{"id":"17832d92-3647-4e11-ab01-ecd79e8e50f3","shortId":"mm2ByT","kind":"skill","title":"ai-md","tagline":"Convert human-written CLAUDE.md into AI-native structured-label format. Battle-tested across 4 models. Same rules, fewer tokens, higher compliance.","description":"# AI.MD v4 — The Complete AI-Native Conversion System\n\n## When to Use This Skill\n\n- Use when your CLAUDE.md is long but AI still ignores your rules\n- Use when token usage is too high from verbose system instructions\n- Use when you want to optimize any LLM system prompt for compliance\n- Use when migrating rules between AI tools (Claude, Codex, Gemini, Grok)\n\n## What Is AI.MD?\n\nAI.MD is a methodology for converting human-written `CLAUDE.md` (or any LLM system instructions)\ninto a structured-label format that AI models follow more reliably, using fewer tokens.\n\n**The paradox we proved:** Adding more rules in natural language DECREASES compliance.\nConverting the same rules to structured format RESTORES and EXCEEDS it.\n\n```\nHuman prose (6 rules, 1 line)  → AI follows 4 of them\nStructured labels (6 rules, 6 lines) → AI follows all 6\nSame content. Different format. Different results.\n```\n\n---\n\n## Why It Works: How LLMs Actually Process Instructions\n\nLLMs don't \"read\" — they **attend**. Understanding this changes everything.\n\n### Mechanism 1: Attention Splitting\n\nWhen multiple rules share one line, the model's attention distributes across all tokens equally.\nEach rule gets a fraction of the attention weight. Some rules get lost.\n\nWhen each rule has its own line, the model processes it as a distinct unit.\nFull attention weight on each rule.\n\n```\n# ONE LINE = attention splits 5 ways (some rules drop to near-zero weight)\nEVIDENCE: no-fabricate no-guess | 禁用詞:應該是/可能是 → 先拿數據 | Read/Grep→行號 curl→數據 | \"好像\"/\"覺得\"→自己先跑test | guess=shame-wall\n\n# FIVE LINES = each rule gets full attention\nEVIDENCE:\n  core: no-fabricate | no-guess | unsure=say-so\n  banned: 應該是/可能是/感覺是/推測 → 先拿數據\n  proof: all-claims-need(data/line#/source) | Read/Grep→行號 | curl→數據\n  hear-doubt: \"好像\"/\"覺得\" → self-test(curl/benchmark) → 禁反問user\n  violation: guess → shame-wall\n```\n\n### Mechanism 2: Zero-Inference Labels\n\nNatural language forces the model to INFER meaning from context.\nLabels DECLARE meaning explicitly. No inference needed = no misinterpretation.\n\n```\n# AI must infer: what does (防搞混) modify? what does 例外 apply to?\nGATE-1: 收到任務→先用一句話複述(防搞混)(長對話中每個新任務都重新觸發) | 例外: signals命中「處理一下」=直接執行\n\n# AI reads labels directly: trigger→action→exception. Zero ambiguity.\nGATE-1 複述:\n  trigger: new-task\n  action: first-sentence=\"你要我做的是___\"\n  persist: 長對話中每個新任務都重新觸發\n  exception: signal=處理一下 → skip\n  yields-to: GATE-3\n```\n\nKey insight: Labels like `trigger:` `action:` `exception:` work across ALL languages.\nThe model doesn't need to parse Chinese/Japanese/English grammar to understand structure.\n**Labels are the universal language between humans and AI.**\n\n### Mechanism 3: Semantic Anchoring\n\nLabeled sub-items create **matchable tags**. When a user's input contains a keyword,\nthe model matches it directly to the corresponding label — like a hash table lookup\ninstead of a full-text search.\n\n```\n# BURIED: AI scans the whole sentence, might miss the connection\n加新功能→第一句問schema | 新增API/endpoint=必確認health-check.py覆蓋\n\n# ANCHORED: label \"new-api:\" directly matches user saying \"加個 API\"\nMOAT:\n  new-feature: 第一句問schema/契約/關聯\n  new-api: 必確認health-check.py覆蓋(GATE-5)\n```\n\n**Real proof:** This specific technique fixed a test case that failed 5 consecutive times\nacross all models. The label `new-api:` raised Codex T5 from ❌→✅ on first try.\n\n---\n\n## The Conversion Process: What Happens When You Give Me a CLAUDE.md\n\nHere's the exact mental model I use when converting natural language instructions to AI.MD format.\n\n### Phase 1: UNDERSTAND — Read Like a Compiler, Not a Human\n\nI read the CLAUDE.md **as if I'm building a state machine**, not reading a document.\n\nFor each sentence, I ask:\n1. **Is this a TRIGGER?** (What input activates this behavior?)\n2. **Is this an ACTION?** (What should the AI do?)\n3. **Is this a CONSTRAINT?** (What should the AI NOT do?)\n4. **Is this METADATA?** (Priority, timing, persistence, exceptions?)\n5. **Is this a HUMAN EXPLANATION?** (Why the rule exists — delete this)\n\nExample analysis:\n\n```\nInput: \"收到任務→先用一句話複述(防搞混)(長對話中每個新任務都重新觸發) | 例外: signals命中「處理一下」=直接執行\"\n\nDecomposition:\n  ├─ TRIGGER:    \"收到任務\" → new-task\n  ├─ ACTION:     \"先用一句話複述\" → first-sentence=\"你要我做的是___\"\n  ├─ DELETE:     \"(防搞混)\" → human motivation, AI doesn't need this\n  ├─ METADATA:   \"(長對話中每個新任務都重新觸發)\" → persist: every-new-task\n  └─ EXCEPTION:  \"例外: signals命中「處理一下」=直接執行\" → exception: signal=處理一下 → skip\n```\n\n### Phase 2: DECOMPOSE — Break Every `|` and `()` Into Atomic Rules\n\nThe #1 source of compliance failure is **compound rules**.