技能 编写 工作流

Purpose

Create or update PM skills without chaos. This workflow turns rough notes, workshop content, or half-baked prompt dumps into compliant skills/<skill-name>/SKILL.md assets that actually pass validation and belong in this repo.

Use it when you want to ship a new skill without "looks good to me" roulette.

Key Concepts

Dogfood First

Use repo-native tools and standards before inventing a custom process:

• scripts/find-a-skill.sh
• scripts/add-a-skill.sh
• scripts/build-a-skill.sh
• scripts/test-a-skill.sh
• scripts/check-skill-metadata.py

Pick the Right Creation Path

• Guided wizard (build-a-skill.sh): Best when you have an idea but not final prose.
• Content-first generator (add-a-skill.sh): Best when you already have source content.
• Manual edit + validate: Best for tightening an existing skill.

Definition of Done (No Exceptions)

A skill is done only when:

1. Frontmatter is valid (name, description, intent, type)
2. Section order is compliant
3. Metadata limits are respected (name <= 64 chars, description <= 200 chars)
4. Description says both what the skill does and when to use it
5. Intent carries the fuller repo-facing summary without replacing the trigger-oriented description
6. Cross-references resolve
7. README catalog counts and tables are updated (if adding/removing skills)

Facilitation Source of Truth

When running this workflow as a guided conversation, use workshop-facilitation as the interaction protocol.

It defines:

• session heads-up + entry mode (Guided, Context dump, Best guess)
• one-question turns with plain-language prompts
• progress labels (for example, Context Qx/8 and Scoring Qx/5)
• interruption handling and pause/resume behavior
• numbered recommendations at decision points
• quick-select numbered response options for regular questions (include Other (specify) when useful)

This file defines the workflow sequence and domain-specific outputs. If there is a conflict, follow this file's workflow logic.

Application

Phase 1: Preflight (Avoid Duplicate Work)

1. Search for overlapping skills:

./scripts/find-a-skill.sh --keyword "<topic>"

2. Decide type:

• Component: one artifact/template
• Interactive: 3-5 adaptive questions + numbered options
• Workflow: multi-phase orchestration

Phase 2: Generate Draft

If you have source material:

./scripts/add-a-skill.sh research/your-framework.md

If you want guided prompts:

./scripts/build-a-skill.sh

Phase 3: Tighten the Skill

Manually review for:

• Clear "when to use" guidance
• One concrete example
• One explicit anti-pattern
• No filler or vague consultant-speak

Phase 4: Validate Hard

Run strict checks before thinking about commit:

./scripts/test-a-skill.sh --skill <skill-name> --smoke
python3 scripts/check-skill-metadata.py skills/<skill-name>/SKILL.md
python3 scripts/check-skill-triggers.py skills/<skill-name>/SKILL.md --show-cases

Phase 5: Integrate with Repo Docs

If this is a new skill:

1. Add it to the correct README category table
2. Update skill totals and category counts
3. Verify link paths resolve

Phase 6: Optional Packaging

If targeting Claude custom skill upload:

./scripts/zip-a-skill.sh --skill <skill-name>
# or zip one category:
./scripts/zip-a-skill.sh --type component --output dist/skill-zips
# or use a curated starter preset:
./scripts/zip-a-skill.sh --preset core-pm --output dist/skill-zips

Examples

Example: Turn Workshop Notes into a Skill

Input: research/pricing-workshop-notes.md
Goal: new interactive advisor

./scripts/add-a-skill.sh research/pricing-workshop-notes.md
./scripts/test-a-skill.sh --skill <new-skill-name> --smoke
python3 scripts/check-skill-metadata.py skills/<new-skill-name>/SKILL.md

Expected result:

• New skill folder exists
• Skill passes structural and metadata checks
• README catalog entry added/updated

Anti-Pattern Example

"We wrote a cool skill, skipped validation, forgot README counts, and shipped anyway."

Result:

• Broken references
• Inconsistent catalog numbers
• Confusion for contributors and users

Common Pitfalls

• Shipping vibes, not standards.
• Choosing workflow when the task is really a component template.
• Bloated descriptions that exceed upload limits.
• Descriptions that say what the skill is but not when Claude should trigger it.
• Descriptions that silently hit the 200-char limit and get cut off mid-thought.
• Letting intent become a substitute for a weak trigger description.
• Forgetting to update README counts after adding a skill.
• Treating generated output as final without review.

References

• README.md
• AGENTS.md
• CLAUDE.md
• docs/Building PM Skills.md
• docs/Add-a-Skill Utility Guide.md
• Anthropic's Complete Guide to Building Skills for Claude
• scripts/add-a-skill.sh
• scripts/build-a-skill.sh
• scripts/find-a-skill.sh
• scripts/test-a-skill.sh
• scripts/check-skill-metadata.py
• scripts/check-skill-triggers.py
• scripts/zip-a-skill.sh