utility-pm-skill-iterate（实用程序 pm 技能迭代）

PM Skill Iterate

This skill improves an existing pm-skills skill by applying targeted changes based on input you provide. It reads the current skill files, proposes changes as before/after previews grouped by file, and writes them on your confirmation. After applying changes, it suggests a version bump class and offers to update HISTORY.md.

The iterator accepts any of these as input:

• A validation report from /pm-skill-validate
• Direct feedback ("the template is missing section X")
• A convention change ("all skills now need a Limitations section")
• A general improvement request ("make the example more realistic")

When to Use

• After running /pm-skill-validate and getting a report with findings
• When you have specific feedback on a skill and want to apply it
• When a repo convention changes and a skill needs to conform
• When a skill's example, template, or instructions need improvement
• When iterating on a skill before a release

When NOT to Use

• To create a new skill from scratch -> use /pm-skill-builder
• To audit a skill before making changes -> use /pm-skill-validate first
• To make bulk convention changes across many skills -> run /pm-skill-validate --all first to triage, then iterate one skill at a time

Instructions

When asked to iterate on a skill, follow these steps:

Step 1: Identify the Target Skill

Accept the skill name in any form:

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

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

If the skill directory does not exist, stop and report: "Skill directory skills/{name}/ does not exist. Use /pm-skill-builder to create it."

Step 2: Read Current Skill Files

Read all files in the skill directory:

Record the exact content of each file at this point. You will compare against this content before writing in Step 5 (stale-preview guard).

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

Step 3: Normalize Input into Intended Changes

Regardless of input type, extract a structured list of intended changes before generating any edits. This normalization step is what makes the unified flow work consistently across all input types.

If the input is a validation report (from /pm-skill-validate):

• Check for Report schema: v1 in the header. If absent or a different schema version, warn: "This report uses an unrecognized schema. I'll do my best but may miss structured fields."
• Parse the ## Recommendations section.
• Split each recommendation line on | to extract:

- Position 1: severity (FAIL, WARN, INFO) - Position 2: check ID - After Target:: file path - After Action: (next line): what to change

• Build the intended changes list from these fields.

If the input is free text (feedback, convention change, improvement request):

• Read the input and identify what needs to change.
• Map each change to a specific target file and section.
• If the input is vague, ask ONE clarifying question before proceeding.

Present the normalized list for user confirmation:

Intended changes:
1. Target: skills/{name}/SKILL.md -> {section}
   Change: {what will change}
   Source: {validation report check ID | user feedback | convention change}
2. Target: skills/{name}/references/EXAMPLE.md -> {section}
   Change: {what will change}
   Source: {source}

If the user wants to modify the list (add, remove, or change items), adjust and re-present before proceeding.

Step 4: Preview Proposed Changes

For each intended change, generate the proposed edit and present it as a before/after block grouped by file:

### skills/{name}/SKILL.md

**{Section name} -- before:**
> {exact current content of the section being changed}

**{Section name} -- after:**
> {proposed new content for this section}

### skills/{name}/references/EXAMPLE.md

**{Section name} -- before:**
> {exact current content}

**{Section name} -- after:**
> {proposed new content}

Preview rules:

• Group all changes by file. Show each file once, with all its changes.
• Show enough surrounding context for the user to understand the change.
• For small edits (a few lines), show the full section before and after.
• For large edits (rewriting most of a section), show the section header and the first/last few lines of before, then the full after.
• Do NOT show files that are not being changed.

Ask: "Apply these changes? [yes / no]"

If the user says no, ask what to adjust and return to Step 3 or Step 4.

Step 5: Apply Changes (with Stale-Preview Guard)

Before writing any file, re-read each target file and compare its content to what you recorded in Step 2.

If any target file has changed since Step 2:

• Do NOT write any files.
• Report: "File {path} has changed since the preview was generated. Regenerating preview with current content."
• Return to Step 2 with the same intended changes list.

If all target files match:

• Write the changes to each target file.
• Update the updated field in SKILL.md frontmatter to today's date. (The updated field tracks when the file was last modified, regardless of whether a version bump is accepted.)
• Report what was written: list each file and a one-line summary of what changed.