\nA single line with 3 rules separated by `|` looks like 1 instruction to AI.\nIt needs to be 3 separate instructions.\n\n**The splitter test:** If you can put \"AND\" between two parts of a sentence,\nthey are separate rules and MUST be on separate lines.\n\n```\n# Input: one sentence hiding 4 rules\n禁用詞:應該是/可能是→先拿數據 | \"好像\"/\"覺得\"→自己先跑test(不是問user)→有數據才能決定\n\n# Analysis: I find 4 hidden rules\nRule 1: certain words are banned → use data instead\nRule 2: hearing doubt words → run self-test\nRule 3: don't ask the user for data → look it up yourself\nRule 4: preference claims → require A/B comparison before accepting\n\n# Output: 4 atomic rules\nbanned: 應該是/可能是/感覺是/推測 → 先拿數據\nhear-doubt: \"好像\"/\"覺得\" → self-test(curl/benchmark)\nself-serve: 禁反問user(自己查)\ncompare: \"覺得A比B好\" → A/B實測先行\n```\n\n### Phase 3: LABEL — Assign Function Labels\n\nEvery atomic rule gets a label that declares its function.\nI use a standard vocabulary of ~12 label types:\n\n| Label | What It Declares | When to Use |\n|-------|-----------------|-------------|\n| `trigger:` | What input activates this | Every gate/rule needs one |\n| `action:` | What the AI must do | The core behavior |\n| `exception:` | When NOT to do it | Override cases |\n| `not-triggered:` | Explicit negative examples | Prevent over-triggering |\n| `format:` | Output format constraint | Position, structure requirements |\n| `priority:` | Override relationship | When rules conflict |\n| `yields-to:` | Which gate takes precedence | Inter-gate priority |\n| `persist:` | Durability across turns | Rules that survive conversation flow |\n| `timing:` | When in the workflow | Before/after/during constraints |\n| `violation:` | Consequence of breaking | Accountability mechanism |\n| `banned:` | Forbidden words/actions | Hard no-go list |\n| `policy:` | Decision heuristic | When judgment is needed |\n\n**The label selection technique:** I pick the label that would help a DIFFERENT AI model\n(not the one being instructed) understand this rule's function if it saw ONLY the label.\nIf `trigger:` clearly tells you \"this is what activates the rule\" without reading anything else,\nit's the right label.\n\n### Phase 4: STRUCTURE — Build the Architecture\n\nI organize rules into a hierarchy:\n\n```\n<gates>    = Hard stops (MUST check before any action)\n<rules>    = Behavioral guidelines (HOW to act)\n<rhythm>   = Workflow patterns (WHEN to do what)\n<conn>     = Connection strings (FACTS — never compress)\n<ref>      = On-demand references (don't load until needed)\n<learn>    = Evolution rules (how the system improves)\n```\n\n**Why this order matters:**\nGates come first because they MUST be checked before anything else.\nThe model processes instructions top-to-bottom. Priority = position.\n\n**Grouping technique:** Rules that share a DOMAIN become sub-items under one heading.\n\n```\n# FLAT (bad): 7 unrelated rules, model treats equally\n1. no guessing\n2. backup before editing\n3. use tables for output\n4. check health after deploy\n5. don't say \"應該是\"\n6. test before reporting\n7. all claims need proof\n\n# GROUPED (good): 3 domains, model understands hierarchy\nEVIDENCE:               ← domain: truthfulness\n  core: no-guess\n  banned: 應該是\n  proof: all-claims-need-data\n\nSCOPE:                  ← domain: safety\n  pre-change: backup\n  pre-run: check-health\n\nOUTPUT:                 ← domain: format\n  format: tables+numbers\n```\n\n### Phase 5: RESOLVE — Handle Conflicts and Edge Cases\n\nThis is the most critical and least obvious phase. Natural language instructions\noften contain **hidden conflicts** that humans resolve with intuition but AI cannot.\n\n**Technique: Conflict Detection Matrix**\n\nI check every pair of gates/rules for conflicts:\n\n```\nGATE-1 (複述: repeat task) vs GATE-3 (保護檔: backup first)\n→ CONFLICT: If user says \"edit .env\", should AI repeat the task first, or backup first?\n→ RESOLUTION: priority: GATE-3 > GATE-1 (safety before courtesy)\n             yields-to: GATE-3 (explicit in GATE-1)\n\nGATE-4 (報結論: cite evidence) vs bug-close (記錄根因: write root cause)\n→ CONFLICT: bug-close requires stating root cause, but GATE-4 bans definitive claims\n→ RESOLUTION: timing: GATE-4 is pre-conclusion brake; bug-close is post-verification record\n             GATE-4 not-triggered when bug already verified\n\nEVIDENCE (no-guess) vs user says \"處理一下\" (just do it)\n→ CONFLICT: should AI verify assumptions or execute immediately?