Set /goal with a testable condition and let Claude decide when it is done

The problem

Long tasks stall because the agent has no standard for done. It finishes a turn, declares success, and waits — so you spend the session nudging it forward one "keep going" at a time.

The hack

Run `/goal <condition>` with a condition that can actually fail. Claude keeps working across turns until it is met. Pair it with a test suite so the condition is machine-checkable rather than a judgement call.

Why it works

A goal turns your session into a loop with a real exit check. When the main agent decides a task is done, a smaller model verifies the work against the requirements and reprompts it to finish whatever it missed, so "done" is not self-assessed.

Setup

1. Write the failing tests for the feature before asking for any code. 2. Run `/goal every test in tests/checkout.spec.ts passes`. 3. Let it work — it will write code, run tests and iterate without you prompting each round. 4. Run `/goal` with no argument to see the current or most recently achieved goal. 5. Clear it early with `/goal clear` (or `stop`, `off`, `cancel`). 6. Add a line to `CLAUDE.md` telling Claude to commit each working version so you can roll back to the last one that passed.

Prompt

/goal npm test passes with zero failures and npx tsc --noEmit reports no errors

Additional details

Source

Related

• → Build a verification-loop skill that must run tests before done
• → Prompt with task, guardrails and exit criteria instead of numbered steps
• → Run recurring checks with /loop and /schedule