Step 6: Suggest Version Bump

After changes are applied, classify the overall change and suggest a version bump class. Do NOT auto-write the version number.

Classification rules (from docs/internal/skill-versioning.md):

Tie-breaker: If a user must do something new to stay compliant with the skill's required contract, classify as major. If the new behavior is additive or optional, classify as minor. If the required behavior is unchanged and only clarified, classify as patch.

Present the suggestion:

Suggested bump: {class} ({reason}).
Current version: {current}.
Bump to {suggested}? [yes / override / skip]

• yes: Write the new version to SKILL.md frontmatter.
• override: Ask for the desired version, validate it's valid SemVer and higher than current, then write it.
• skip: Leave the version unchanged. The user may bump it later during release prep.

Step 7: Offer HISTORY.md Update

After the version decision, produce a change summary and handle HISTORY.md based on the current state:

If HISTORY.md exists and version was bumped:

1. Read HISTORY.md and validate its format:

- Has a summary table with | Version | Date | Release |... header - Versions are newest-first in the table - Each table version has a corresponding ## X.Y.Z section below

1. If format is valid: offer to append. "Would you like me to add this version to HISTORY.md? [yes / no]" On yes: add a new row to the summary table (newest-first) and a new ## X.Y.Z section with the change summary.
2. If format is invalid: warn and show the proposed content without writing. "HISTORY.md doesn't follow the expected format. Here's what I would add -- you can paste it manually:" Then show the proposed table row and version section.

If HISTORY.md does not exist and this is the skill's second version: Offer to create it. "This is the second version of this skill. Would you like me to create HISTORY.md with entries for both versions? [yes / no]" On yes: create HISTORY.md with the format from docs/internal/skill-versioning.md, including entries for both the original version (from release history or effort brief) and the new version.

If HISTORY.md does not exist and version was not bumped: No offer. HISTORY.md is premature until the skill has shipped a second version.

If HISTORY.md exists but version was not bumped (skip): No offer. The change summary is available in the conversation for the user to use at their discretion.

Step 8: Report Summary

Present a final summary:

## Iteration Complete: {skill-name}

**Files changed:**
- skills/{name}/SKILL.md -- {summary}
- skills/{name}/references/EXAMPLE.md -- {summary}

**Version:** {current} -> {new} ({class}) | or: unchanged (skipped)
**HISTORY.md:** updated | created | skipped | not applicable

**Next steps:**
- Run `/pm-skill-validate {name}` to verify the changes pass
- Run local CI: `bash scripts/lint-skills-frontmatter.sh`
- If satisfied, commit the changes

Degraded Mode

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

1. Validation-report-driven iteration is preferred in this mode. The report carries the context (check IDs, target paths, actions).
2. For free-text iteration, ask the user to paste the content of each relevant file before proposing changes.
3. The stale-preview guard (Step 5) cannot run -- note in the summary: "Applied without stale-preview check (file system not available)."
4. HISTORY.md operations require the user to paste the current file content or confirm it does not exist.

Output Contract

The iterator MUST:

• Normalize input into a structured intended-changes list before editing
• Present all proposed changes as before/after previews grouped by file
• Require explicit user confirmation before writing any file
• Re-read target files before writing to guard against stale previews
• Update the updated frontmatter field on every apply
• Suggest a version bump class without auto-writing the version number
• Handle HISTORY.md according to the rules in Step 7

The iterator MUST NOT:

• Write files without showing a preview first
• Write files without user confirmation
• Auto-increment the version number without explicit confirmation
• Create HISTORY.md for a skill still on its first version
• Append to HISTORY.md without validating its format first

Quality Checklist

Before completing the iteration, verify:

• Input was normalized into an intended-changes list before editing
• All proposed changes were shown as before/after previews
• User confirmed before any files were written
• Stale-preview guard ran before writing (or noted as unavailable)
• updated date was set to today in SKILL.md frontmatter
• Version bump class was suggested with correct reasoning
• Version was only written after explicit user confirmation
• HISTORY.md was handled correctly per Step 7 rules
• Final summary was presented with next steps

Examples

See references/EXAMPLE.md for a completed iteration demonstrating a validation-report-driven improvement to a real shipped skill.