\n→ RESOLUTION: signal \"處理一下\" = user has decided, skip confirmation\n```\n\n**Technique: Not-Triggered Lists**\n\nFor any rule that could over-trigger, I add explicit negative examples:\n\n```\nGATE-4 報結論:\n  trigger: 最終歸因/根因判定/不可逆建議\n  not-triggered: 中間進度數字 | 純指標查詢 | 工具原始輸出 | 已知事實 | 轉述文件\n```\n\nThis was discovered because Gemini 2.5 Pro kept triggering GATE-4 on simple number queries\nlike \"成功率怎麼樣?\". Adding `not-triggered: 純指標查詢` fixed it immediately.\n\n### Phase 6: TEST — Multi-Model Validation (Non-Negotiable)\n\n**This is not optional.** Every conversion MUST be validated by 2+ different LLM models.\n\nWhy? Because a format that works perfectly for Claude might confuse GPT, and vice versa.\nThe whole point of AI.MD is that it works ACROSS models.\n\n**The exam protocol we developed:**\n\n1. Write 8 test inputs that simulate REAL user behavior (not textbook examples)\n2. Include \"trap\" questions where two rules conflict\n3. Include \"negative\" tests where a rule should NOT trigger\n4. DO NOT hint which rules are being tested (the AI shouldn't know)\n5. Run each model independently\n6. Score each answer: ✅ full compliance, ⚠️ partial, ❌ miss\n7. If ANY model's score drops after conversion → revert that specific change\n\n**The 8-question template we used:**\n\n```\nT1: Simple task (does GATE-1 trigger?)\nT2: Database write attempt (does GATE-2 catch it?)\nT3: Protected file edit (does GATE-3 fire FIRST, before GATE-1?)\nT4: Root cause analysis (does GATE-4 require all 4 questions?)\nT5: Business API addition (does AI mention health-check.py?)\nT6: User says \"好像X比Y好\" (does AI run comparison or just accept it?)\nT7: User says \"處理一下\" (does AI skip GATE-1 confirmation?)\nT8: Simple metric query (does GATE-4 NOT trigger?)\n```\n\n---\n\n## Special Techniques Discovered During Battle-Testing\n\n### Technique 1: Bilingual Label Strategy\n\nLabels in English, output strings in the user's language.\nEnglish labels are shorter AND more universally understood by all models.\nBut the actual text the AI produces must stay in the user's language.\n\n```\naction: first-sentence=\"你要我做的是___\"    ← AI outputs Chinese\nformat: must-be-line-1                      ← structural constraint in English\nbanned: 應該是/可能是                        ← forbidden words stay in original language\n```\n\n**Why this works:** English label vocabulary (`trigger`, `action`, `exception`) maps directly\nto concepts in every model's training data. Chinese grammar labels (觸發條件, 執行動作, 例外情況)\nare less standardized across models.\n\n### Technique 2: State Machine Gates\n\nInstead of treating rules as a flat list, model them as a **state machine**:\n- Each gate has a `trigger` (input state)\n- Each gate has an `action` (transition)\n- Gates have `priority` (which fires first when multiple match)\n- Gates have `yields-to` (explicit conflict resolution)\n\nThis gives AI a clear execution model:\n```\nInput arrives → Check GATE-3 first (highest priority) → Check GATE-1 → Check GATE-2 → ...\n```\n\nInstead of:\n```\nInput arrives → Read all rules → Try to figure out which one applies → Maybe miss one\n```\n\n### Technique 3: XML Section Tags for Semantic Boundaries\n\nUsing `<gates>`, `<rules>`, `<rhythm>`, `<conn>` as section delimiters\ncreates hard boundaries that prevent rule-bleed (where the model confuses\nwhich section a rule belongs to).\n\n```xml\n<gates label=\"硬性閘門 | 優先序: gates>rules>rhythm | 缺一項=STOP\">\n...gates here...\n</gates>\n\n<rules>\n...rules here...\n</rules>\n```\n\nThe `label` attribute on the opening tag serves as a section-level instruction:\n\"these are hard gates, this is their priority, missing = stop\"\n\n### Technique 4: Cross-Reference Instead of Duplicate\n\nWhen the same concept appears in multiple rules, DON'T repeat it.\nUse a cross-reference label.\n\n```\n# BAD: health-check mentioned in 3 places\nGATE-5: ...check health-check.py...\nMOAT: ...must check health-check.py...\nSCOPE: ...verify health-check.py exists...\n\n# GOOD: single source of truth + cross-reference\nGATE-5 驗收:\n  checks:\n    新增API → 確認health-check.py覆蓋\n\nMOAT:\n  new-api: 必確認health-check.py覆蓋(GATE-5)    ← cross-ref, not duplicate\n```\n\n### Technique 5: The \"What Not Why\" Principle\n\nDelete ALL text that exists to explain WHY a rule exists.\nAI needs WHAT to do, not WHY.\n\n```\n# DELETE these human explanations:\n(防搞混)                     → motivation\n(不是大爆破,是每次順手一點)    → metaphor\n(想清楚100倍後才做現在的)     → backstory\n(因為用戶是非工程師)          → justification\n\n# KEEP only the actionable instruction:\naction: first-sentence=\"你要我做的是___\"\nrefactor: 同區塊連續第3次修改 → extract\n```\n\nEvery deleted explanation saves tokens AND removes noise that could confuse the model\nabout what it should actually DO.