machine via the local `vardoger` CLI (`pipx install vardoger`); no network calls and no uploads. Triggers: 'personalize my copilot', 'analyze my copilot history', 'tailor copilot to me', 'run vardoger', 'update my copilot instructions from history', 'make copilot learn my style'.","Apache-2.0","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fvardoger-analyze\u002FSKILL.md",[],[116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135],"vardoger analyze","analyze","directory","cli","github copilot","copilot","run","update","learn","network","my style","install","conversation","me","local","github copilot cli","pipx","my","at","copilot instructions",{"skillNo":137,"name":138,"nameZh":6,"description":139,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":140,"categories":141,"keywords":146,"publishedAt":51},"01M6NN3CZGQP3XAJ04BM1NNYGT","transloadit-media-processing","Process media files (video, audio, images, documents) using Transloadit. Use when asked to encode video to HLS\u002FMP4, generate thumbnails, resize or watermark images, extract audio, concatenate clips, add subtitles, OCR documents, or run any media processing pipeline. Covers 86+ processing robots for file transformation at scale.","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Ftransloadit-media-processing\u002FSKILL.md",[142,145],{"categoryNo":143,"categoryName":144,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000016","图像与多媒体",{"categoryNo":23,"categoryName":24,"taxonomyType":20,"taxonomyTypeDesc":21},[147,148,149,150,151,152,153,154,155,156,157,158,159,101,160,161,162,163,106,164],"transloadit media processing","Media","media processing","processing","MCP","mcp server","documents","subtitles","pipeline","ocr","Video","Audio","mp4","images","files","transformation","free","hls",{"skillNo":166,"name":167,"nameZh":6,"description":168,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":169,"categories":170,"keywords":174,"publishedAt":51},"01M6NN28E92F9GH4WFYQ2H1Y15","terraform-azurerm-set-diff-analyzer","Analyze Terraform plan JSON output for AzureRM Provider to distinguish between false-positive diffs (order-only changes in Set-type attributes) and actual resource changes. Use when reviewing terraform plan output for Azure resources like Application Gateway, Load Balancer, Firewall, Front Door, NSG, and other resources with Set-type attributes that cause spurious diffs due to internal ordering changes.","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fterraform-azurerm-set-diff-analyzer\u002FSKILL.md",[171],{"categoryNo":172,"categoryName":173,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000011","运维与部署",[175,176,177,178,179,180,181,182,183,101,184,185,186,187,188,189,190,191,192,106],"terraform","SET","terraform azurerm set diff analyzer","diff","analyzer","diffs","azure","gateway","firewall","only","resource changes","provider","resources","load balancer","attributes","front","load","balancer",{"skillNo":194,"name":195,"nameZh":6,"description":196,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":197,"categories":198,"keywords":208,"publishedAt":51},"01M6NN1QMH2TDYWTH3XY3N5XPV","technical-job-search","Use this skill when a software engineer asks for help with job search tasks: parsing or analyzing a job description, tailoring a CV\u002Fresume, writing a cover letter, evaluating a job offer, or drafting a post-interview follow-up email. Do not activate for general career advice unrelated to an active job search action.","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Ftechnical-job-search\u002FSKILL.md",[199,202,205],{"categoryNo":200,"categoryName":201,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":203,"categoryName":204,"taxonomyType":28,"taxonomyTypeDesc":29},"01JBSEEDCAT000000000000036","内容创作者",{"categoryNo":206,"categoryName":207,"taxonomyType":28,"taxonomyTypeDesc":29},"01JBSEEDCAT000000000000038","人力资源",[209,210,211,212,213,214,215,216,117,217,101,218,219,220,221,222,223,224,106,225],"search","Job Search","technical job search","technical","writing","tasks","email","parsing","interview","up","engineer","general","activate","do","task","cv","letter",{"skillNo":227,"name":228,"nameZh":6,"description":229,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":230,"categories":231,"keywords":232,"publishedAt":51},"01M6NMZFA67XYS1GS09WT42235","steno-mode","Shorthand-first response compression that cuts ~40% of response tokens while preserving technical precision and exact literals. Use when the user says \"steno mode\", \"shorthand mode\", \"compressed responses\", \"token reduction\", \"brief structured output\", or invokes \u002Fsteno. Supports four compression levels: lite, brief, court, machine. Do not trigger for requests needing polished prose such as onboarding\u002Ftutorial content, stakeholder or customer-facing copy, or teaching-focused explanations.","https:\u002F\u002Fgithub.com\u002FJiayiV\u002Fquestion_mixing\u002Fblob\u002Fmain\u002Fskills\u002Fsteno-mode\u002FSKILL.md",[],[233,234,235,236,237,212,238,101,222,239,240,241,242,243,244,245,246,247,106,248],"steno mode","mode","customer","brief","first","tutorial","teaching","tokens","as","trigger","structured output","machine","output","court","requests","response compression"]