Self Message

Self-Message

You can send yourself a continuation message to start a fresh agentic loop within the same session. This resets your reasoning frame while preserving full conversation context.

This is NOT a subagent. It's a new turn for YOU — same session, same context, new frame.

How It Works

1. Send your self-message using the bundled script
2. Finish your turn normally (stop responding)
3. A Stop hook intercepts your stop, reads the message, deletes it
4. The hook returns {"decision": "block", "reason": "<your message>"}
5. You receive the message as context and start a new agentic loop

bash ${CLAUDE_SKILL_DIR}/scripts/self-message.sh "Reassess: the recursive approach hit a dead end. Try iterative with memoization instead."

Then just stop. That's it. The hook handles session isolation automatically.

When to Use

• Reframing: Your current approach isn't working and you need to pivot
• Phase transitions: You finished research, now need a clean start for synthesis
• Accumulated context: You've gathered scattered information and need to step back and connect the dots
• Course correction: You notice you've drifted from the user's actual request
• Complex multi-step tasks: Breaking work into deliberate phases with a clear handoff message between each

When NOT to Use

• Simple, straightforward tasks — just do them
• When you're already on the right track — keep going
• As a substitute for thinking harder in the current turn
• For delegation — use subagents instead
• For retrying the same thing — that's a Ralph Wiggum loop, not a self-message

Writing Good Self-Messages

Your self-message is the most important part. It's your instruction to your future self. Be specific:

Bad:

Continue working on the task.

Good:

Phase 1 complete: identified 3 API endpoints that need refactoring (users.ts:45, auth.ts:120, billing.ts:89). The common pattern is they all do N+1 queries. Phase 2: implement batch loading for each, starting with users.ts since it's the simplest case.

Good:

Reassess: spent 3 tool calls trying to fix the type error by widening the type. Wrong direction — the real issue is the caller in main.ts:230 passing the wrong shape. Fix the caller, not the type.

Include:

• What you learned / what happened
• Why you're self-messaging (what needs to change)
• Clear direction for the next phase

Limits

• Max 200 consecutive self-messages by default (configurable via SELF_MSG_MAX_DEPTH env var) — enough for ~12 hours of autonomous work
• Counter resets automatically when you stop normally (no self-message file)
• If you hit the limit, you stop normally — write a good summary for the user

Important

• The self-message file is deleted after being read. Write a new one each time.
• You do NOT need user permission to self-message. It's your tool.
• The user sees your work across all loops — nothing is hidden.
• This is a metacognitive tool. Use it deliberately, not reflexively.