[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6C397V7JEN45CVWECA2MBZH":3,"skill-same:01M6C397V7JEN45CVWECA2MBZH":76},{"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":22,"publishedAt":50,"linkOnly":51,"markdownContent":52,"markdownBody":53,"frontMatter":54,"refName":55,"commitSha":56,"skillDirectory":57,"forksCount":58,"repositoryHtmlUrl":59,"repositoryHomepageUrl":6,"ownerHtmlUrl":60,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":61,"install":62,"resources":75},"01M6C397V7JEN45CVWECA2MBZH","llm-model-capability-audit",null,"Audit which LLM models actually support the capabilities claimed in their config (code execution, structured JSON, tool calling) BEFORE promoting them to defaults, allowlists, or dropdowns. Covers Gemini code_execution, OpenAI reasoning_token footprint, OpenRouter routing fallbacks, and tail-latency cliffs. Triggers on any \"make X the default model\", \"add model Y to dropdown\", \"is this model safe to enable for dice\u002Fnarrative\u002Fdice?\", or \"we just shipped a regression on model Z\" question.","jleechanorg\u002Fjleechan-skills","jleechanorg","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F221404966?v=4","01M6C1VJ99XG950X1DWPEBCBGX","01M6C1VJDK5YVDNT3362RT5WSB",6,"MIT","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002Fhermes\u002Fskills\u002Fdevops\u002Fllm-model-capability-audit\u002FSKILL.md",[17],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49],"model","capability","llm model capability audit","audit","openrouter","gemini","add model","which","llm models","narrative","add","token","default","support","question","just","routing","regression","safe","json","tool","execution","config","openai","capabilities","reasoning","code","2026-11-10 10:12:30",false,"---\nname: llm-model-capability-audit\ndescription: Audit which LLM models actually support the capabilities claimed in their config (code execution, structured JSON, tool calling) BEFORE promoting them to defaults, allowlists, or dropdowns. Covers Gemini code_execution, OpenAI reasoning_token footprint, OpenRouter routing fallbacks, and tail-latency cliffs. Triggers on any \"make X the default model\", \"add model Y to dropdown\", \"is this model safe to enable for dice\u002Fnarrative\u002Fdice?\", or \"we just shipped a regression on model Z\" question.\n---\n\n# LLM model capability audit\n\n> **What this skill is:** the pre-promotion gate that catches \"model claims X in marketing docs but actually returns `TOO_MANY_TOOL_CALLS` \u002F empty content \u002F 30k reasoning tokens on a trivial prompt\" regressions BEFORE they ship as a default. Distinct from `llm-model-upgrade-survey` (which is for surveying new options); this is for verifying that a model actually does what the catalog says before flipping the default.\n\n**Read first:** canonical examples in `references\u002Fworldarchitect-gemini-flash-lite-2026-07-30.md` — the live evidence behind this skill.\n\n---\n\nFor new Gemini selections or promotions, only consider Gemini 3.8 Flash or later; never promote Gemini Pro or an earlier Gemini model. Older model IDs below describe historical findings.\n\n## When to use this skill\n\nUse it whenever ANY of the following lands in your session:\n\n- \"make X the new default model\"\n- \"add Y to the model dropdown \u002F settings \u002F ALLOWED_MODELS\"\n- \"why is the dice tool slow \u002F using the wrong path?\"\n- \"we just shipped a regression — users picked model X and got different behavior\"\n- A PR title like \"feat: switch default to X\" \u002F \"feat: add Y to allowlist\"\n- A code change touching `MODELS_WITH_CODE_EXECUTION`, `ALLOWED_MODELS`, `DEFAULT_*_MODEL`, or equivalent per-provider allowlists\n- A user report of \"the LLM is taking forever\" \u002F \"model returned empty\" \u002F \"cost jumped 5× after we changed defaults\"\n\n**Do NOT use it** for: pure creative prose evaluation, narrative-quality review, brand-tone critique — those are not capability audits.\n\n---\n\n## The audit workflow\n\n### Phase 1 — Catalog snapshot (no LLM calls, 5 minutes)\n\nFor every model the change touches, capture:\n- Catalog entries (`ALLOWED_MODELS`, `MODELS_WITH_CODE_EXECUTION`, model mapping tables, frontend dropdown options)\n- The **claimed** capability comment for each (e.g. \"inherits 3.x code execution + JSON single-pass\")\n- Whether the model is in any allowlist that gates critical features (dice, structured output, code execution)\n\nCross-reference: is every model in the dropdown also in the relevant allowlists? If not, **that's the latent bug** — users picking it silently degrade.