Docs Seeker

Documentation Discovery & Analysis

Overview

Intelligent discovery and analysis of technical documentation through multiple strategies:

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

Single URL:

• WebFetch to retrieve content
• Extract and present information

Multiple URLs (3+):

• 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

Example:

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

When llms.txt not found:

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

   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

Repomix benefits:

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

Phase 4: Fallback Research

When no GitHub repository exists:

• 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

Latest (default):

• Search without version specifier
• Use current documentation paths

Specific version:

• 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

Tool selection:

• 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

Popular llms.txt locations (try context7.com first):

• 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

Fallback to official sites if context7.com unavailable:

• 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

For comprehensive guides, examples, and best practices:

Workflows:

• WORKFLOWS.md — Detailed workflow examples and strategies

Reference guides:

• 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

Detailed Workflows & Examples

This document provides comprehensive workflow examples for the docs-seeker skill.

Parallel Exploration Strategy

When to Use Multiple Agents

Deploy parallel agents when:

• llms.txt contains more than 3 URLs
• Repository has multiple documentation directories
• Need to check multiple versions
• Comprehensive coverage required

How to Launch Parallel Agents

Use Task tool with Explore subagent:

Example for 5 URLs:
1. Launch all 5 Explore agents in single message
2. Each agent gets specific URLs to read
3. Each agent extracts relevant information
4. Wait for all agents to complete
5. Aggregate results

Agent Distribution Guidelines

Small documentation sets (1-3 URLs):

• Deploy 2 Explore agents to handle all URLs
• Simple, straightforward extraction
• Fastest for small amounts

Medium documentation sets (4-10 URLs):

• Deploy 3-6 Explore agents
• Balance workload evenly
• Group related URLs together

Large documentation sets (11+ URLs):

• Deploy 7-15 Explore agents (max)
• Prioritize most relevant URLs first
• Group related URLs together
• Balance workload evenly
• Avoid over-parallelization

Best Distribution Practices

1. Group related content: Keep related URLs with same agent 2. Balance workload: Distribute URLs evenly by estimated size 3. Prioritize critical docs: Assign core docs first 4. Avoid over-parallelization: Max 7 agents to avoid overwhelming 5. Sequential batches: For 15+ URLs, use two sequential batches

Workflow Examples

Example 1: Library with llms.txt (Simple)

Scenario: User requests documentation for Astro

Step 1: Initial Search (PRIORITIZE context7.com)
→ Try context7.com first: https://context7.com/withastro/astro/llms.txt
→ WebFetch: Read llms.txt content
→ Result: Contains 8+ documentation URLs (success!)

Alternative if context7.com fails:
→ WebSearch: "Astro llms.txt site:docs.astro.build"
→ Result: https://docs.astro.build/llms.txt found

Step 2: Process llms.txt
→ Already fetched in Step 1
→ Result: Contains 8 documentation URLs

Step 3: Parallel Exploration
→ Launch 3 Explorer agents simultaneously:

  Agent 1 (URLs 1-3):
  - https://docs.astro.build/en/getting-started/
  - https://docs.astro.build/en/install/
  - https://docs.astro.build/en/editor-setup/

  Agent 2 (URLs 4-6):
  - https://docs.astro.build/en/core-concepts/project-structure/
  - https://docs.astro.build/en/core-concepts/astro-components/
  - https://docs.astro.build/en/core-concepts/layouts/

  Agent 3 (URLs 7-8):
  - https://docs.astro.build/en/guides/configuring-astro/
  - https://docs.astro.build/en/reference/configuration-reference/

Step 4: Aggregate Findings
→ Collect results from all 3 agents
→ Synthesize into cohesive documentation

Step 5: Present Report
→ Format using standard output structure
→ Include source attribution
→ Note any gaps or limitations

Example 2: Library without llms.txt on context7 (Repository Analysis)

Scenario: User requests documentation for obscure library

Step 1: Try context7.com first
→ Attempt: https://context7.com/org/library-name/llms.txt
→ Result: Not found (404)

Step 2: Find GitHub Repository
→ WebSearch: "[library-name] github repository"
→ Result: https://github.com/org/library-name

Step 2a: Try context7.com with GitHub info
→ Attempt: https://context7.com/org/library-name/llms.txt
→ Result: Still not found

Step 3: Verify Repository
→ Check if it's official/active
→ Note star count, last update, license

Step 4: Check Repomix Installation
→ Bash: which repomix || npm install -g repomix

Step 5: Clone and Process Repository
→ Bash: git clone https://github.com/org/library-name /tmp/docs-analysis
→ Bash: cd /tmp/docs-analysis && repomix --output repomix-output.xml

Step 6: Analyze Repomix Output
→ Read: /tmp/docs-analysis/repomix-output.xml
→ Extract sections: README, docs/, examples/, CONTRIBUTING.md

Step 7: Present Findings
→ Format extracted documentation
→ Highlight key sections: installation, usage, API, examples
→ Note repository health: stars, activity, issues

Example 3: Topic-Specific Search (context7.com feature)

Scenario: User asks "How do I use the date picker in shadcn/ui?"