\n\n---\n\n## Two-Stage Workflow\n\n### Stage 1: PREVIEW — Measure, Don't Touch\n\n```bash\necho \"=== Current Token Burn ===\"\nclaude_md=$(wc -c < ~/.claude/CLAUDE.md 2>/dev/null || echo 0)\nrules=$(cat ~/.claude/rules/*.md 2>/dev/null | wc -c || echo 0)\ntotal=$((claude_md + rules))\ntokens=$((total / 4))\necho \"CLAUDE.md:     $claude_md bytes\"\necho \"rules/*.md:    $rules bytes\"\necho \"Total:         $total bytes ≈ $tokens tokens/turn\"\necho \"50-turn session: ≈ $((tokens * 50)) tokens on instructions alone\"\n```\n\nThen: Read all auto-loaded files. Identify redundancy, prose overhead, and duplicate rules.\n\n**Ask user before proceeding: \"Want to distill?\"**\n\n### Stage 2: DISTILL — Convert with Safety Net\n\n1. **Backup**: `cp ~/.claude/CLAUDE.md ~/.claude/CLAUDE.md.bak-pre-distill`\n2. **Phase 1-5**: Run the full conversion process above\n3. **Phase 6**: Run multi-model test (minimum 2 models, 8 questions)\n4. **Report**: Show before/after scores\n\n```\n=== AI.MD Conversion Complete ===\n\nBefore: {old} bytes ({old_score} compliance)\nAfter:  {new} bytes ({new_score} compliance)\nSaved:  {percent}% bytes, +{delta} compliance points\n\nBackup: ~/.claude/CLAUDE.md.bak-pre-distill\nRestore: cp ~/.claude/CLAUDE.md.bak-pre-distill ~/.claude/CLAUDE.md\n```\n\n---\n\n## AI-Native Template\n\n```xml\n# PROJECT-NAME | lang:xx | for-AI-parsing | optimize=results-over-format\n\n<user>\nidentity, tone, signals, decision-style (key: value pairs)\n</user>\n\n<gates label=\"硬性閘門 | 優先序: gates>rules>rhythm | 缺一項=STOP\">\n\nGATE-1 name:\n  trigger: ...\n  action: ...\n  exception: ...\n  yields-to: ...\n\nGATE-2 name:\n  trigger: ...\n  action: ...\n  policy: ...\n\n</gates>\n\n<rules>\n\nRULE-NAME:\n  core: ...\n  banned: ...\n  hear-X: ... → action\n  violation: ...\n\n</rules>\n\n<rhythm>\nworkflow patterns as key: value pairs\n</rhythm>\n\n<conn>\nconnection strings (keep exact — NEVER compress facts/credentials/URLs)\n</conn>\n\n<ref label=\"on-demand Read only\">\nfile-path → purpose\n</ref>\n\n<learn>\nhow system evolves over time\n</learn>\n```\n\n---\n\n## Anti-Patterns\n\n| Don't | Do Instead | Why |\n|-------|------------|-----|\n| Human prose in CLAUDE.md | Structured labels | Prose requires inference; labels are direct |\n| Multiple rules on one line | One concept per line | Attention splits across dense lines |\n| Parenthetical explanations | Remove them | AI needs \"what\" not \"why\" |\n| Same rule in 3 places | Single source + cross-ref | Duplicates can diverge and confuse |\n| 20+ flat rules | 5-7 domains with sub-items | Hierarchy helps model organize behavior |\n| Compress without testing | Validate with 2+ models | What works for Claude might fail for GPT |\n| Assume format doesn't matter | Test it — it does | Same content, different format = different compliance |\n| Chinese-only labels | English labels + native output | English labels are more universal across models |\n| Flat rule list | State machine with priorities | Clear execution order prevents missed rules |\n\n---\n\n## Real-World Results\n\nTested 2026-03, washinmura.jp CLAUDE.md, 5 rounds, 4 models:\n\n| Round | Change | Codex (GPT-5.3) | Gemini 2.5 Pro | Claude Opus 4.6 |\n|-------|--------|-----------------|----------------|-----------------|\n| R1 (baseline prose) | — | 8/8 | 7/8 | 8/8 |\n| R2 (added rules) | +gates +examples | 7/8 | 6/8 | — |\n| R3 (refined prose) | +exceptions +non-triggers | 6/8 | 6.5/8 | — |\n| R4 (AI-native convert) | structured labels | **8/8** | **7/8** | **8/8** |\n\nKey findings:\n1. **More prose rules = worse compliance** (R1→R3: scores dropped as rules grew)\n2. **Structured format = restored + exceeded** (R4: back to max despite more rules)\n3. **Cross-model consistency**: Format that works for one model works for all (except Grok)\n4. **Semantic anchoring**: The `new-api:` label fix was the single most impactful change\n\n**The uncomfortable truth: Your beautiful, carefully-written CLAUDE.md\nmight be HURTING your AI's performance. Structure > Prose. Always.**\n\n## Limitations\n- Use this skill only when the task clearly matches the scope described above.\n- Do not treat the output as a substitute for environment-specific validation, testing, or expert review.