\n\nOutput: a verdict table with 4 columns (model, in dropdown, in allowlist, capability claim).\n\n### Phase 2 — Direct API probe (real LLM calls, 10 minutes per model)\n\nFor each model the change touches, hit the **real provider API** (not just docs) with the EXACT request shape the app sends:\n\n- Same tools array (e.g. `{\"code_execution\": {}}` for Gemini)\n- Same response format\n- Same temperature + system prompt structure\n- Same approximate prompt length the app sends (often much larger than test prompts)\n\nCapture: `finish_reason`, `code_parts` \u002F tool-call count, `total_tokens`, `usage.completion_tokens_details.reasoning_tokens`, and the actual text output. **Critical: reasoning_tokens count on trivial prompts**. A model that emits 200 reasoning tokens on a one-line JSON request is 5-10× more expensive than its base rate suggests.\n\n**Auth-source priority** (in order, per `dual-probe-secrets` SOUL.md rule):\n1. `os.environ.get(\"PROVIDER_API_KEY\")` from bashrc-sourced shell\n2. macOS Keychain via `security find-generic-password -s \u003Cservice> -w`\n3. GCP Secret Manager via `gcloud secrets versions access latest --secret=\u003Cname> --project=\u003Cproject>`\n4. Local secret file (`~\u002F.gemini_api_key_secret`, etc.)\n\nNever trust a single source — dual-probe both local file AND GCP secret-manager when both exist.\n\n### Phase 3 — Latency benchmark (12 turns minimum, 10 minutes)\n\nFor latency-sensitive decisions (default model selection, dice provider choice), run 12 turns per model with a representative prompt:\n- Capture mean, p50, p95 latency\n- Capture cost per 12 turns\n- Capture total tokens + reasoning tokens\n- Capture finish_reason distribution (any non-`stop` is a smell)\n\n**Tail-latency cliff check:** if p95 > 3× mean, the model has unstable latency under load. Worth flagging.\n\nFor structured-output workloads (dice, JSON schema validation), also run a `response_format={\"type\": \"json_object\"}` test — some models accept the parameter but emit markdown-wrapped JSON instead.\n\n### Phase 4 — Promotion gate (3 questions)\n\nBefore approving any \"make X the default\" PR, answer:\n1. **Is X in every allowlist that gates critical features?** (dice, code execution, structured JSON)\n2. **Does the direct-API probe show clean finish + correct output + reasonable token footprint?**\n3. **Is the p95 latency within 2× the existing default?**\n\nIf any answer is \"no\" — the change is NOT safe to ship. Either patch the allowlist first, fix the model choice, or document the limitation in the catalog comment.\n\n---\n\n## Common pitfalls (verified in production)\n\n### Pitfall 1 — Allowlist drift after model add\n\nWhen a PR adds a new model to `ALLOWED_MODELS` but forgets to add it to `MODELS_WITH_CODE_EXECUTION`, users picking that model silently get a slower fallback path (e.g. native two-phase dice instead of code-execution dice). Symptom is silent: the LLM still returns valid JSON, just via a slower + more expensive code path.\n\n**Defense:** Phase 1 catalog snapshot must cross-check every allowlist the model affects.\n\n### Pitfall 2 — Comment claims capabilities that don't exist\n\nA code comment like \"inherits 3.x code execution + JSON single-pass\" is documentation debt. If the model doesn't actually inherit the capability (e.g. `gemini-3.5-flash-lite` loops on code-execution calls), the comment becomes a lie that future agents trust. Always verify the comment against the direct-API probe.\n\n**Verified example:** PR #8512 (your-project.com) added `gemini-3.5-flash-lite` and `gemini-3.6-flash` with the claim that both inherit code execution. Direct API test: both return `TOO_MANY_TOOL_CALLS` with 48 code_parts and 20-27k tokens of empty output.\n\n### Pitfall 3 — Reasoning-token footgun\n\n\"Fast\" tier models often emit reasoning tokens on trivial prompts. `gemini-3.5-flash` emits 220 reasoning tokens on a 1-line JSON request. `gemini-3.6-flash` emits 237. That makes them 20-30× more expensive per turn than their headline price suggests. Always capture `usage.completion_tokens_details.reasoning_tokens` separately.