AgentSkillsCN

docs-seeker

利用 llms.txt 标准、通过 Repomix 搜索 GitHub 仓库,并结合并行探索,在互联网上查找技术文档。当用户有如下需求时可使用此技能:(1) 获取最新库/框架文档;(2) 查找 llms.txt 格式的文档;(3) 分析 GitHub 仓库;(4) 查找未直接支持 llms.txt 格式的文档;(5) 同时从多个文档来源中进行并行检索。

SKILL.md
--- frontmatter
name: docs-seeker
description: "Searching internet for technical documentation using llms.txt standard, GitHub repositories via Repomix, and parallel exploration. Use when user needs: (1) Latest documentation for libraries/frameworks, (2) Documentation in llms.txt format, (3) GitHub repository analysis, (4) Documentation without direct llms.txt support, (5) Multiple documentation sources in parallel"
version: 2.0.0

Documentation Discovery

Find technical documentation through: llms.txtRepomixResearch

Decision Flow

code
1. Try context7.com llms.txt
   ├─ Found → Process URLs with Explorer agents
   └─ Not found ↓
2. Find GitHub repository
   ├─ Found → Use Repomix
   └─ Not found ↓
3. Deploy Researcher agents

Method 1: llms.txt (Priority)

context7.com Patterns

SourcePattern
GitHub repohttps://context7.com/{org}/{repo}/llms.txt
Websitehttps://context7.com/websites/{normalized-path}/llms.txt
Topic search...llms.txt?topic={query}

Examples:

code
github.com/vercel/next.js     → context7.com/vercel/next.js/llms.txt
docs.imgix.com                → context7.com/websites/imgix/llms.txt
shadcn/ui + date picker       → context7.com/shadcn-ui/ui/llms.txt?topic=date

Fallback: WebSearch

code
"[library] llms.txt site:[docs domain]"

Processing URLs

URL CountStrategy
1-3Single Explorer agent
4-103-5 Explorer agents
11+5-7 agents in phases

Launch all agents in single message (parallel, not sequential).


Method 2: Repomix

When llms.txt unavailable, analyze GitHub repository directly.

Quick Commands

bash
# Remote (recommended - no clone needed)
npx repomix --remote https://github.com/owner/repo

# Docs only
npx repomix --remote [url] --include "docs/**,*.md"

# Source only
npx repomix --remote [url] --include "src/**" --exclude "**/*.test.*"

# Specific version
npx repomix --remote [url] --remote-branch v2.0.0

# With security check
npx repomix --remote [url] --security-check

Key Options

OptionPurpose
--remote [url]Analyze without cloning
--remote-branch [ref]Specific branch/tag/commit
--include "pattern"Only include matching files
--exclude "pattern"Skip matching files
--style markdownHuman-readable output
--security-checkScan for secrets
--copyCopy to clipboard

Language-Specific Patterns

bash
# JavaScript/TypeScript
--include "src/**,lib/**,*.md" --exclude "node_modules/**,dist/**,*.test.*"

# Python
--include "src/**,*.py,*.md" --exclude "__pycache__/**,venv/**"

# Go
--include "**/*.go,go.mod,*.md" --exclude "vendor/**,*_test.go"

# PHP/Laravel
--include "app/**,config/**,routes/**,*.md" --exclude "vendor/**,storage/**"

Error Handling

ProblemSolution
Repo too large--include "docs/**,*.md"
TimeoutClone with --depth 1, then local repomix
Private repoClone with token, then local repomix
Not installedUse npx repomix

Method 3: Research (Fallback)

When no structured docs exist, deploy 3-4 Researcher agents:

  1. Official sources: Package registry, website, releases
  2. Tutorials: Blog posts, guides
  3. Community: Stack Overflow, GitHub discussions
  4. API/Reference: Auto-generated docs, code examples

Output Format

markdown
# [Library] Documentation

## Source
- Method: llms.txt / Repomix / Research
- URL: [source]
- Version: [version]
- Date: [date]

## Content
[Organized by topic, not by agent]

## Limitations
[Missing sections, version caveats]

Best Practices

  1. context7.com first - fastest path to documentation
  2. Use ?topic= when applicable - reduces noise
  3. Parallel agents - never sequential for 3+ URLs
  4. Verify official sources - check domain, version, date
  5. Aggregate by topic - synthesize, don't concatenate
  6. Report methodology - tell user how you found info
  7. Handle versions explicitly - don't assume latest
  8. Fail fast - 30s timeout, then try next method