| name | skill-test |
| description | Auto-test a skill by generating scenarios and running them. Args: [skill-name]. Use when user says 'skill-test', 'test skill', 'validate skill'. |
| effort | slow |
| context | fork |
/skill-test — Skill Auto-Testing
Generates test scenarios for a skill and validates them through subagents.
Algorithm
1. Validate skill exists
Read the skill name from args. Search for SKILL.md in:
agents/skills/{name}/SKILL.md (core skills).claude/skills/{name}/SKILL.md (installed skills)
If not found — list available skills and stop with error.
2. Analyze skill
Read the SKILL.md file. Extract:
- Triggers: what phrases invoke it
- Algorithm: what steps it performs
- Edge cases: documented gotchas
- Inputs/outputs: what it expects and produces
3. Generate test scenarios
Create 3-5 test scenarios covering:
- Happy path: standard invocation with typical input
- Edge case: boundary condition or unusual input
- Error handling: missing required data, invalid state
- Integration: interaction with TAUSIK DB (tasks, sessions)
- Idempotency: running twice should not break state (if applicable)
Each scenario must have:
- Name: short descriptive label
- Setup: preconditions (e.g., "active task exists")
- Input: what the user says or passes
- Expected behavior: what should happen (not exact output — behavioral check)
4. Run scenarios
For each scenario, launch a subagent:
Agent(prompt: "You are testing the /{name} skill.
Scenario: {scenario.name}
Setup: {scenario.setup}
Input: {scenario.input}
Expected: {scenario.expected}
Read the skill at agents/skills/{name}/SKILL.md.
Simulate running this skill mentally (do NOT actually execute MCP tools or CLI).
Analyze: would the skill's algorithm produce the expected behavior given this input?
Check for: missing error handling, unclear instructions, logic gaps.
Report:
- PASS: skill handles this correctly
- FAIL: describe what would go wrong and why
- WARN: handles it but with potential issues",
subagent_type: "general-purpose")
5. Report results
Show a summary table:
| # | Scenario | Result | Details |
|---|
| 1 | Happy path | PASS | ... |
| 2 | Error handling | FAIL | Missing check for... |
Overall verdict:
- ALL PASS: "Skill is solid."
- Any FAIL: "Skill needs fixes. Suggest creating a task with
/plan fix-{name}-skill."
- WARN only: "Skill works but has minor gaps."
Gotchas
- This is a dry-run analysis, not live execution. Subagents analyze the SKILL.md logic, they don't run actual commands.
- Core skills (in
agents/skills/) should be edited at root, not in .claude/skills/ (generated copy).
- For live testing of skills, use
/dispatch with real task scenarios instead.
