Empirically tests a skill with subagent scenarios to verify it fires on the right prompts, stays silent on the wrong ones, and performs its job when loaded. Use when validating a new or modified SKILL.md before publishing - "test this skill", "does this trigger correctly", "verify the skill works". Do NOT use to statically grade or rewrite skill quality (use skill-auditor) - this runs live behavioral tests.
Click to play with sound.
---
name: Skill Tester
description: Empirically tests a skill with subagent scenarios to verify it fires on the right prompts, stays silent on the wrong ones, and performs its job when loaded. Use when validating a new or modified SKILL.md before publishing - "test this skill", "does this trigger correctly", "verify the skill works". Do NOT use to statically grade or rewrite skill quality (use skill-auditor) - this runs live behavioral tests.
---
# Skill Tester
Verifies empirically that a skill fires on the right prompts, stays silent on unrelated ones, and produces the behavior its body promises when loaded. Static review (skill-auditor) judges the text; this skill judges the behavior. A skill that reads perfectly and never fires is still broken.
## Inputs to collect
1. The SKILL.md under test, from source.
2. The neighbor skills whose territory borders it (for negative scenarios).
3. The intended user phrasing, if the author collected any - real phrasing makes the best positive scenarios.
## Operating procedure
### Step 1 - Write the scenario matrix
For each skill, write at minimum:
- **Three positive trigger scenarios** - prompts that must fire the skill. Paraphrase the description's WHEN clause in different words; do not copy it verbatim, or you test string matching instead of routing.
- **Two negative trigger scenarios** - adjacent tasks the skill must NOT hijack. The best negatives are the named neighbors from the description's NOT clause; the second-best are near-miss requests one topic over.
- **Two performance scenarios** - full realistic requests where you will judge the output against the skill's own quality bar and deliverable sections.
… install to load the full skill