Loop/Prompt Library / Running AI agents well
🔁 Champion vs challenger
Keeps the current best prompt as champion, and a new version only replaces it if it wins on fresh test cases it never saw.
Use it when: You are tuning a prompt or setting and want the final call made on fresh examples.
🛑 Stops when: The challenger beats the champion on unseen cases, or the budget runs out.
Copy this loop
Improve [prompt or setting]. Save the current version as the champion with its score, a practice set, a separate untouched test set, and must-pass checks. Each round, change one thing based on a real failure. Promote the challenger only if it beats the champion on the untouched set by [margin] and passes every must-pass check. Otherwise keep the champion. Run at most [10] rounds. Return the winner, scores, and a log of every attempt.
Review every change the agent makes before you keep it. Set a budget or round limit — the prompt includes one.
Idea from Forward Future's Loop Library · by Jose C. Munoz. Prompt rewritten in plain English by FixKit AI.
More loops like this
Agent workspace health check
Reads a project's notes and state, without changing anything, and grades whether a new agent could safely pick up the work.
Stop and check before continuing
After an agent finishes a task, it reports what is done and whether the next step is safe, then waits for you.
Clean handoff note
Before a session ends, writes a note so the next person or agent can continue without guessing.
Predict, act, check
Before every meaningful action, the agent writes what it expects to happen, then checks if that is what really happened.
Lessons that stick
Keeps a playbook of lessons from repeated runs, tests one lesson at a time, and only keeps lessons that keep working.
Fair-test experiments
Improves a prompt or piece of code through saved, fairly compared experiments that can pick up again later.