\n- Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.","tags":["antigravity","awesome","skills","sickn33","agent-skills","agentic-skills","ai-agent-skills","ai-agents","ai-coding","ai-workflows","antigravity-skills","claude-code"],"capabilities":["skill","source-sickn33","skill-ai-md","topic-agent-skills","topic-agentic-skills","topic-ai-agent-skills","topic-ai-agents","topic-ai-coding","topic-ai-workflows","topic-antigravity","topic-antigravity-skills","topic-claude-code","topic-claude-code-skills","topic-codex-cli","topic-codex-skills"],"categories":["antigravity-awesome-skills"],"synonyms":[],"warnings":[],"endpointUrl":"https://skills.sh/sickn33/antigravity-awesome-skills/ai-md","protocol":"skill","transport":"skills-sh","auth":{"type":"none","details":{"cli":"npx skills add sickn33/antigravity-awesome-skills","source_repo":"https://github.com/sickn33/antigravity-awesome-skills","install_from":"skills.sh"}},"qualityScore":"0.700","qualityRationale":"deterministic score 0.70 from registry signals: · indexed on github topic:agent-skills · 34964 github stars · SKILL.md body (17,650 chars)","verified":false,"liveness":"unknown","lastLivenessCheck":null,"agentReviews":{"count":0,"score_avg":null,"cost_usd_avg":null,"success_rate":null,"latency_p50_ms":null,"narrative_summary":null,"summary_updated_at":null},"enrichmentModel":"deterministic:skill-github:v1","enrichmentVersion":1,"enrichedAt":"2026-04-25T00:50:24.587Z","embedding":null,"createdAt":"2026-04-18T21:30:41.997Z","updatedAt":"2026-04-25T00:50:24.587Z","lastSeenAt":"2026-04-25T00:50:24.587Z","tsv":"'-03':2415 '-1':368,387,1258,1288,1300,1576,1598,1638,1819,2232 '-2':1584,1822,2241 '-3':408,1264,1286,1296,1593,1813 '-4':1302,1324,1331,1346,1400,1424,1605,1646 '-5':525,1934,1954,1971,2151 '-5.3':2426 '-7':2340 '/.claude/claude.md':2067,2146,2202 '/.claude/claude.md.bak-pre-distill':2147,2198,2201 '/.claude/rules':2074 '/8':2455 '/dev/null':2069,2077 '/endpoint':494 '/source':310 '0':2071,2081 '1':149,191,583,613,722,740,797,1141,1494,1657,1709,2052,2143,2150,2468 '12':886 '2':331,623,713,806,1144,1459,1507,1754,2068,2076,2137,2148,2167,2356,2481 '2.5':1419,2428 '20':2336 '2026':2414 '3':442,633,734,748,815,865,1148,1174,1515,1841,1931,2158,2324,2493 '4':21,153,644,779,793,828,837,1045,1153,1525,1608,1900,2088,2171,2420,2509 '4.6':2432 '5':247,537,652,1158,1214,1539,1978,2339,2418 '50':2106,2110 '6':147,158,160,165,1163,1440,1544,2160 '6.5':2454 '6/8':2445,2453 '7':1135,1167,1552 '7/8':2437,2444,2464 '8':1496,1566,2169 '8/8':2436,2438,2463,2465 'a/b':832,862 'accept':835,1628 'account':976 'across':20,205,417,540,958,1487,1751,2309,2394 'act':1067 'action':382,393,414,627,681,905,1062,1696,1730,1783,2018,2020,2235,2244,2254 'activ':620,899,1032 'actual':177,1684,2045 'ad':126,1431,2440 'add':1395 'addit':1613 'ai':2,11,34,50,83,114,151,162,355,377,440,482,631,641,691,743,908,1006,1243,1275,1367,1535,1615,1623,1635,1687,1701,1804,1995,2204,2215,2316,2458,2537 'ai-md':1 'ai-nat':10,33,2203,2457 'ai.md':29,91,92,580,1482,2176 'all-claims-ne':305 'all-claims-need-data':1189 'alon':2114 'alreadi':1352 'alway':2542 'ambigu':385 'analysi':665,790,1602 'anchor':444,499,2511 'answer':1547 'anti':2279 'anti-pattern':2278 'anyth':1037,1107 'api':503,509,519,547,1612,1965,2515 'appear':1911 'appli':365,1836 'architectur':1049 'arriv':1810,1826 'ask':612,818,2129,2576 'assign':867 'assum':2366 'assumpt':1369 'atom':719,838,871 'attempt':1581 'attend':185 'attent':192,203,216,238,245,285,2307 'attribut':1877 'auto':2119 'auto-load':2118 'back':2487 'backstori':2012 'backup':1145,1200,1266,1281,2144,2197 'bad':1134,1925 'ban':298,801,840,978,1186,1325,1714,2250 'baselin':2434 'bash':2058 'battl':18,1654 'battle-test':17,1653 'beauti':2528 'becom':1126 'before/after':2174 'before/after/during':970 'behavior':622,913,1063,1503,2350 'belong':1868 'bilingu':1658 'bleed':1859 'bottom':1116 'boundari':1847,1854,2584 'brake':1336 'break':715,975 'bug':1308,1316,1338,1351 'bug-clos':1307,1315,1337 'build':600,1047 'buri':481 'burn':2062 'busi':1611 'byte':2093,2098,2102,2181,2187,2193 'c':2066,2079 'cannot':1244 'care':2530 'carefully-written':2529 'case':534,921,1220 'cat':2073 'catch':1585 'caus':1313,1321,1601 'certain':798 'chang':188,1199,1564,2423,2523 'check':497,522,1059,1105,1154,1205,1250,1811,1817,1820,1928,1935,1939,1956,1960,1968 'check-health':1204 'chines':1703,1742,2382 'chinese-on':2381 'chinese/japanese/english':427 'cite':1304 'claim':307,830,1169,1191,1327 'clarif':2578 'claud':85,1471,2063,2083,2091,2361,2430 'claude.md':8,46,101,565,595,2090,2289,2417,2532 'clear':1026,1806,2403,2551 'close':1309,1317,1339 'codex':86,549,2424 'come':1099 'compar':860 'comparison':833,1625 'compil':588 'complet':32,2178 'complianc':28,77,133,725,1549,2184,2190,2195,2380,2473 'compound':728 'compress':1078,2267,2351 'concept':1735,1910,2304 'conclus':1335 'confirm':1380,1639 'conflict':944,1217,1236,1246,1256,1268,1314,1365,1514,1800 'confus':1473,1863,2038,2335 'connect':490,1074,2262 'consecut':538 'consequ':973 'consist':2497 'constraint':637,935,971,1711 'contain':457,1234 'content':167,2376 'context':345 'convers':36,556,963,1454,1560,2155,2177 'convert':4,97,134,575,2139,2460 