Step 1: Identify library and topic
→ Library: shadcn/ui
→ Topic: date picker

Step 2: Construct context7.com URL with topic parameter
→ URL: https://context7.com/shadcn-ui/ui/llms.txt?topic=date
→ WebFetch: Read filtered content
→ Result: Returns ONLY date-related documentation (highly targeted!)

Step 3: Present Findings
→ Focused documentation on date picker component
→ Installation instructions
→ Usage examples
→ API reference for date components
→ Much faster than reading entire documentation

Benefits of topic search:
- Reduces context usage (only relevant docs loaded)
- Faster results (no need to filter manually)
- More accurate (context7 filters for you)

Example 4: Multiple Versions Comparison

Scenario: User wants to compare v1 and v2 documentation

Step 1: Identify Version Requirements
→ User needs: v1.x and v2.x comparison
→ Primary focus: migration path and breaking changes

Step 2: Search Both Versions
→ WebSearch: "[library] v1 llms.txt"
→ WebSearch: "[library] v2 llms.txt"

Step 3: Launch Parallel Version Analysis
→ Deploy two sets of Explorer agents:

  Set A - v1 Documentation (3 agents):
  Agent 1: Core concepts v1
  Agent 2: API reference v1
  Agent 3: Examples v1

  Set B - v2 Documentation (3 agents):
  Agent 4: Core concepts v2
  Agent 5: API reference v2
  Agent 6: Examples v2

Step 4: Compare Findings
→ Analyze differences in:
  - Core concepts changes
  - API modifications
  - Breaking changes
  - New features in v2
  - Deprecated features from v1

Step 5: Present Side-by-Side Analysis
→ Migration guide format:
  - What changed
  - What's new
  - What's deprecated
  - Migration steps
  - Code examples (before/after)

Example 4: No Official Documentation (Research Fallback)

Scenario: Library with scattered documentation

Step 1: Exhaust Structured Sources
→ WebSearch: llms.txt (not found)
→ WebSearch: GitHub repo (not found or no docs)
→ WebSearch: Official website (minimal content)

Step 2: Deploy Researcher Agents
→ Launch 4 Researcher agents in parallel:

  Researcher 1: Official sources
  - Package registry page (npm, PyPI, etc.)
  - Official website
  - Release notes

  Researcher 2: Tutorial content
  - Blog posts
  - Getting started guides
  - Video tutorials

  Researcher 3: Community resources
  - Stack Overflow discussions
  - Reddit threads
  - GitHub issues/discussions

  Researcher 4: API & reference
  - Auto-generated docs
  - Code examples in wild
  - Community examples

Step 3: Aggregate Diverse Sources
→ Collect findings from all researchers
→ Cross-reference information
→ Identify consistent patterns
→ Note conflicting information

Step 4: Present Consolidated Report
→ Structure findings:
  - Overview (from multiple sources)
  - Installation (verified approach)
  - Basic usage (community examples)
  - Common patterns (from discussions)
  - Known issues (from GitHub/SO)
  - Caveats about source quality

Example 5: Large Documentation Set (Two-Phase)

Scenario: Framework with 20+ documentation pages

Step 1: Analyze Documentation Structure
→ WebFetch: llms.txt
→ Result: Contains 24 URLs across multiple categories

Step 2: Prioritize URLs
→ Categorize by importance:
  - Critical (8): Getting started, core concepts, API
  - Important (10): Guides, integrations, examples
  - Supplementary (6): Advanced topics, internals

Step 3: Phase 1 - Critical Documentation
→ Launch 5 Explorer agents:
  Agent 1: URLs 1-2 (Getting started)
  Agent 2: URLs 3-4 (Installation & setup)
  Agent 3: URLs 5-6 (Core concepts)
  Agent 4: URLs 7-8 (Basic API)
  Agent 5: URL 9 (Configuration)

→ Wait for completion
→ Quick review of coverage

Step 4: Phase 2 - Important Documentation
→ Launch 5 Explorer agents:
  Agent 6: URLs 10-11 (Routing guide)
  Agent 7: URLs 12-13 (Data fetching)
  Agent 8: URLs 14-15 (Authentication)
  Agent 9: URLs 16-17 (Deployment)
  Agent 10: URLs 18-19 (Integrations)

Step 5: Evaluate Need for Phase 3
→ Assess user needs
→ If supplementary topics required:
  - Launch final batch for advanced topics
→ If basics sufficient:
  - Note additional resources in report

