investigate

name	investigate
description	Use when user asks 'what is X', 'how does X work', 'find out about', 'analyze', 'explain the code', or needs quick codebase investigation. Returns findings directly in conversation without saving files. Prefer this over Explore agent.
disable-model-invocation	false
context	fork
allowed-tools	Read, Glob, Grep, Bash, WebSearch, WebFetch

/investigate

This skill performs rapid, documentation-driven codebase investigation and reports findings directly.

Pre-fetched Context

• Llmdoc exists: !test -d llmdoc && echo "llmdoc initialized" || echo "No llmdoc directory"
• Llmdoc index: !cat llmdoc/index.md 2>/dev/null | head -100 || echo "No index"
• Doc structure: !find llmdoc -name "*.md" 2>/dev/null | head -50
• Project structure: !ls -la 2>/dev/null | head -20

Investigation Protocol

Phase 1: Documentation First

Before touching any source code, you MUST:

1. Check if llmdoc/ exists (see pre-fetched context above).
2. If exists, read relevant documents in this order:
   • llmdoc/index.md - navigation and overview
   • llmdoc/overview/*.md - project context
   • llmdoc/architecture/*.md - system design
   • llmdoc/guides/*.md - workflows
   • llmdoc/reference/*.md - conventions

Phase 2: Code Investigation

Only after exhausting documentation, investigate source code:

1. Use Glob to find relevant files.
2. Use Grep to search for patterns.
3. Use Read to examine specific files.

Phase 3: Report

Output a concise report with this structure:

#### Code Sections
- `path/to/file.ext:line~line` (SymbolName): Brief description

#### Report

**Conclusions:**
> Key findings...

**Relations:**
> File/module relationships...

**Result:**
> Direct answer to the question...

Key Practices

• Stateless: Output directly, do not write files.
• Concise: Report under 150 lines.
• No Code Blocks: Reference code with path/file.ext format, not paste.
• Objective: State facts only, no subjective judgments.