Utility Pm Skill Builder

<!-- PM-Skills | https://github.com/product-on-purpose/pm-skills | Apache 2.0 -->

PM Skill Builder

This skill creates new PM skills for the pm-skills library. It produces a Skill Implementation Packet - a complete design document with draft files in a staging area for review before promotion to canonical locations.

When to Use

• When you have an idea for a new PM skill
• When you want to add a domain skill (phase-specific), foundation skill (cross-cutting), or utility skill (meta/tooling) to the pm-skills library
• When a contributor needs guided skill creation that follows repo conventions

When NOT to Use

• To modify or improve an existing skill → use utility-pm-skill-iterate
• To audit an existing skill against conventions → use utility-pm-skill-validate
• To create a skill for a non-pm-skills context → use a general agent skill builder
• To create a workflow → workflows are authored directly, not via this builder

Instructions

When asked to create a new PM skill, follow these steps:

Step 1: Understand the Idea

Accept the idea in either form:

• Problem-first: "What PM problem does this skill solve? Who runs into this problem, and what do they currently produce (or fail to produce)?"
• Skill-first: "Describe the skill you want to create. What artifact does it produce? What PM activity does it support?"

Both entry points produce the same downstream flow. If the user provides one form, do not ask for the other - extract what you need and proceed.

If the idea is vague, ask ONE follow-up question to clarify the artifact type and target audience before proceeding.

Step 2: Gap Analysis

Check ALL existing skills for overlap. Use the Current Library Reference below (derive the live inventory from skill-manifest.json or AGENTS.md) AND scan the skills/ directory for the latest inventory. Include all four families - domain, foundation, utility, and tool - the tool family is easy to miss since it has no phase prefix pattern in common conversation.

Present findings with specificity:

• Name each overlapping skill and explain what it covers
• Identify the specific gap this new skill would fill
• If overlap is high, trigger the Why Gate (see below)

Why Gate (triggers when overlap is found): Ask the user: "Name 2-3 specific prompts or scenarios where the existing skills fail to produce what you need."

Kill Gate: If the user cannot articulate convincing gaps, recommend an alternative:

• "Revise [existing skill] to cover this case"
• "Create a workflow combining [skill A] + [skill B]"
• "Add a command variant, not a new skill"
• "This is a documentation improvement, not a new skill"

Do not proceed past the kill gate without either convincing evidence of a gap or explicit user override.

Step 3: Scope Check

Evaluate whether the idea should be ONE skill or MULTIPLE skills.

Splitting signals:

• The idea produces multiple distinct artifact types
• The idea crosses Triple Diamond phases (e.g., Discover + Deliver)
• The description naturally contains "and" connecting two activities

If splitting is warranted, present the recommendation: "This seems to cover two distinct PM activities:

1. [Activity A] → produces [Artifact A]
2. [Activity B] → produces [Artifact B] These work better as separate skills that can be chained via a workflow. Want to proceed with just [Activity A] for now?"

Step 4: Classification + Repo-Fit

Determine the skill's classification and naming:

Domain skills (phase-specific PM activities):

• Phase: discover | define | develop | deliver | measure | iterate
• Directory: {phase}-{skill-name}
• Frontmatter: phase: {phase} (required), no classification field

Foundation skills (cross-cutting, used across phases):

• No phase
• Directory: foundation-{skill-name}
• Frontmatter: classification: foundation (required), no phase field
• Use when: the skill applies to multiple phases equally

Utility skills (meta-skills, repo tooling):

• No phase
• Directory: utility-{skill-name}
• Frontmatter: classification: utility (required), no phase field
• Use when: the skill operates on the repo, workflow, or other skills

Dialect and exemplar selection: First, pick the skill's dialect from Skeleton Canon: The Three Sanctioned Dialects (classic, contract-shaped, or tool-family) - this fixes the required heading skeleton before any exemplar is chosen. Do not default to "mirror the closest exemplar": an exemplar can itself carry drift (a case variant, an extra or missing section), and a straight mirror copies that drift forward into the new skill.

Then identify 1-2 existing skills in the SAME dialect that are the closest structural match:

• Same phase > same category > similar artifact type
• Read their SKILL.md to confirm instruction style, output-contract wording, and quality-checklist pattern (the canon fixes the required headings; the exemplar informs everything else - tone, step granularity, and any skill-specific extra sections)
• Name the exemplars explicitly: "Modeled after [skill] - same phase, [category] category, [dialect] dialect"

Present the classification, dialect, and exemplar selection for user confirmation.

