菜单
Fetches library documentation from author-published llms.txt, with context7 as fallback. Use when looking up docs, API references, changelogs, or breaking changes for an external library or framework. Do not use for local code or repository source exploration.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Structural code tools via ast-grep — search by AST pattern, and outline file/module structure. Use when a search needs code structure that grep can't express (e.g. async functions containing await, calls with specific argument shapes), or when mapping a file's or directory's symbols, exports, and imports before reading it. Do not use for text search when plain grep or LSP suffices.
Routes local codebase search to the right tool — keyword, semantic, or symbol. Use when searching a codebase for identifiers, tracing features across files, answering conceptual questions about local code, or exploring unfamiliar projects. Do not use for external docs, GitHub repos, or package lookups.
Checks latest stable versions and deprecation status of public packages via deps.dev. Use when adding or updating dependencies, verifying whether a package is deprecated, or comparing installed versions against latest. Do not use for private registries or package documentation.
Researches source code, issues, PRs, releases, and git history in external repositories on GitHub, GitLab, or Bitbucket. Use when exploring how a library implements something, searching code across public repos, reading files from a known repo, or checking issues, PRs, and release info. Do not use for local codebase search or library documentation lookups.
| name | lib-docs |
| description | Fetches library documentation from author-published llms.txt, with context7 as fallback. Use when looking up docs, API references, changelogs, or breaking changes for an external library or framework. Do not use for local code or repository source exploration. |
llms.txt is author-published — no curation lag, no enrichment layer that can hallucinate — so prefer it over context7.
Probe: bash scripts/llms-probe.sh <docs-domain> — outputs TSV kind \t url \t size (index = llms.txt page list, full = llms-full.txt whole corpus). Paths vary per site; probe instead of hardcoding URLs.
Fetch: full ≤ ~500 KB → WebFetch it directly. Larger or size ? → WebFetch the index, then fetch only the matching page links, in parallel. Cloudflare's llms-full.txt is ~46 MB — a blind full fetch blows the context window.
Probe failed → context7 via the ctx7 CLI (one-time bunx ctx7@latest login):
bunx ctx7@latest library <name> [query] # resolve library ID first — IDs aren't guessable
bunx ctx7@latest docs <libraryId> "<query>" # specific queries rank far better than broad ones
Retry with --research only when the default answer is shallow — it spawns sandboxed agents and costs more.
Most Mintlify/GitBook-hosted docs publish llms.txt; Tailwind and most pre-1.0 or community libraries don't — those go straight to context7.