\n\n### Pitfall 4 — Default flip without allowlist patch\n\nPR #8571 (your-project.com) made `gemini-3.5-flash-lite` the default. The model wasn't in `MODELS_WITH_CODE_EXECUTION`, so every user silently got native two-phase dice. The PR was reverted 8 hours later by PR #8590, which only fixed the **default**, not the allowlist. The latent bug persists for users who explicitly pick Lite or 3.6.\n\n**Defense:** Phase 4 promotion gate must explicitly ask \"does this model work for every allowlist that the current default works in?\"\n\n### Pitfall 5 — Tail-latency cliff\n\n`gemini-3.6-flash` has mean=3.20s, p50=2.06s, p95=**11.84s** on a 12-turn dice benchmark. The model is \"fine\" on average and terrible on the 5th percentile. Always run p95, not just mean.\n\n### Pitfall 6 — Bashrc-vs-Keychain vs GCP secret-manager divergence\n\nThe \"where's the key\" question has 4 valid answers and they often disagree:\n- bashrc export (often unset or commented out for security)\n- macOS Keychain via `security find-generic-password`\n- GCP Secret Manager via `gcloud secrets versions access`\n- Local secret file (`~\u002F.gemini_api_key_secret` style)\n\n`dual-probe-secrets` SOUL.md rule: probe at least 2 of these before declaring \"key is broken\". The local secret file in this repo was stale (`AIzaSyDQ6a...` returns `API_KEY_INVALID`) but the GCP secret-manager version (`AIzaSyAvyb...`) was live.\n\n### Pitfall 7 — Codex is a UI, not an API\n\nGPT-6 Luna is only exposed in the Codex picker — but the underlying API is OpenRouter. `POST https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1\u002Fchat\u002Fcompletions` with `model=\"openai\u002Fgpt-6-luna\"` works directly. Same for any \"Codex-only\" model: it's almost always OpenRouter under the hood.\n\n### Pitfall 8 — `claudem` worker with stale `.agent_prompt_*.txt`\n\nWhen dispatching a `claudem -p` worker via `git worktree add`, the worktree may inherit a stale `.agent_prompt_\u003Cprior-branch>.txt` from a prior session. The worker reads the stale prompt and either loops silently or exits with no evidence. Always check `ls .agent_prompt_*.txt` after `worktree add` and delete if it's not for the current branch. (Pattern verified across sessions `20260722_155550_273bd3e1`, `20260730_*`).\n\n---\n\n## The audit produces 4 deliverables\n\n1. **Verdict table** — model × capability matrix with `pass`\u002F`fail`\u002F`partial` per row\n2. **Direct-API evidence** — raw JSON responses saved to `evidence\u002F\u003Cmodel>_\u003Ccapability>.json`\n3. **Latency benchmark** — `evidence\u002Flatency_\u003Cworkload>.json` with per-turn + aggregate stats\n4. **Promotion recommendation** — explicit `safe-to-ship: yes\u002Fno\u002Fwith-fixes` + the specific fixes if `no`\n\nIf any deliverable is missing, the audit is not complete. Don't ship the model change.\n\n---\n\n## Reference\n\n- `references\u002Fworldarchitect-gemini-flash-lite-2026-07-30.md` — verified reproduction of the Lite\u002F3.6 Flash infinite-loop + the PR #8571 → #8590 history + the latency\u002Fcost table.\n","# LLM model capability audit\n\n> **What this skill is:** the pre-promotion gate that catches \"model claims X in marketing docs but actually returns `TOO_MANY_TOOL_CALLS` \u002F empty content \u002F 30k reasoning tokens on a trivial prompt\" regressions BEFORE they ship as a default. Distinct from `llm-model-upgrade-survey` (which is for surveying new options); this is for verifying that a model actually does what the catalog says before flipping the default.\n\n**Read first:** canonical examples in `references\u002Fworldarchitect-gemini-flash-lite-2026-07-30.md` — the live evidence behind this skill.\n\n---\n\nFor new Gemini selections or promotions, only consider Gemini 3.8 Flash or later; never promote Gemini Pro or an earlier Gemini model. Older model IDs below describe historical findings.\n\n## When to use this skill\n\nUse it whenever ANY of the following lands in your session:\n\n- \"make X the new default model\"\n- \"add Y to the model dropdown \u002F settings \u002F ALLOWED_MODELS\"\n- \"why is the dice tool slow \u002F using the wrong path?\"\n- \"we just shipped a regression — users picked model X and got different behavior\"\n- A PR title like \"feat: switch default to X\" \u002F \"feat: add Y to allowlist\"\n- A code change touching `MODELS_WITH_CODE_EXECUTION`, `ALLOWED_MODELS`, `DEFAULT_*_MODEL`, or equivalent per-provider allowlists\n- A user report of \"the LLM is taking forever\" \u002F \"model returned empty\" \u002F \"cost jumped 5× after we changed defaults\"\n\n**Do NOT use it** for: pure creative prose evaluation, narrative-quality review, brand-tone critique — those are not capability audits.