Step 4.5: Eval Readiness (the eval contract)

A new skill must ship eval-ready so coverage never falls behind: routing health and output quality are both measurable from day one. Decide the eval contract here, then emit its assets in Step 5. Four parts:

A. Nearest neighbors (C-2). Name the new skill's 1-3 nearest neighbors - the existing skills whose triggers most overlap. Derive them from the Step 2 gap analysis plus same-phase / same-category siblings. These neighbors drive the boundary pointers (below), the near-miss trigger fixtures (Step 5), and the collision probe (Step 7).

B. Reciprocal boundary pointers (C-3). The draft SKILL.md MUST include a ## When NOT to Use section that names each neighbor and says when to use that neighbor instead. For every neighbor, also add the reciprocal pointer back: a "When NOT to Use" bullet in the neighbor's SKILL.md pointing to the new skill. If the overlap is strong enough to be a measured collision pair, add the pair to COLLISION_PAIRS in scripts/check-trigger-fixtures.mjs so the reciprocity gate (check-reciprocal-boundary-pointers.mjs) and the collision probe both cover it. Reciprocal pointers are what kept the v2.26.0 rewrites collision-clean.

C. Output-eval family (C-4). Map the skill to an output-eval family rubric so its artifact quality is measurable. Pick by phase/category:

Family rubrics live at docs/internal/eval-rubrics/{family}.md. If the skill fits a family that has a rubric, use that family value in the scenario frontmatter. If it opens a NEW family (meeting/tool skills have none yet), note "no family rubric yet - author one before the skill enters the output-eval roster" and skip the family value.

D. Fixture + scenario plan (C-1, C-4). Plan the two eval-asset files emitted in Step 5: a trigger-fixtures.json (routing) and an output-scenarios/<id>.md (output quality). Their shapes are specified in Step 5 items 13-14.

Present the neighbors + family + reciprocal-pointer plan for user confirmation before generating the packet.

Step 5: Generate Skill Implementation Packet

Produce the complete packet using references/TEMPLATE.md as the format. The packet includes:

1. Decision - recommendation + Why Gate evidence (if applicable)
2. Classification - type, phase (if domain), category, directory name
3. Overlap Analysis - what was found, why this skill is still needed
4. Exemplar Skills - which existing skills modeled, why
5. Draft Frontmatter - complete, valid YAML block. The frontmatter MUST begin with --- at byte 0 of the file (no preceding content of any kind, including HTML comments, BOM, or whitespace). Place any attribution comment AFTER the closing --- fence, never before. Reference: library/skill-output-samples/SAMPLE_CREATION.md Section 5.
6. Draft SKILL.md - full content (not an outline), following the Step 4 dialect's canon heading skeleton and the chosen exemplars' instruction style
7. Draft TEMPLATE.md - section headers with guidance comments
8. Draft EXAMPLE.md - complete, realistic example (150-300 lines) with a specific PM scenario, every section filled, optional sections demonstrated both filled and skipped
9. Draft Command - command frontmatter
10. AGENTS.md Entry - exact text to add
11. Validation Checklist - all CI rules checked against the draft
12. Next Steps - local CI, testing, contribution workflow
13. Draft trigger-fixtures.json (C-1, routing eval) - evals/trigger-fixtures.json: a JSON object { "schema": 1, "skill": "{name}", "runs_per_query": 3, "trigger_threshold": 0.5, "queries": [...] }. The queries array needs >= 16 total, >= 8 with "expect": "trigger" (drawn from the skill's real intents, NOT just artifact keywords - include intent-only asks) and >= 8 with "expect": "no-trigger", of which >= 2 are near-misses aimed at the Step 4.5 neighbors (mark them "near_miss_of": "{neighbor}"). Split each class ~60/40 across "split": "train" / "split": "validation". This is the B-4 fixture contract (scripts/check-trigger-fixtures.mjs).
14. Draft output-scenario (C-4, output-quality eval) - evals/output-scenarios/{id}.md: frontmatter scenario: {id} / skill: {name} / family: {family from Step 4.5} / created: {date}, then a realistic input brief (>= 100 chars of body) that gives the skill arm and a freehand control the same raw material. This is the B-7 asset contract (scripts/check-output-eval-assets.mjs). Omit family only if Step 4.5 found no rubric yet.