The Explore Subagent — Context-Efficient Discovery

Core 7 min +30 XP

💡

THE ANALOGY

Sending a scout ahead before moving the army. The scout maps terrain, identifies obstacles, returns with a briefing. The army acts on the briefing — not the raw scouting experience. The Explore subagent is that scout.

⚠️ EXAM TRAP — The Wrong Answer People Choose

Running extensive codebase exploration directly in the main session, which fills context with raw file contents and intermediate output, leaving no room for implementation work.

KEY POINTS

1 Explore subagent: isolated sub-session for verbose discovery — returns only a structured summary.

2 context: fork creates isolation — the Explore subagent's file reading doesn't accumulate in main context.

3 Explore output is a structured summary (JSON or organized markdown) — NOT raw file contents.

4 Use /compact to compress main session context mid-task — preserves key decisions, removes verbose history.

5 After Explore returns, main session has intelligence without the noise — full context window for implementation.

Writing an Effective Explore Prompt

Explore the authentication system in this codebase.

## What to Find
1. All auth entry points (middleware, decorators, route guards)
2. Token creation and validation locations
3. Session management approach
4. Security concerns (TODOs, known issues, incomplete implementations)

## Output Format (JSON — not raw file contents)
{
  "auth_entry_points": [
    {"file": "...", "type": "middleware|decorator", "mechanism": "...", "applies_to": "..."}
  ],
  "token_handling": {
    "creation": "file:function()",
    "validation": "file:function()",
    "expiry": "description of how handled"
  },
  "session_approach": "stateless JWT | server-side sessions | ...",
  "files_examined": 0,
  "security_concerns": ["...", "..."]
}

Return ONLY this JSON. Do not include file contents or grep output.

Context Management with /compact

When the main session grows too large:

/compact

Claude Code summarizes the conversation history — preserving:

Key decisions made
Important findings
Current task state
What has and hasn’t been done

Removing:

Verbose intermediate output
Raw tool results already processed
Repetitive context

Use /compact when context is above 50% to preserve room for implementation.

The Explore → Implement Flow

1. Main session starts: receives task
2. Main session spawns Explore subagent
3. Explore subagent:
   - Reads 40+ files
   - Runs multiple greps
   - Builds comprehensive map
   - Returns: structured 200-line summary
4. Main session receives: 200-line summary (not 10,000 lines of raw files)
5. Main session implements: with full context window available

Key Takeaways

Explore pattern isolates verbose discovery from implementation context
Structured summary output — JSON or organized markdown, not raw content
context: fork creates the isolation — don’t skip it
/compact for compressing a full session mid-task
Explore before large implementations — map first, build second