Autoresearch - Turn Prompt Tuning Into Measured Experiments

Most AI skills and prompts work... sometimes. They nail it on one run, drift on the next, and you end up rewriting by gut feel until something sticks. That's not optimization. That's guessing with extra steps.

Autoresearch turns prompt improvement into an actual experiment. It takes any skill or prompt you've built, runs it repeatedly against real test inputs, scores every output with binary pass/fail evals (not vibes), and then mutates one variable at a time - keeping only changes that measurably improve results.

Think of it as A/B testing for your AI workflows, except the AI runs the tests, tracks the scores, and makes the edits for you.

What makes this different:

• Binary evals, not taste. You define what good means as yes/no checks.
• One change at a time. Every mutation is isolated.
• Live dashboard. Watch the optimization happen in real time.
• Full research log. Every experiment is logged.
• It actually stops. Built-in convergence detection.