\n\n---\n\n## The audit workflow\n\n### Phase 1 — Catalog snapshot (no LLM calls, 5 minutes)\n\nFor every model the change touches, capture:\n- Catalog entries (`ALLOWED_MODELS`, `MODELS_WITH_CODE_EXECUTION`, model mapping tables, frontend dropdown options)\n- The **claimed** capability comment for each (e.g. \"inherits 3.x code execution + JSON single-pass\")\n- Whether the model is in any allowlist that gates critical features (dice, structured output, code execution)\n\nCross-reference: is every model in the dropdown also in the relevant allowlists? If not, **that's the latent bug** — users picking it silently degrade.\n\nOutput: a verdict table with 4 columns (model, in dropdown, in allowlist, capability claim).\n\n### Phase 2 — Direct API probe (real LLM calls, 10 minutes per model)\n\nFor each model the change touches, hit the **real provider API** (not just docs) with the EXACT request shape the app sends:\n\n- Same tools array (e.g. `{\"code_execution\": {}}` for Gemini)\n- Same response format\n- Same temperature + system prompt structure\n- Same approximate prompt length the app sends (often much larger than test prompts)\n\nCapture: `finish_reason`, `code_parts` \u002F tool-call count, `total_tokens`, `usage.completion_tokens_details.reasoning_tokens`, and the actual text output. **Critical: reasoning_tokens count on trivial prompts**. A model that emits 200 reasoning tokens on a one-line JSON request is 5-10× more expensive than its base rate suggests.\n\n**Auth-source priority** (in order, per `dual-probe-secrets` SOUL.md rule):\n1. `os.environ.get(\"PROVIDER_API_KEY\")` from bashrc-sourced shell\n2. macOS Keychain via `security find-generic-password -s \u003Cservice> -w`\n3. GCP Secret Manager via `gcloud secrets versions access latest --secret=\u003Cname> --project=\u003Cproject>`\n4. Local secret file (`~\u002F.gemini_api_key_secret`, etc.)\n\nNever trust a single source — dual-probe both local file AND GCP secret-manager when both exist.\n\n### Phase 3 — Latency benchmark (12 turns minimum, 10 minutes)\n\nFor latency-sensitive decisions (default model selection, dice provider choice), run 12 turns per model with a representative prompt:\n- Capture mean, p50, p95 latency\n- Capture cost per 12 turns\n- Capture total tokens + reasoning tokens\n- Capture finish_reason distribution (any non-`stop` is a smell)\n\n**Tail-latency cliff check:** if p95 > 3× mean, the model has unstable latency under load. Worth flagging.\n\nFor structured-output workloads (dice, JSON schema validation), also run a `response_format={\"type\": \"json_object\"}` test — some models accept the parameter but emit markdown-wrapped JSON instead.\n\n### Phase 4 — Promotion gate (3 questions)\n\nBefore approving any \"make X the default\" PR, answer:\n1. **Is X in every allowlist that gates critical features?** (dice, code execution, structured JSON)\n2. **Does the direct-API probe show clean finish + correct output + reasonable token footprint?**\n3. **Is the p95 latency within 2× the existing default?**\n\nIf any answer is \"no\" — the change is NOT safe to ship. Either patch the allowlist first, fix the model choice, or document the limitation in the catalog comment.\n\n---\n\n## Common pitfalls (verified in production)\n\n### Pitfall 1 — Allowlist drift after model add\n\nWhen a PR adds a new model to `ALLOWED_MODELS` but forgets to add it to `MODELS_WITH_CODE_EXECUTION`, users picking that model silently get a slower fallback path (e.g. native two-phase dice instead of code-execution dice). Symptom is silent: the LLM still returns valid JSON, just via a slower + more expensive code path.\n\n**Defense:** Phase 1 catalog snapshot must cross-check every allowlist the model affects.\n\n### Pitfall 2 — Comment claims capabilities that don't exist\n\nA code comment like \"inherits 3.x code execution + JSON single-pass\" is documentation debt. If the model doesn't actually inherit the capability (e.g. `gemini-3.5-flash-lite` loops on code-execution calls), the comment becomes a lie that future agents trust. Always verify the comment against the direct-API probe.