/goal in Claude Code: set the stop condition and walk away

TL;DR /goal <condition> leaves a persistent note: "don't return control to me until this holds". After every turn, Haiku reads the transcript and decides yes/no against your condition. No → Claude starts another turn automatically. Yes → control returns to you. Use it for sweeping refactors, migrations, or draining an issue queue — anything with a verifiable "done".

There are tasks you know will finish, but they force you to type "keep going" every two turns. Migrating an import across the repo. Getting a legacy test suite to pass. Draining a labeled issue queue. After each turn Claude stops and waits for your nod. /goal breaks that cycle: you state the stop condition and stop being a human cron.

How it works

/goal wraps a session-scoped prompt-based Stop hook. Every time Claude finishes a turn:

1. The condition and the conversation so far are sent to the small fast model (Haiku by default).
2. The evaluator returns yes/no plus a short reason.
3. "No" → Claude continues, with the reason injected as guidance for the next turn.
4. "Yes" → the goal clears itself and control returns to you.

The evaluator does NOT call tools — it only reads what Claude has already surfaced in the conversation. Conditions must be verifiable from the transcript.

Result preview

> /goal all tests in test/auth pass and the lint step is clean

◎ /goal active · 0m

# Claude runs npm test → 3 failures
# Evaluator: condition not met — 3 tests still failing in test/auth/*
# Claude fixes the first test → next turn
# Claude fixes the other two → next turn
# npm test green, npm run lint clean
# Evaluator: condition met

◎ /goal cleared — achieved in 4m 12s, 6 turns

How to use it

1. Set the condition

/goal all tests in test/auth pass and the lint step is clean

Setting the goal starts a turn immediately, with the condition as the directive — no extra prompt needed. Up to 4,000 characters per condition.

2. Write conditions verifiable from the transcript

The evaluator only sees what Claude has surfaced in the conversation:

• "npm test exits 0" → Claude runs it, the output lands in the transcript
• "every getUser call site now reads fetchUser" → Claude greps, the result is visible
• "the app is production-ready" → not verifiable, don't use it

Add constraints ("don't touch migrations") and an upper bound ("or stop after 20 turns") when there's a real risk of looping.

3. Status and cancel

/goal         # status: condition, turns, time, tokens, last evaluator reason
/goal clear   # cancel (aliases: stop, off, reset, none, cancel)

Running /clear to start a new conversation also drops the active goal.

4. Resume across sessions

If the session ended with an active goal, claude --resume or --continue restores it. The condition persists; the turn counter and timer reset.

5. Headless (CI / cron / scripts)

claude -p "/goal CHANGELOG.md has an entry for every PR merged this week"

Single invocation; runs until the condition holds or you hit Ctrl+C.

/goal vs /loop vs your own Stop hook

/goal is the closest thing to "Claude, don't come back to me until this is done".

Reference

Official docs: Keep Claude working toward a goal