Improve

Set a finish-line condition for any output, then run an autonomous improve → evaluate loop until the condition holds. Modeled on Claude Code's /goal — the condition itself is the directive, and a fresh self-evaluation after each iteration decides whether to keep going.

How It Works (TL;DR)

1. Finish line — Establish ONE measurable completion condition with an explicit check and a bound
2. Improve — Make one focused change targeting the condition
3. Evaluate — Self-check against the condition: met / not met + one-sentence reason
4. Loop or stop — If met → deliver. If not → the reason becomes the directive for the next iteration.
5. Bound lives in the condition — e.g., "…or stop after 5 iterations"

This skill stays autonomous between iterations: no user checkpoint mid-loop. The condition is the contract. The user can interrupt at any time with "stop" / "good enough".

Tools