\n\n**Verified example:** PR #8512 (your-project.com) added `gemini-3.5-flash-lite` and `gemini-3.6-flash` with the claim that both inherit code execution. Direct API test: both return `TOO_MANY_TOOL_CALLS` with 48 code_parts and 20-27k tokens of empty output.\n\n### Pitfall 3 — Reasoning-token footgun\n\n\"Fast\" tier models often emit reasoning tokens on trivial prompts. `gemini-3.5-flash` emits 220 reasoning tokens on a 1-line JSON request. `gemini-3.6-flash` emits 237. That makes them 20-30× more expensive per turn than their headline price suggests. Always capture `usage.completion_tokens_details.reasoning_tokens` separately.\n\n### Pitfall 4 — Default flip without allowlist patch\n\nPR #8571 (your-project.com) made `gemini-3.5-flash-lite` the default. The model wasn't in `MODELS_WITH_CODE_EXECUTION`, so every user silently got native two-phase dice. The PR was reverted 8 hours later by PR #8590, which only fixed the **default**, not the allowlist. The latent bug persists for users who explicitly pick Lite or 3.6.\n\n**Defense:** Phase 4 promotion gate must explicitly ask \"does this model work for every allowlist that the current default works in?\"\n\n### Pitfall 5 — Tail-latency cliff\n\n`gemini-3.6-flash` has mean=3.20s, p50=2.06s, p95=**11.84s** on a 12-turn dice benchmark. The model is \"fine\" on average and terrible on the 5th percentile. Always run p95, not just mean.\n\n### Pitfall 6 — Bashrc-vs-Keychain vs GCP secret-manager divergence\n\nThe \"where's the key\" question has 4 valid answers and they often disagree:\n- bashrc export (often unset or commented out for security)\n- macOS Keychain via `security find-generic-password`\n- GCP Secret Manager via `gcloud secrets versions access`\n- Local secret file (`~\u002F.gemini_api_key_secret` style)\n\n`dual-probe-secrets` SOUL.md rule: probe at least 2 of these before declaring \"key is broken\". The local secret file in this repo was stale (`AIzaSyDQ6a...` returns `API_KEY_INVALID`) but the GCP secret-manager version (`AIzaSyAvyb...`) was live.\n\n### Pitfall 7 — Codex is a UI, not an API\n\nGPT-6 Luna is only exposed in the Codex picker — but the underlying API is OpenRouter. `POST https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1\u002Fchat\u002Fcompletions` with `model=\"openai\u002Fgpt-6-luna\"` works directly. Same for any \"Codex-only\" model: it's almost always OpenRouter under the hood.\n\n### Pitfall 8 — `claudem` worker with stale `.agent_prompt_*.txt`\n\nWhen dispatching a `claudem -p` worker via `git worktree add`, the worktree may inherit a stale `.agent_prompt_\u003Cprior-branch>.txt` from a prior session. The worker reads the stale prompt and either loops silently or exits with no evidence. Always check `ls .agent_prompt_*.txt` after `worktree add` and delete if it's not for the current branch. (Pattern verified across sessions `20260722_155550_273bd3e1`, `20260730_*`).\n\n---\n\n## The audit produces 4 deliverables\n\n1. **Verdict table** — model × capability matrix with `pass`\u002F`fail`\u002F`partial` per row\n2. **Direct-API evidence** — raw JSON responses saved to `evidence\u002F\u003Cmodel>_\u003Ccapability>.json`\n3. **Latency benchmark** — `evidence\u002Flatency_\u003Cworkload>.json` with per-turn + aggregate stats\n4. **Promotion recommendation** — explicit `safe-to-ship: yes\u002Fno\u002Fwith-fixes` + the specific fixes if `no`\n\nIf any deliverable is missing, the audit is not complete. Don't ship the model change.\n\n---\n\n## Reference\n\n- `references\u002Fworldarchitect-gemini-flash-lite-2026-07-30.md` — verified reproduction of the Lite\u002F3.6 Flash infinite-loop + the PR #8571 → #8590 history + the latency\u002Fcost table.\n",{"name":5,"description":7},"main","42ccd02cbf5e8e47839dc3eab6fc754dc6b339cf","hermes\u002Fskills\u002Fdevops\u002Fllm-model-capability-audit",0,"https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills","https:\u002F\u002Fgithub.com\u002Fjleechanorg","2026-11-14 14:55:35",{"source":8,"skillName":5,"commands":63,"prompt":73,"zipAvailable":74},[64,67,70],{"tool":65,"command":66},"npx","npx skills add jleechanorg\u002Fjleechan-skills --skill llm-model-capability-audit",{"tool":68,"command":69},"bunx","bunx skills add jleechanorg\u002Fjleechan-skills --skill llm-model-capability-audit",{"tool":71,"command":72},"pnpm","pnpm dlx skills add jleechanorg\u002Fjleechan-skills --skill