| name | verify-skill |
| description | Use when validating a new or changed Claude Code skill, checking structural and repository policy compliance, testing whether descriptions trigger accurately, or grading representative skill outputs before deployment. |
| model | opus |
| context | fork |
Verify Skill
Boundaries
- Claude Code's validator owns manifest and frontmatter schema correctness.
quick_validate.py owns repository policies only; it must not duplicate the
evolving Claude schema.- Functional and trigger checks are required when behavior or discovery changed,
not for a wording-only edit with unchanged meaning.
Shared authoring policy lives at
${CLAUDE_SKILL_DIR}/../../constitution/references/authoring-invariants.md.
Inputs
- Required: a
SKILL.md, skill directory, plugin, marketplace, or plugins path.
- Optional mode:
structural, functional, or full (default full).
- Optional evaluation data:
evals/evals.yaml and representative prompts.
Workflow
- Resolve the target and enumerate affected skills and plugin roots.
- Resolve each affected plugin root from the target path, then run official
strict validation for every affected plugin. Report Claude's
errors verbatim enough to identify their source; do not reinterpret valid
fields through a private schema.
- Run
${CLAUDE_SKILL_DIR}/scripts/quick_validate.py on the target. Review body length,
description budget, unresolved local references, and placeholders.
- For functional or full mode, load existing evals or derive a small set from
the owned outcome. Follow
references/functional-mode.md for execution and
grading details.
- Test positive trigger prompts and nearby negative prompts. A pass requires
the intended skill to trigger without stealing work from its neighbors.
- Aggregate evidence by skill. When fixes are requested, change only reported
causes and rerun the failed checks.
The validator does not require fixed headings, personas, diagrams, subagent
prompts, XML report envelopes, or a section literally named “Skill Completion.”
Verification
claude plugin validate --strict <plugin-path>
python3 "${CLAUDE_SKILL_DIR}/scripts/quick_validate.py" <skill-or-plugin-path>
python3 "${CLAUDE_SKILL_DIR}/scripts/run_trigger_eval.py" --help
Run the first two commands after every fix iteration. Trigger and functional
results must include the prompt, observed behavior, and pass/fail rationale.
Completion
Return a concise per-skill verdict with official validation status, policy
issues, functional/trigger evidence when applicable, and actionable failures.