'core':287,912,1182,2249 'correspond':467 'could':1390,2037 'courtesi':1291 'cp':2145,2200 'creat':449,1852 'criteria':2587 'critic':1225 'cross':1902,1922,1951,1973,2329,2495 'cross-model':2494 'cross-ref':1972,2328 'cross-refer':1901,1921,1950 'curl':270,313 'curl/benchmark':323,854 'current':2060 'data':803,822,1193,1741 'data/line':309 'databas':1579 'decid':1378 'decis':987,2226 'decision-styl':2225 'declar':347,877,892 'decompos':714 'decomposit':675 'decreas':132 'definit':1326 'delet':662,687,1984,2002,2029 'delimit':1851 'delta':2194 'demand':1081 'dens':2310 'deploy':1157 'describ':2555 'despit':2490 'detect':1247 'develop':1493 'differ':168,170,1005,1460,2377,2379 'direct':380,464,504,1733,2297 'discov':1416,1651 'distil':2135,2138 'distinct':235 'distribut':204 'diverg':2333 'document':607 'doesn':422,692,2368 'domain':1125,1175,1180,1195,1208,2341 'doubt':317,808,848 'drop':251,1558,2477 'duplic':1906,1976,2127,2331 'durabl':957 'echo':2059,2070,2080,2089,2094,2099,2105 'edg':1219 'edit':1147,1272,1590 'els':1038,1108 'english':1663,1671,1713,1726,2385,2389 'env':1273 'environ':2567 'environment-specif':2566 'equal':208,1140 'everi':700,716,870,901,1251,1453,1737,2028 'every-new-task':699 'everyth':189 'evid':257,286,1179,1305,1354 'evolut':1088 'evolv':2275 'exact':569,2265 'exam':1490 'exampl':664,927,1398,1506,2443 'exceed':143,2485 'except':383,400,415,651,703,708,914,1731,2236,2449,2507 'execut':1371,1807,2404 'exist':661,1944,1988,1994 'expert':2572 'explain':1990 'explan':657,2005,2030,2313 'explicit':349,925,1297,1396,1799 'extract':2027 'fabric':260,290 'fact':1076 'facts/credentials/urls':2268 'fail':536,2363 'failur':726 'featur':513 'fewer':25,120 'figur':1832 'file':1589,2121,2270 'file-path':2269 'find':792,2467 'fire':1594,1789 'first':395,553,684,1100,1267,1279,1282,1595,1698,1790,1814,2022 'first-sent':394,683,1697,2021 'five':279 'fix':531,1436,2517 'flat':1133,1764,2337,2396 'flow':964 'follow':116,152,163 'for-ai-pars':2213 'forbidden':979,1717 'forc':338 'format':16,112,140,169,581,932,934,1209,1210,1466,1704,2221,2367,2378,2483,2498 'fraction':213 'full':237,284,478,1548,2154 'full-text':477 'function':868,879,1017 'gate':367,386,407,524,949,954,1098,1257,1263,1285,1287,1295,1299,1301,1323,1330,1345,1399,1423,1575,1583,1592,1597,1604,1637,1645,1757,1773,1780,1785,1794,1812,1818,1821,1871,1892,1933,1953,1970,2231,2240,2442 'gate/rule':902 'gates/rules':1254 'gemini':87,1418,2427 'get':211,220,283,873 'give':562,1803 'go':984 'good':1173,1945 'gpt':1474,2365,2425 'grammar':428,1743 'grew':2480 'grok':88,2508 'group':1119,1172 'guess':263,275,293,326,1143,1185,1357 'guidelin':1064 'handl':1216 'happen':559 'hard':981,1056,1853,1891 'hash':471 'head':1132 'health':1155,1206,1927 'health-check':1926 'health-check.py':1617,1936,1940,1943 'hear':316,807,847,2252 'hear-doubt':315,846 'hear-x':2251 'help':1003,2347 'heurist':988 'hidden':794,1235 'hide':778 'hierarchi':1055,1178,2346 'high':61 'higher':27 'highest':1815 'hint':1528 'human':6,99,145,438,591,656,689,1238,2004,2286 'human-written':5,98 'hurt':2535 'ident':2222 'identifi':2122 'ignor':52 'immedi':1372,1438 'impact':2522 'improv':1093 'includ':1508,1516 'independ':1543 'infer':334,342,351,357,2294 'input':456,619,666,775,898,1498,1777,1809,1825,2581 'insight':410 'instead':474,804,1758,1823,1904,2284 'instruct':65,106,179,578,741,750,1012,1112,1232,1888,2019,2113 'inter':953 'inter-g':952 'intuit':1241 'item':448,1129,2345 'judgment':990 'justif':2014 'keep':2015,2264 'kept':1421 'key':409,2228,2259,2466 'keyword':459 'know':1538 'label':15,111,157,335,346,379,411,432,445,468,500,544,866,869,875,887,889,994,1000,1023,1043,1659,1661,1672,1727,1744,1876,1924,2291,2295,2384,2386,2390,2462,2516 'lang':2211 'languag':131,337,419,436,577,1231,1670,1695,1722 'least':1227 'less':1749 'level':1887 'like':412,469,586,739,1429 'limit':2543 'line':150,161,199,228,244,280,732,774,1708,2302,2306,2311 'list':985,1385,1765,2398 'llm':73,104,1461 'llms':176,180 'load':1085,2120 'long':48 'look':738,823 'lookup':473 'lost':221 'm':599 'machin':603,1756,1771,2400 'map':1732 'match':462,505,1793,2552 'matchabl':450 'matrix':1248 'matter':1097,2370 'max':2489 'mayb':1837 'md':3,2064,2075,2084,2092,2096 'mean':343,348 'measur':2054 'mechan':190,330,441,977 'mental':570 'mention':1616,1929 'metadata':647,696 'metaphor':2010 'methodolog':95 'metric':1642 'might':487,1472,2362,2533 'migrat':80 'minimum':2166 'misinterpret':354 'miss':488,1551,1838,1897,2407,2589 'moat':510,1937,1962 'model':22,115,201,230,340,421,461,542,571,1007,1110,1138,1176,1444,1462,1488,1542,1555,1681,1738,1752,1766,1808,1862,2040,2164,2168,2348,2357,2395,2421,2496,2503 'modifi':361 'motiv':690,2007 