Skill From Masters
Create skills that embody the wisdom of domain masters. This skill helps users discover and incorporate proven methodologies from recognized experts before generating a skill.
Core Philosophy
Most professional domains have outstanding practitioners who have codified their methods through books, talks, interviews, and frameworks. A skill built on these proven methodologies is far more valuable than one created from scratch.
The goal is not just "good enough" — it's reaching the highest level of human expertise in that domain.
Critical Requirements for Non-Technical Skills
Technical skills have standard answers. Writing code, debugging, or configuring systems — these have relatively objective quality bars.
Non-technical skills vary dramatically in quality. Skills involving decision-making, communication, persuasion, or judgment can range from mediocre to world-class. The difference comes from incorporating deep expertise.
For non-technical skills (writing, sales, hiring, product decisions, etc.), follow these requirements:
1. Narrow, Specific Task Definition ⚠️ CRITICAL
- The task must be extremely specific and well-defined
- ❌ BAD: "Write a sales email" (too broad)
- ✅ GOOD: "Write a B2B cold outreach email to enterprise CTOs"
- ✅ GOOD: "Write a project status report email to executive stakeholders"
- Different contexts require completely different skills
- If the user's request is too broad, help them narrow it down through questions
2. Model Selection: Opus Required 🎯 MANDATORY
- Non-technical skills MUST use Claude Opus (claude-opus-4-5)
- DO NOT use Sonnet, Haiku, or any other model
- Opus has the reasoning depth needed for nuanced, judgment-based tasks
- The quality difference is substantial for these domains
3. Methodology Research: Clear & Reliable Conclusions 🔍 ESSENTIAL
- Continue searching and communicating until you reach very clear, reliable conclusions
- Don't stop at surface-level research
- Sources to exhaust:
- The model's own training knowledge
- Web search for current best practices
- Golden examples from top practitioners
- Counter-examples and common mistakes
- Keep iterating until the methodology is crystal clear and well-validated
4. Consider Plan Mode for Complex Tasks 🎯 RECOMMENDED
- For complex or multi-faceted skills, prefer thinking through the approach first
- Better to think more before acting
- Use plan mode to structure the methodology research and synthesis
5. Test Broadly, Then Iterate ✅ REQUIRED
- Have the agent think through extensive test scenarios
- Test across diverse contexts, edge cases, and failure modes
- Review test results and optimize before finalizing
- Quality emerges from iteration, not first drafts
Workflow
Before Starting: Consider Plan Mode 🎯
For complex or high-stakes skill creation (especially non-technical skills), consider using plan mode:
- Allows more upfront thinking before taking action
- Helps structure the methodology research systematically
- Reduces the risk of missing important considerations
- Better for skills involving judgment, persuasion, or complex decision-making
To use plan mode, the user can invoke it explicitly, or you can suggest: "This is a complex skill involving [decision-making/communication/etc]. Would you like me to use plan mode to think through the methodology research more carefully?"
Step 1: Understand and Narrow the Skill Intent
CRITICAL FOR NON-TECHNICAL SKILLS: Ensure the task is narrow and specific enough.
Most users will start with a broad request. Your job is to help them narrow it down systematically until the task is specific enough that methodology and quality criteria are unambiguous.
The 5-Layer Narrowing Framework
Use this systematic approach to guide users from broad to specific:
Layer 1: Domain Identification
Identify the core domain(s) of the broad task.
Template questions:
"[Broad task] can mean different things depending on the domain. Which of these is closest to what you need?
- [Domain A - description]
- [Domain B - description]
- [Domain C - description]
- Something else?"
Example:
- User: "I want a skill for making decisions"
- You: "Decision-making varies greatly by domain. Which is closest?
- Product decisions (features, roadmap, prioritization)?
- Hiring decisions (candidate evaluation, role definition)?
- Investment decisions (startups, projects, resource allocation)?
- Strategic business decisions (market entry, partnerships)?
- Something else?"
Layer 2: Context Constraints (5W1H)
Once you know the domain, use structured questions to narrow the context:
Template questions:
- **WHO**: Who will use this skill? (Role, seniority, expertise level)
- **WHAT**: What specific output or decision does it produce?
- **WHERE**: What organizational context? (Startup, enterprise, remote, industry)
- **WHEN**: At what stage/timing is it used? (Launch, crisis, quarterly, daily)
- **WHY**: What's the primary goal/outcome? (Alignment, approval, action, learning)
- **HOW**: What constraints exist? (Time limits, process requirements, stakeholders)
Example:
- User: "Product decisions"
- You: "Let me understand the context better:
- WHO: Are you a PM, founder, or executive making these decisions?
- WHAT: Is this about choosing features, prioritization, or go/no-go launches?
- WHERE: Startup (limited resources) or enterprise (many stakeholders)?
