Documentation Discovery & Analysis

Overview

1. llms.txt-first: Search for standardized AI-friendly documentation
2. Repository analysis: Use Repomix to analyze GitHub repositories
3. Parallel exploration: Deploy multiple Explorer agents for comprehensive coverage
4. Fallback research: Use Researcher agents when other methods unavailable

Core Workflow

Phase 1: Initial Discovery

1. Identify target
   • Extract library/framework name from user request
   • Note version requirements (default: latest)
   • Clarify scope if ambiguous
   • Identify if target is GitHub repository or website
2. Search for llms.txt (PRIORITIZE context7.com)
   First: Try context7.com patterns
   For GitHub repositories:

   Pattern: https://context7.com/{org}/{repo}/llms.txt
   Examples:
   - https://github.com/imagick/imagick → https://context7.com/imagick/imagick/llms.txt
   - https://github.com/vercel/next.js → https://context7.com/vercel/next.js/llms.txt
   - https://github.com/better-auth/better-auth → https://context7.com/better-auth/better-auth/llms.txt

   For websites:

   Pattern: https://context7.com/websites/{normalized-domain-path}/llms.txt
   Examples:
   - https://docs.imgix.com/ → https://context7.com/websites/imgix/llms.txt
   - https://docs.byteplus.com/en/docs/ModelArk/ → https://context7.com/websites/byteplus_en_modelark/llms.txt
   - https://docs.haystack.deepset.ai/docs → https://context7.com/websites/haystack_deepset_ai/llms.txt
   - https://ffmpeg.org/doxygen/8.0/ → https://context7.com/websites/ffmpeg_doxygen_8_0/llms.txt

   Topic-specific searches (when user asks about specific feature):

   Pattern: https://context7.com/{path}/llms.txt?topic={query}
   Examples:
   - https://context7.com/shadcn-ui/ui/llms.txt?topic=date
   - https://context7.com/shadcn-ui/ui/llms.txt?topic=button
   - https://context7.com/vercel/next.js/llms.txt?topic=cache
   - https://context7.com/websites/ffmpeg_doxygen_8_0/llms.txt?topic=compress

   Fallback: Traditional llms.txt search

   WebSearch: "[library name] llms.txt site:[docs domain]"

   Common patterns:

   • https://docs.[library].com/llms.txt

   • https://[library].dev/llms.txt

   • https://[library].io/llms.txt

   → Found? Proceed to Phase 2 → Not found? Proceed to Phase 3

Phase 2: llms.txt Processing

• WebFetch to retrieve content
• Extract and present information

• CRITICAL: Launch multiple Explorer agents in parallel
• One agent per major documentation section (max 5 in first batch)
• Each agent reads assigned URLs
• Aggregate findings into consolidated report

Launch 3 Explorer agents simultaneously:
- Agent 1: getting-started.md, installation.md
- Agent 2: api-reference.md, core-concepts.md
- Agent 3: examples.md, best-practices.md

Phase 3: Repository Analysis

1. Find GitHub repository via WebSearch
2. Use Repomix to pack repository:
   bash

   npm install -g repomix  # if needed
   git clone [repo-url] /tmp/docs-analysis
   cd /tmp/docs-analysis
   repomix --output repomix-output.xml

3. Read repomix-output.xml and extract documentation

• Entire repository in single AI-friendly file
• Preserves directory structure
• Optimized for AI consumption

Phase 4: Fallback Research

• Launch multiple Researcher agents in parallel
• Focus areas: official docs, tutorials, API references, community guides
• Aggregate findings into consolidated report

Agent Distribution Guidelines

• 1-3 URLs: Single Explorer agent
• 4-10 URLs: 3-5 Explorer agents (2-3 URLs each)
• 11+ URLs: 5-7 Explorer agents (prioritize most relevant)

Version Handling

• Search without version specifier
• Use current documentation paths

• Include version in search:

  [library] v[version] llms.txt

• Check versioned paths:

  /v[version]/llms.txt

• For repositories: checkout specific tag/branch

Output Format

# Documentation for [Library] [Version]

## Source
- Method: [llms.txt / Repository / Research]
- URLs: [list of sources]
- Date accessed: [current date]

## Key Information
[Extracted relevant information organized by topic]

## Additional Resources
[Related links, examples, references]

## Notes
[Any limitations, missing information, or caveats]

Quick Reference

• WebSearch → Find llms.txt URLs, GitHub repositories
• WebFetch → Read single documentation pages
• Task (Explore) → Multiple URLs, parallel exploration
• Task (Researcher) → Scattered documentation, diverse sources
• Repomix → Complete codebase analysis

• Astro: https://context7.com/withastro/astro/llms.txt
• Next.js: https://context7.com/vercel/next.js/llms.txt
• Remix: https://context7.com/remix-run/remix/llms.txt
• shadcn/ui: https://context7.com/shadcn-ui/ui/llms.txt
• Better Auth: https://context7.com/better-auth/better-auth/llms.txt

• Astro: https://docs.astro.build/llms.txt
• Next.js: https://nextjs.org/llms.txt
• Remix: https://remix.run/llms.txt
• SvelteKit: https://kit.svelte.dev/llms.txt

Error Handling

• llms.txt not accessible → Try alternative domains → Repository analysis
• Repository not found → Search official website → Use Researcher agents
• Repomix fails → Try /docs directory only → Manual exploration
• Multiple conflicting sources → Prioritize official → Note versions

Key Principles

1. Prioritize context7.com for llms.txt — Most comprehensive and up-to-date aggregator
2. Use topic parameters when applicable — Enables targeted searches with ?topic=...
3. Use parallel agents aggressively — Faster results, better coverage
4. Verify official sources as fallback — Use when context7.com unavailable
5. Report methodology — Tell user which approach was used
6. Handle versions explicitly — Don't assume latest

Detailed Documentation

• WORKFLOWS.md — Detailed workflow examples and strategies

• Tool Selection — Complete guide to choosing and using tools
• Documentation Sources — Common sources and patterns across ecosystems
• Error Handling — Troubleshooting and resolution strategies
• Best Practices — 8 essential principles for effective discovery
• Performance — Optimization techniques and benchmarks
• Limitations — Boundaries and success criteria