'multi':1443,2163 'multi-model':1442,2162 'multipl':195,1792,1913,2298 'must':356,770,909,1058,1103,1455,1689,1706,1938 'must-be-lin':1705 'name':2210,2233,2242,2248 'nativ':12,35,2205,2387,2459 'natur':130,336,576,1230 'near':254 'near-zero':253 'need':308,352,424,694,745,903,992,1087,1170,1192,1996,2317 'negat':926,1397,1517 'negoti':1448 'net':2142 'never':1077,2266 'new':391,502,512,518,546,679,701,1964,2186,2188,2514 'new-api':501,517,545,1963,2513 'new-featur':511 'new-task':390,678 'no-fabr':258,288 'no-go':982 'no-guess':261,291,1183,1355 'nois':2035 'non':1447,2451 'non-negoti':1446 'non-trigg':2450 'not-trigg':922,1347,1382,1406,1432 'number':1212,1427 'obvious':1228 'often':1233 'old':2180,2182 'on-demand':1079 'one':198,243,776,904,1010,1131,1835,1839,2301,2303,2502 'open':1880 'optim':71,2217 'option':1452 'opus':2431 'order':1096,2405 'organ':1051,2349 'origin':1721 'output':836,933,1152,1207,1664,1702,2388,2561 'over-trigg':929,1391 'overhead':2125 'overrid':920,940 'pair':1252,2230,2261 'paradox':123 'parenthet':2312 'pars':426,2216 'part':761 'partial':1550 'path':2271 'pattern':1069,2257,2280 'per':2305 'percent':2192 'perfect':1469 'perform':2539 'permiss':2582 'persist':398,650,698,956 'phase':582,712,864,1044,1213,1229,1439,2149,2159 'pick':998 'place':1932,2325 'point':1480,2196 'polici':986,2245 'posit':936,1118 'post':1342 'post-verif':1341 'pre':1198,1202,1334 'pre-chang':1197 'pre-conclus':1333 'pre-run':1201 'preced':951 'prefer':829 'prevent':928,1856,2406 'preview':2053 'principl':1983 'prioriti':648,939,955,1117,1284,1787,1816,1896,2402 'pro':1420,2429 'proceed':2132 'process':178,231,557,1111,2156 'produc':1688 'project':2209 'project-nam':2208 'prompt':75 'proof':304,527,1171,1188 'prose':146,2124,2287,2292,2435,2448,2470,2541 'protect':1588 'protocol':1491 'prove':125 'purpos':2272 'put':757 'py覆蓋':498,523,1961,1969 'queri':1428,1643 'question':1510,1567,1609,2170 'r1':2433,2474 'r2':2439 'r3':2446,2475 'r4':2456,2486 'rais':548 'read':183,378,585,593,605,1036,1827,2116 'read/grep':268,311 'real':526,1501,2410 'real-world':2409 'record':1344 'redund':2123 'ref':1974,2330 'refactor':2025 'refer':1082,1903,1923,1952 'refin':2447 'relationship':941 'reliabl':118 'remov':2034,2314 'repeat':1260,1276,1917 'report':1166,2172 'requir':831,938,1318,1606,2293,2580 'resolut':1283,1328,1373,1801 'resolv':1215,1239 'restor':141,2199,2484 'result':171,2219,2412 'results-over-format':2218 'revert':1561 'review':2573 'right':1042 'root':1312,1320,1600 'round':2419,2422 'rule':24,54,81,128,137,148,159,196,210,219,224,242,250,282,660,720,729,735,768,780,795,796,805,814,827,839,872,943,960,1015,1034,1052,1089,1121,1137,1388,1513,1521,1530,1761,1829,1858,1867,1873,1914,1993,2072,2085,2095,2097,2128,2247,2299,2322,2338,2397,2408,2441,2471,2479,2492 'rule-ble':1857 'rule-nam':2246 'run':810,1203,1540,1624,2152,2161 'safeti':1196,1289,2141,2583 'save':2031,2191 'saw':1020 'say':296,507,1161,1271,1360,1620,1632 'say-so':295 'scan':483 'scope':1194,1941,2554 'score':1545,1557,2175,2183,2189,2476 'search':480 'section':1843,1850,1865,1886 'section-level':1885 'select':995 'self':321,812,852,856 'self-serv':855 'self-test':320,811,851 'semant':443,1846,2510 'sentenc':396,486,610,685,764,777,1699,2023 'separ':736,749,767,773 'serv':857,1882 'session':2108 'shame':277,328 'shame-wal':276,327 'share':197,1123 'shorter':1674 'shouldn':1536 'show':2173 'signal':401,709,1374,2224 'signals命中':374,672,705 'simpl':1426,1572,1641 'simul':1500 'singl':731,1946,2326,2520 'skill':42,2546 'skill-ai-md' 'skip':403,711,1379,1636 'sourc':723,1947,2327 'source-sickn33' 'special':1649 'specif':529,1563,2568 'split':193,246,2308 'splitter':752 'stage':2049,2051,2136 'standard':883,1750 'state':602,1319,1755,1770,1778,2399 'stay':1690,1719 'still':51 'stop':1057,1898,2574 'strategi':1660 'string':1075,1665,2263 'structur':14,110,139,156,431,937,1046,1710,2290,2461,2482,2540 'structured-label':13,109 'style':2227 'sub':447,1128,2344 'sub-item':446,1127,2343 'substitut':2564 'success':2586 'surviv':962 'system':37,64,74,105,1092,2274 't1':1571 't2':1578 't3':1587 't4':1599 't5':550,1610 't6':1618 't7':1630 't8':1640 'tabl':472,1150,1211 'tag':451,1844,1881 'take':950 'task':392,680,702,1261,1278,1573,2550 'techniqu':530,996,1120,1245,1381,1650,1656,1753,1840,1899,1977 'tell':1027 'templat':1568,2206 'test':19,322,533,753,813,853,1164,1441,1497,1518,1533,1655,2165,2353,2371,2413,2570 'text':479,1685,1986 'textbook':1505 'time':539,649,965,1329,2277 'token':26,57,121,207,2032,2061,2086,2103,2109,2111 'tokens/turn':2104 'tone':2223 'tool':84 'top':1114 'top-to-bottom':1113 'topic-agent-skills' 'topic-agentic-skills' 'topic-ai-agent-skills' 'topic-ai-agents' 