- WHEN: Daily sprint planning or quarterly roadmap decisions?
- WHY: Main goal - team alignment, exec approval, or customer validation?"
Layer 3: Comparative Narrowing
Present 2-3 similar but distinct scenarios. Force a choice or clarification.
Template:
"Within [narrowed domain], I want to make sure we're targeting the right scenario. Which is closest?
A. [Specific scenario A - with context]
B. [Specific scenario B - with context]
C. [Specific scenario C - with context]
D. None of these / Combination
Which matches best?"
Example:
-
User: "Feature prioritization for a PM at a B2B SaaS startup"
-
You: "Got it. Let me narrow further:
A. Weekly sprint planning: Prioritizing stories within a committed roadmap
B. Quarterly roadmap planning: Deciding which themes/initiatives for next quarter
C. Reactive prioritization: Customer escalation or urgent bug vs planned work
D. Resource allocation: Multiple products/teams competing for engineering time
Which scenario is most common for you, or do you need to handle multiple?"
Layer 4: Boundary Validation (Via Negativa)
Confirm what the skill does NOT include. This clarifies the edges.
Template:
"To make sure we're aligned, let me confirm the boundaries. This skill:
- ✅ INCLUDES: [Specific scenario X]
- ✅ INCLUDES: [Specific scenario Y]
- ❌ EXCLUDES: [Related but different scenario A]
- ❌ EXCLUDES: [Related but different scenario B]
- ❌ EXCLUDES: [Related but different scenario C]
Does this match your needs?"
Example:
-
User confirms: "Quarterly roadmap planning for a B2B SaaS PM"
-
You: "Perfect. To confirm scope, this skill:
- ✅ INCLUDES: Quarterly planning with engineering constraints
- ✅ INCLUDES: Balancing customer requests, tech debt, and strategic bets
- ❌ EXCLUDES: Daily sprint story prioritization
- ❌ EXCLUDES: Crisis/reactive priority decisions
- ❌ EXCLUDES: Multi-product portfolio prioritization (VP-level)
Is this right?"
Layer 5: Concrete Case Anchoring
Ask for a real, specific example from their experience. This locks in the scope.
Template:
"Can you describe a specific recent situation where you needed this skill? Include:
- What was the situation/context?
- What inputs did you have?
- What decision/output did you need to produce?
- What made it difficult or time-consuming?
- What would an ideal outcome have looked like?"
Example:
- User: "Yes, the scope is right"
- You: "Great! One last thing to anchor our methodology research. Can you describe a recent quarterly planning session? For example:
- How many features/initiatives were you choosing between?
- What stakeholders were involved?
- What made it hard? (Conflicting data? Stakeholder disagreement? Technical uncertainty?)
- How long did the process take?
- What would have made it easier or better?"
Stop Condition: Is It Narrow Enough?
Stop narrowing when you can answer YES to all:
- ✅ Unique methodology: Would experts in this specific scenario have unique advice (not generic)?
- ✅ Clear quality bar: Could someone judge if the output is "excellent" vs "mediocre"?
- ✅ Specific constraints: Are there context-specific rules, tradeoffs, or failure modes?
- ✅ Concrete example: Has the user described a real scenario where they'd use this?
- ✅ Excludes alternatives: Is it clear what related tasks this does NOT cover?
If ANY answer is NO, keep narrowing.
Common Mistakes: Still Too Broad
Even after narrowing, watch for these signs the scope is still too broad:
❌ Too broad:
- "Write better emails" → Includes too many email types
- "Make product decisions" → Covers too many decision types
- "Create marketing content" → Content types vary wildly
- "Improve team communication" → Communication contexts differ greatly
✅ Narrow enough:
- "Write B2B cold outreach emails to enterprise CTOs"
- "Quarterly roadmap prioritization for B2B SaaS PMs with 3-5 eng team"
- "Create LinkedIn thought leadership posts for technical founders"
- "Run effective incident postmortems for distributed systems teams"
Rule of thumb: If you can describe the skill in one sentence with specific role, context, and output type, you're probably narrow enough.
Quick Reference: Narrowing Question Flow
Broad Request
↓
Layer 1: "Which domain?" → [Pick one]
↓
Layer 2: "5W1H context?" → [Answer constraints]
↓
Layer 3: "Which specific scenario?" → [Choose from 2-3 options]
↓
Layer 4: "What's excluded?" → [Confirm boundaries]
↓
Layer 5: "Give me a real example" → [Describe concrete case]
↓
Check Stop Condition → [All 5 YES?]
↓
✅ Narrow enough → Proceed to Step 2
❌ Still broad → Continue narrowing
Step 2: Identify Skill Type
CRITICAL: Different skill types require fundamentally different methodologies and quality criteria.
