Keep-or-revert eval loop
Single-metric experiments where only measured gains survive. Works for prompts, pages, build times and more.
How each pass goes
- Pick one metric and how to measure it.
- Try one change.
- Keep it only if the metric improves.
Stops when
The target is reached or the experiment budget is spent.
Starter prompt
Our metric is: [metric and the command that measures it]. Record the baseline. Then loop: try one change, measure, keep it only if it beats the best result, otherwise revert. Log each attempt and its number. Stop after [N] attempts or when we reach [target].