Skip to main content

​
Task contract

Start from one concrete failure or explicit creation request. Preserve the current package, define expected results, and compare the candidate against a no-skill or previous-skill baseline in separate contexts.

​
Procedure

  1. Identify the exact skill revision and the observed problem.
  2. Add representative cases under evals/, including an edge case.
  3. Preview the proposed package diff.
  4. Run the same cases against both comparison arms.
  5. Inspect outputs and assertion evidence, not only aggregate scores.
  6. Recommend adoption only when the evidence supports it and regressions are understood.

​
Stop conditions

  • there is no representative failure or expected result;
  • the candidate changes unrelated package behavior;
  • comparison arms use different cases or environments;
  • assertions pass or fail identically in both arms without explanation; or
  • the user has not approved applying or publishing the candidate.

​
Expected result

Return the problem statement, exact revisions, cases, outputs, grading evidence, limitations, regressions, and a reversible recommendation. Do not call a candidate improved solely because a score increased.