Step 6: Comprehensive Report
→ Synthesize all phases
→ Organize by topic
→ Cross-reference related sections
→ Highlight critical workflows

Performance Optimization Strategies

Minimize Sequential Operations

Bad approach:

1. Read URL 1 with WebFetch
2. Wait for result
3. Read URL 2 with WebFetch
4. Wait for result
5. Read URL 3 with WebFetch
6. Wait for result
Time: 3x single URL fetch time

Good approach:

1. Launch 3 Explorer agents simultaneously
2. Each reads one URL
3. All complete in parallel
4. Aggregate results
Time: ~1x single URL fetch time

Batch Related Operations

Group by topic:

Agent 1: Authentication (login.md, oauth.md, sessions.md)
Agent 2: Database (models.md, queries.md, migrations.md)
Agent 3: API (routes.md, middleware.md, validation.md)

Group by content type:

Agent 1: Tutorials (getting-started.md, quickstart.md)
Agent 2: Reference (api-ref.md, config-ref.md)
Agent 3: Guides (best-practices.md, troubleshooting.md)

Use Caching Effectively

Repository analysis:

1. First request: Clone + Repomix (slow)
2. Save repomix-output.xml
3. Subsequent requests: Reuse saved output (fast)
4. Refresh only if repository updated

llms.txt content:

1. First fetch: WebFetch llms.txt
2. Store URL list in session
3. Reuse for follow-up questions
4. Re-fetch only if user changes version

Fail Fast Strategy

Set timeouts:

1. WebSearch: 30 seconds max
2. WebFetch: 60 seconds max
3. Repository clone: 5 minutes max
4. Repomix processing: 10 minutes max

Quick fallback:

1. Try llms.txt (30 sec timeout)
2. If fails → immediately try repository
3. If fails → immediately launch researchers
4. Don't retry failed methods

Common Pitfalls & Solutions

Pitfall 1: Over-Parallelization

Problem: Launching 15 agents at once Impact: Slow, overwhelming, hard to track Solution: Max 7 agents per batch, use phases for large sets

Pitfall 2: Unbalanced Workload

Problem: Agent 1 gets 1 URL, Agent 2 gets 10 URLs Impact: Agent 1 finishes fast, Agent 2 bottleneck Solution: Distribute evenly or by estimated size

Pitfall 3: Ignoring Errors

Problem: Agent fails, continue without checking Impact: Incomplete documentation, missing sections Solution: Check all agent outputs, retry or note failures

Pitfall 4: Poor Aggregation

Problem: Concatenating agent outputs without synthesis Impact: Redundant, disorganized information Solution: Synthesize findings, organize by topic, deduplicate

Pitfall 5: Not Verifying Sources

Problem: Using first result without verification Impact: Outdated or unofficial documentation Solution: Check official status, version, date

Decision Trees

Choosing Documentation Strategy

Start
  ↓
Does llms.txt exist?
  ↓
YES → How many URLs?
  ↓
  1-3 URLs → Single WebFetch/Explorer
  4+ URLs → Parallel Explorers
  ↓
NO → Is there GitHub repo?
  ↓
  YES → Is Repomix feasible?
    ↓
    YES → Use Repomix
    NO → Manual exploration with Explorers
  ↓
  NO → Deploy Researcher agents

Choosing Agent Count

URL Count < 3
  ↓
Single Explorer
  ↓
URL Count 4-10
  ↓
3-5 Explorers
  ↓
URL Count 11-20
  ↓
5-7 Explorers (or two phases)
  ↓
URL Count > 20
  ↓
Two-phase approach:
  Phase 1: 5 agents (critical)
  Phase 2: 5 agents (important)

Advanced Scenarios

Scenario: Multi-Language Documentation

Challenge: Documentation in multiple languages

Approach: 1. Identify target language from user 2. Search for language-specific llms.txt 3. If not found, search for English version 4. Note language limitations in report 5. Offer to translate key sections if needed

Scenario: Framework with Plugins

Challenge: Core framework + 50 plugin docs

Approach: 1. Focus on core framework first 2. Ask user which plugins they need 3. Launch targeted search for specific plugins 4. Avoid trying to document everything 5. Note available plugins in report

Scenario: Documentation Under Construction

Challenge: New release with incomplete docs

Approach: 1. Note documentation status upfront 2. Combine available docs with repository analysis 3. Check GitHub issues for documentation requests 4. Provide code examples from tests/examples 5. Clearly mark sections as "inferred from code"

Scenario: Conflicting Information

Challenge: Multiple sources with different approaches

Approach: 1. Identify primary official source 2. Note version differences between sources 3. Present both approaches with context 4. Recommend official/latest approach 5. Explain why conflict exists (e.g., version change)