Utility Pm Skill Validate

PM Skill Validate

This skill audits an existing pm-skills skill against the repo's structural conventions and quality criteria. It produces a validation report that a human can scan and that utility-pm-skill-iterate can consume as input.

The validator checks two tiers:

• Tier 1 (Structural) - deterministic checks that mirror CI: frontmatter, naming, file presence, description word count.
• Tier 2 (Quality) - LLM-assessed coherence checks: does the output contract reference the template? Is the example complete? Are checklist items testable?

When to Use

• After creating a skill with utility-pm-skill-builder, before shipping
• After manually editing a skill, to confirm it still passes conventions
• Before running utility-pm-skill-iterate, to identify what needs improvement
• When a convention changes, to audit which skills need updating (batch mode)
• When reviewing a contributed skill for quality and completeness

When NOT to Use

• To create a new skill from scratch -> use utility-pm-skill-builder
• To fix or improve a skill -> use utility-pm-skill-iterate (feed it this report)
• To run CI checks in a pipeline -> use scripts/lint-skills-frontmatter.sh (this skill is for interactive, deeper-than-CI validation)

Instructions

When asked to validate a skill, follow these steps:

Step 1: Identify the Target

Accept the skill name in any form:

• Directory name: deliver-prd
• Full path: skills/deliver-prd/SKILL.md
• Slash command: deliver-prd

Resolve to the canonical directory path: skills/{name}/.

If the skill directory does not exist, report immediately:

# Validation Report: {input}
Result: FAIL
Skill directory `skills/{input}/` does not exist.

Batch mode: If the input is --all, run Tier 1 structural checks across all skills and produce a summary table (see Step 5). Do not run Tier 2 in batch mode.

Step 2: Read Skill Files

Read all files in the skill directory:

Also read:

• The corresponding command file: commands/{command-name}.md
• The AGENTS.md entry for this skill

If reading files is not possible (MCP/embedded environment), ask the user to paste the content of each file before proceeding (see Degraded Mode).

Step 3: Run Tier 1 - Structural Checks

Run these deterministic checks. Each produces a PASS or FAIL line.

Step 4: Run Tier 2 - Quality Checks

Run these LLM-assessed checks. Each produces a PASS, WARN, or INFO line. Tier 2 findings are capped at WARN unless objectively grounded (placeholder leakage is the exception - it can FAIL).

Quality standard framing: These checks validate against current library conventions - what the shipped library actually does today. Findings graded WARN or INFO represent the v2.8 quality standard that newer skills (built with utility-pm-skill-builder) meet. Older skills may legitimately receive these findings until iterated through the lifecycle.

Step 5: Produce the Validation Report

Assemble the report using this exact structure. F-11 (utility-pm-skill-iterate) parses this report by section headings and pipe-delimited fields.

# Validation Report: {skill-name}
Date: {YYYY-MM-DD}
Skill version: {version from frontmatter}
Validator version: 1.0.0
Report schema: v1
Result: {PASS | WARN | FAIL}

## Summary
{1-2 sentence overall assessment.}
Errors: {n} | Warnings: {n} | Info: {n}

> Tier 2 findings are heuristic quality assessments and may require human review.

## Structural Checks
- {STATUS} | structural | {check-id} | {message}
- {STATUS} | structural | {check-id} | {message}
...

## Quality Checks
- {STATUS} | quality | {check-id} | {message}
- {STATUS} | quality | {check-id} | {message}
...

## Recommendations
1. {STATUS} | {check-id} | Target: {file-path}
   Action: {what to do}
2. {STATUS} | {check-id} | Target: {file-path}
   Action: {what to do}
...

Report rules:

• Result = worst severity found: any FAIL → FAIL, else any WARN → WARN, else PASS.
• Structural Checks: one line per Tier 1 check. STATUS is PASS or FAIL.
• Quality Checks: one line per Tier 2 check. STATUS is PASS, WARN, or INFO.
• Recommendations: only include checks that did NOT pass. Each recommendation includes the check ID, the target file path, and a specific action.
• If all checks pass, the Recommendations section should say: "No issues found."
• Omit passing checks from Recommendations - only list findings that need action.

Batch mode output (when input is --all):

Run Tier 1 structural checks only across all skills. Produce a summary table:

# Batch Validation Summary
Date: {YYYY-MM-DD}
Validator version: 1.0.0
Report schema: v1
Skills checked: {n}

| Skill | Result | Errors | Warnings |
|-------|--------|--------|----------|
| deliver-prd | PASS | 0 | 0 |
| define-hypothesis | WARN | 0 | 1 |
| foundation-persona | FAIL | 1 | 0 |
...

Skills passing: {n}/{total}
Run `utility-pm-skill-validate {skill}` for a detailed report.

Degraded Mode

If you cannot read skill files directly (e.g., running via MCP or in an embedded environment without file system access):

1. Ask the user to provide the content of each required file:
   • skills/{name}/SKILL.md
   • skills/{name}/references/TEMPLATE.md
   • skills/{name}/references/EXAMPLE.md
2. Run all checks against the provided content.
3. Note in the report: "Validated from user-provided content (file system not available)."
4. Batch mode is not available in degraded mode - single skill only.

Output Contract

The validator MUST produce a validation report following the format in Step 5.

The report:

• Uses the exact section headings: ## Summary, ## Structural Checks, ## Quality Checks, ## Recommendations
• Uses pipe-delimited check lines: STATUS | TIER | CHECK-ID | message
• Uses pipe-delimited recommendations: STATUS | CHECK-ID | Target: path followed by Action: description on the next line
• Includes Report schema: v1 in the header for F-11 compatibility
• Includes the Tier 2 caveat line in the Summary section

Quality Checklist

Before delivering the report, verify:

• All Tier 1 structural checks were run (not skipped), including the eval-asset checks (eval-trigger-fixtures, eval-output-scenarios, reciprocal-boundary-pointers); absent eval assets are INFO, not FAIL, except a one-directional declared collision pair which FAILs
• All Tier 2 quality checks were run (not skipped) - single skill mode only
• Report follows the exact section and line format from Step 5
• Every non-passing check appears in Recommendations with a target file path
• Result field reflects the worst severity found
• Tier 2 findings are capped at WARN (except placeholder-leakage which can FAIL)
• No Tier 2 check was marked FAIL unless objectively grounded

Examples

See references/EXAMPLE.md for a completed validation report demonstrating both Tier 1 and Tier 2 checks against a real shipped skill.