autoresearch

Autoresearch — Skill Self-Optimization

The system improves its own skills through iterative loops.

Process

1. Baseline

Run /eval on the target skill. Record trigger accuracy and fidelity.

2. Diagnosis

For each failure:

• Trigger miss → description too narrow
• False trigger → description too broad
• Fidelity fail → instructions in body incorrect

3. Mutation

ONE change per iteration. Never more than one variable at a time.

• For trigger: modify the description
• For fidelity: modify the body
• For stale eval: update the tests

4. Re-eval

Run /eval again. Compare with baseline:

• Improved? Keep the mutation.
• Same or worse? Revert, try a different mutation.

5. Report

Baseline → final score, what changed, how many iterations.

Rules

• One variable at a time
• The eval is the judge, not your opinion
• Max 5 iterations — if it does not converge, flag for human review
• Never change the purpose of the skill — only refine trigger/fidelity

Eval

Trigger Tests

"optimize my deploy skill" -> activates

"skill X is failing tests" -> activates

"improve skill accuracy" -> activates

"run eval" -> does NOT activate (that is /eval)

"deploy" -> does NOT activate

Fidelity Tests

Given skill with 2 trigger failures: mutates description, re-tests, reports improvement

Given skill with 0 failures: reports "already optimal, no mutations needed"

Given skill that does not converge in 5 iterations: flags for human review