llm-model-capability-audit","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002Fhermes\u002Fskills\u002Fdevops\u002Fllm-model-capability-audit\u002FSKILL.md",true,[],[77,95,117,131,145,160],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":50},"01M716Y5MTJ6GYJZ2ATXCJK8G4","instinct","Use when the user invokes \u002Finstinct or \u002Fin, or asks to \"message Instinct\", \"ask Instinct\", \"tell Instinct\", \"check Instinct\", or read replies from the configured peer assistant. iMessage first, WhatsApp Web as fallback.","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Finstinct\u002FSKILL.md",[],[79,84,85,86,87,88,89,90,91,92,93,94],"user","when","replies","check","use","identity","whatsapp","gmail","read","message","ask",{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":106,"publishedAt":50},"01M716Y4HF7FVVMJSA2CCVN5VF","hr","Compatibility alias for history-resume; produce the same fresh-state, provenance-aware resume briefing.","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fhr\u002FSKILL.md",[101],{"categoryNo":102,"categoryName":103,"taxonomyType":104,"taxonomyTypeDesc":105},"01JBSEEDCAT000000000000038","人力资源","PROFESSION","适用职业",[97,107,108,109,110,111,112,113,114,115,116],"resume","compatibility","provenance","for","history","produce","alias","briefing","state","fresh",{"skillNo":118,"name":119,"nameZh":6,"description":120,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":121,"categories":122,"keywords":124,"publishedAt":50},"01M716Y1FFNFKFJYYKF8HQA6T8","history-resume","Resume briefing for the current repository branch or worktree using fresh state and admitted history sources.","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fhistory-resume\u002FSKILL.md",[123],{"categoryNo":102,"categoryName":103,"taxonomyType":104,"taxonomyTypeDesc":105},[111,107,125,126,84,127,114,128,110,129,130,115,116],"history resume","branch","analysis","repository","worktree","sources",{"skillNo":132,"name":133,"nameZh":6,"description":134,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":135,"categories":136,"keywords":137,"publishedAt":50},"01M716XWXJ8EPHK4RH9T14TR85","grok","Alias for the canonical grokbot browser conversation workflow; \u002Fgrok talks to the Grok bot the same way \u002Fdot talks to Dot.","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fgrok\u002FSKILL.md",[],[133,138,139,140,110,141,113,142,143,144],"dot","bot","workflow","canonical","conversation","browser","grokbot",{"skillNo":146,"name":147,"nameZh":6,"description":148,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":149,"categories":150,"keywords":151,"publishedAt":50},"01M70N9Q07DTE0EW5KFXMMMSJ0","codexa-bench","Compare `codexa` and normal `codex` on a fair split of comparable work by default; use `--strict` for duplicate runs of the exact same task.","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fcodexa-bench\u002FSKILL.md",[],[152,153,154,88,35,110,155,156,157,158,159],"codexa bench","work","split","duplicate","task","compare","normal","codex",{"skillNo":161,"name":162,"nameZh":6,"description":163,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":164,"licenseFileUrl":165,"sourceHtmlUrl":166,"categories":167,"keywords":177,"publishedAt":50},"01M6C24S2X60SP8BBFGAEKPEGW","frontend-design","Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, artifacts, posters, or applications (examples include websites, landing pages, dashboards, React components, HTML\u002FCSS layouts, or when styling\u002Fbeautifying any web UI). Generates creative, polished code and UI design that avoids generic AI aesthetics.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffrontend-design\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fjleechanorg\u002Fjleechan-skills\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffrontend-design\u002FSKILL.md",[168,171,174],{"categoryNo":169,"categoryName":170,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000009","前端开发",{"categoryNo":172,"categoryName":173,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000015","视觉设计",{"categoryNo":175,"categoryName":176,"taxonomyType":104,"taxonomyTypeDesc":105},"01JBSEEDCAT000000000000032","设计师",[178,179,180,88,181,182,183,84,184,185,49,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200,85,201],"frontend","design","frontend design","generic","ui design","react","html","components","artifacts","ui","aesthetics","Production","web components","landing","applications","quality","web","css","create","examples","with","react components","web ui","pages"]