Evaluate a Codex skill with a small task set

Choose three representative tasks before changing the skill. For each, record the input and the behavior you expect. Include a normal task, a missing-input task, and a task outside the skill’s purpose. These cases help expose problems that a single successful demonstration hides.

Keep examples small enough to inspect yourself. If you cannot judge the result, ask a knowledgeable reviewer to define the acceptance criteria before treating the skill as ready.

Note whether the procedure was used, whether it followed the supplied requirement, and whether it produced the required output. A review might need a concrete trigger, a file location, and a consequence for each supported finding. The absence of findings is acceptable when the evidence supports it.

Case:
Input:
Expected behavior:
Observed behavior:
Checks actually run:
Missing or unsupported claims:
Decision: keep / revise / investigate

When a case fails, identify whether the problem is selection, missing context, procedure, or output. Revise the smallest relevant section. Repeat the failed case and one neighboring case to check that the fix did not make the trigger too broad.

Keep the previous version of the skill and your results together. A version number without a record of the changed behavior tells a future maintainer very little.

Describe the environments and task types you actually evaluated. Do not call a procedure universally reliable after a few examples. A useful maintenance note says what worked, what remains uncertain, and what would trigger another evaluation.

• Download the evidence checklist
• Troubleshoot an unexpected result