Spawn Investigator

Launch an isolated investigation subagent to answer a codebase question without polluting the main conversation context.

Why This Exists

Reading many files to answer a question fills up the main context window. By spawning an Explore-type subagent, the investigation happens in a separate context. Only the concise answer comes back, keeping the main context clean for actual work.

Usage

• /spawn-investigator <question> — User-invoked investigation
• Claude can also auto-invoke this when it recognizes a question requiring broad codebase exploration

Procedure

Step 1: Formulate the Question

Parse $ARGUMENTS into a clear, specific investigation question. If the question is vague, refine it:

• BAD: "How does auth work?"
• GOOD: "What authentication middleware is used, where is it configured, and what are the auth routes?"

Step 2: Spawn the Subagent

Use the Task tool with these parameters:

subagent_type: Explore
description: "Investigate: [short summary]"
prompt: |
  Investigate this question about the codebase:

  **Question:** [the question]

  Instructions:
  - Search broadly first (Glob for file patterns, Grep for keywords)
  - Read the most relevant files
  - Trace the code path if needed (follow imports, function calls)
  - Return a CONCISE answer (max 500 words) with:
    1. Direct answer to the question
    2. Key file paths discovered
    3. Any important caveats or gotchas
  - Do NOT include full file contents in your response

Step 3: Report Back

When the subagent returns, present the findings to the user:

## Investigation Results: [question summary]

[Subagent's concise answer]

### Key Files
- `path/to/file1` — [what it does]
- `path/to/file2` — [what it does]

If the answer is insufficient, offer to spawn a follow-up investigator with a more specific question.

When to Auto-Invoke

Claude should consider auto-invoking this skill when:

• The user asks a broad "how does X work?" question
• Answering requires reading 5+ files
• The main context is already large (approaching compaction)
• The question is about understanding, not modifying, code