Utility Pm Skill Iterate

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 utility-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 utility-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 utility-pm-skill-builder
• To audit a skill before making changes -> use utility-pm-skill-validate first
• To make bulk convention changes across many skills -> run utility-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: deliver-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 utility-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 utility-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.)
• Byte-0 preservation: verify each rewritten file still has --- at byte 0 of the file (no preceding content). If the previous content violated byte-0 placement (e.g., HTML attribution comment on line 1), surface this defensively before applying other changes and offer to fix the placement as part of the same write. Reference: library/skill-output-samples/SAMPLE_CREATION.md Section 5.
• 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.