'topic-ai-coding' 'topic-ai-workflows' 'topic-antigravity' 'topic-antigravity-skills' 'topic-claude-code' 'topic-claude-code-skills' 'topic-codex-cli' 'topic-codex-skills' 'total':2082,2087,2100,2101 'touch':2057 'train':1740 'transit':1784 'trap':1509 'treat':1139,1760,2559 'tri':554,1830 'trigger':381,389,413,617,676,896,924,931,1025,1349,1384,1393,1402,1408,1422,1434,1524,1577,1648,1729,1776,2234,2243,2452 'truth':1181,1949,2526 'turn':959,2107 'two':760,1512,2048 'two-stag':2047 'type':888 'uncomfort':2525 'understand':186,430,584,1013,1177 'understood':1678 'unit':236 'univers':435,1677,2393 'unrel':1136 'unsur':294 'usag':58 'use':40,43,55,66,78,119,573,802,881,895,1149,1570,1848,1919,2544 'user':454,506,820,1270,1359,1376,1502,1619,1631,1668,1693,2130 'v4':30 'valid':1445,1457,2354,2569 'valu':2229,2260 'verbos':63 'verif':1343 'verifi':1353,1368,1942 'versa':1477 'vice':1476 'violat':325,972,2255 'vocabulari':884,1728 'vs':1262,1306,1358 'wall':278,329 'want':69,2133 'washinmura.jp':2416 'way':248 'wc':2065,2078 'weight':217,239,256 'whole':485,1479 'without':1035,2352 'word':799,809,1718 'words/actions':980 'work':174,416,1468,1486,1725,2359,2500,2504 'workflow':969,1068,2050,2256 'world':2411 'wors':2472 'would':1002 'write':1311,1495,1580 'written':7,100,2531 'x':2253 'xml':1842,1870,2207 'xx':2212 'yield':405,946,1293,1797,2238 'yields-to':404,945,1292,1796,2237 'zero':255,333,384 'zero-infer':332 '不可逆建議':1405 '不是問user':788 '不是大爆破':2008 '中間進度數字':1409 '你要我做的是':397,686,1700,2024 '例外':364,373,671,704 '例外情況':1747 '保護檔':1265 '先拿數據':267,303,784,845 '先用一句話複述':370,668,682 '加個':508 '加新功能':491 '可能是':266,300,783,842,1716 '同區塊連續第3次修改':2026 '因為用戶是非工程師':2013 '執行動作':1746 '報結論':1303,1401 '契約':515 '好像':272,318,785,849 '好像x比y好':1621 '實測先行':863 '工具原始輸出':1411 '已知事實':1412 '必確認health':496,521,1967 '必確認health-check':495,520,1966 '想清楚100倍後才做現在的':2011 '感覺是':301,843 '應該是':265,299,782,841,1162,1187,1715 '成功率怎麼樣':1430 '推測':302,844 '收到任務':369,667,677 '數據':271,314 '新增api':493,1957 '是每次順手一點':2009 '最終歸因':1403 '有數據才能決定':789 '根因判定':1404 '直接執行':376,674,707 '確認health':1959 '確認health-check':1958 '禁反問user':324,858 '禁用詞':264,781 '第一句問schema':492,514 '純指標查詢':1410,1435 '自己先跑test':274,787 '自己查':859 '處理一下':375,402,673,706,710,1361,1375,1633 '行號':269,312 '複述':388,1259 '覺得':273,319,786,850 '覺得a比b好':861 '觸發條件':1745 '記錄根因':1310 '轉述文件':1413 '長對話中每個新任務都重新觸發':372,399,670,697 '關聯':516 '防搞混':360,371,669,688,2006 '驗收':1955","prices":[{"id":"e1d9fd8c-51d9-4466-9995-a84245305bda","listingId":"17832d92-3647-4e11-ab01-ecd79e8e50f3","amountUsd":"0","unit":"free","nativeCurrency":null,"nativeAmount":null,"chain":null,"payTo":null,"paymentMethod":"skill-free","isPrimary":true,"details":{"org":"sickn33","category":"antigravity-awesome-skills","install_from":"skills.sh"},"createdAt":"2026-04-18T21:30:41.997Z"}],"sources":[{"listingId":"17832d92-3647-4e11-ab01-ecd79e8e50f3","source":"github","sourceId":"sickn33/antigravity-awesome-skills/ai-md","sourceUrl":"https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/ai-md","isPrimary":false,"firstSeenAt":"2026-04-18T21:30:41.997Z","lastSeenAt":"2026-04-25T00:50:24.587Z"}],"details":{"listingId":"17832d92-3647-4e11-ab01-ecd79e8e50f3","quickStartSnippet":null,"exampleRequest":null,"exampleResponse":null,"schema":null,"openapiUrl":null,"agentsTxtUrl":null,"citations":[],"useCases":[],"bestFor":[],"notFor":[],"kindDetails":{"org":"sickn33","slug":"ai-md","github":{"repo":"sickn33/antigravity-awesome-skills","stars":34964,"topics":["agent-skills","agentic-skills","ai-agent-skills","ai-agents","ai-coding","ai-workflows","antigravity","antigravity-skills","claude-code","claude-code-skills","codex-cli","codex-skills","cursor","cursor-skills","developer-tools","gemini-cli","gemini-skills","kiro","mcp","skill-library"],"license":"mit","html_url":"https://github.com/sickn33/antigravity-awesome-skills","pushed_at":"2026-04-24T06:41:17Z","description":"Installable GitHub library of 1,400+ agentic skills for Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity, and more. Includes installer CLI, bundles, workflows, and official/community skill collections.","skill_md_sha":"8deb9cc84c0c02f7090bb26400d5f6d7446cb7ee","skill_md_path":"skills/ai-md/SKILL.md","default_branch":"main","skill_tree_url":"https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/ai-md"},"layout":"multi","source":"github","category":"antigravity-awesome-skills","frontmatter":{"name":"ai-md","description":"Convert human-written CLAUDE.md into AI-native structured-label format. Battle-tested across 4 models. Same rules, fewer tokens, higher compliance."},"skills_sh_url":"https://skills.sh/sickn33/antigravity-awesome-skills/ai-md"},"updatedAt":"2026-04-25T00:50:24.587Z"}}