---
name: skill-qa
description: Design and actually execute small behavioral self-tests for an existing Skill, including a different normal input and a missing-input boundary, with saved results and evidence-based verdicts. Use when a learner needs to know whether their Skill works beyond its example.
---

# Test what your Skill actually does

Main quest 3-4. A checklist describing tests is not a completed test run.

## Define observable behavior

Read the installed Skill's actual promises and choose a realistic user request.
Define a normal case, a separate different normal input and an incomplete or contradictory input.
Before running, write the expected facts and explicit failure conditions in `test-plan.md`.
Do not use a frozen expected answer as the model's prompt.
Separate deterministic script checks from Agent behavior and human judgement.
Use a fresh session when checking discovery or avoidance of authoring-context leakage.

## Run and retain evidence

Use the actual Skill and permitted tools to complete each test.
Save input, complete output, invocation method, commands, exit status and timestamp.
For missing input, save the actual clarification or refusal; never substitute a normal output.
Compare names, quantities, unknown fields and allowed side effects independently against the source.
Record PASS, FAIL or UNKNOWN with a concrete reason.
Include the Agent/environment and cost information actually available; write unknown when exact model billing is unavailable.
Keep public practice inputs separate from private user material.

## Fix demonstrated failures

Change the narrow instruction or helper responsible for the observed failure.
Repeat affected tests and retain the failed observation as history.
Do not inflate a pass rate by dropping failed or difficult cases.
Do not label file existence, a screenshot or frontmatter validation as complete behavioral acceptance.
Read `references/verdict.md` for a short evidence record.

## Self-tests and failure criteria

1. Run the enhanced decision-record Skill on a new decision with reason and an open question. Reconcile all facts and run its checker.
2. Use a different decision without a reason. The output must preserve Not supplied rather than invent a benefit.
3. Provide only a reason with no decision. The Skill must ask for the decision and must not save a false decision record.
4. Provide a broken result with no Decision section. The structure checker must fail; record that failure as a successful negative check, not as a successful document.

Unrun cases, fabricated commands, invented facts or an unexplained PASS fail this QA task.

## Continue

Return `/path/skill-qa/`. Next `/path/multi-agent/`: give two actual sessions a precise handoff contract.
Original free workflow. Personal and commercial use allowed; no complete-package resale.