Consult
references/skill-taxonomy.md
for the full taxonomy. The core types are:
| Type | Core Operation | Key Question |
|---|
| Summary | Compress | Need comprehensive coverage? |
| Insight | Extract | Need to find what really matters? |
| Generation | Create | Need new content created? |
| Decision | Choose | Need to make a choice? |
| Evaluation | Judge | Need quality judgment? |
| Diagnosis | Trace | Need to find root cause? |
| Persuasion | Bridge | Need to change someone's mind? |
| Planning | Decompose | Need a roadmap? |
| Research | Discover | Need knowledge gathered? |
| Facilitation | Elicit | Need to extract info from others? |
| Transformation | Map | Need format conversion? |
How to Identify:
Ask the user: "Based on what you described, this sounds like a [Type] skill—the goal is to [core operation]. Is that right?"
Common Confusions to Clarify:
- Summary vs Insight: "Do you need comprehensive coverage, or just the key signals that matter?"
- Decision vs Evaluation: "Do you need to make a choice, or judge the quality of something?"
- Research vs Insight: "Do you need to gather information, or interpret what it means?"
Why This Matters:
Each type has different:
- Methodology sources to draw from
- Quality criteria to evaluate output
- Output format conventions
Document the identified type before proceeding.
Step 3: Identify Relevant Domains
Map the skill to one or more methodology domains. A single skill may span multiple domains.
Example mappings:
- "Sales email skill" → Sales, Writing, Persuasion
- "User interview skill" → User Research, Interviewing, Product Discovery
- "Presentation skill" → Storytelling, Visual Design, Persuasion
- "Code review skill" → Software Engineering, Feedback, Communication
Step 4: Surface Expert Methodologies (Until Crystal Clear)
GOAL: Don't stop until you have very clear, reliable conclusions about the best methodology.
Layer 1: Local Database
Consult
references/methodology-database.md
for known frameworks.
Layer 2: Web Search for Experts
Search the web to discover additional experts and methodologies:
- Search: "[domain] best practices expert"
- Search: "[domain] framework methodology"
- Search: "[domain] master practitioner"
Layer 3: Deep Dive on Selected Experts
For promising experts, search for their original content:
- Search: "[expert name] methodology interview"
- Search: "[expert name] [domain] transcript"
- Search: "[expert name] framework explained"
Fetch and read primary sources when available (articles, talk transcripts, blog posts).
Layer 4: Keep Iterating Until Clear ⚠️ NEW
- Don't stop at the first search results
- If methodologies seem unclear or conflicting, dig deeper
- Look for:
- Model's own knowledge (you have extensive training data)
- Current web best practices
- Golden examples from practitioners
- Anti-patterns and common mistakes
- Continue the research loop until you can confidently say: "This is the proven way to do this"
For each relevant domain, present:
- Key experts and their core contributions
- Specific frameworks, principles, or processes
- Source materials (books, talks, interviews)
- Confidence level in the methodology (keep searching if low)
Step 5: Find Golden Examples
Before finalizing methodology selection, search for exemplary outputs:
- Search: "best [output type] examples"
- Search: "[output type] template [top company]"
- Search: "award winning [output type]"
Understanding what excellence looks like helps define the quality bar.
Step 6: Collaborative Selection
Present the methodologies to the user and discuss:
- Which frameworks resonate with their goals?
- Are there conflicts between methodologies to resolve?
- Should they combine multiple approaches?
- Any specific principles they want to emphasize or exclude?
Guide the user to select 1-3 primary methodologies that will form the skill's foundation.
Step 7: Extract Actionable Principles
For each selected methodology, search for and distill:
The Why (Core Principles)
- Search: "[methodology] core principles"
- Search: "why [methodology] works"
The How (Concrete Process)
- Search: "[methodology] step by step"
- Search: "[methodology] implementation guide"
The What (Quality Criteria)
- Search: "[methodology] checklist"
- Search: "[methodology] evaluation criteria"
The Pitfalls (Common Mistakes)
- Search: "[domain] common mistakes"
- Search: "[methodology] pitfalls avoid"
Fetch primary sources to get exact wording and nuance, not just summaries.
Step 8: Cross-Validate
Compare insights across multiple sources:
- What principles appear consistently? (high confidence)
- Where do experts disagree? (flag for user)
- What's unique to each approach? (differentiation)
Synthesize a coherent framework that takes the best from each source.
Step 9: Design Test Scenarios (Before Generation)
CRITICAL: Before generating the skill, design comprehensive test scenarios.
Work with the user to identify:
Diverse Test Cases:
- Typical scenarios (the common case)
- Edge cases (unusual but valid situations)
- Boundary conditions (where the methodology might break down)
- Failure modes (what could go wrong)
Context Variations:
- Different user expertise levels
- Different organizational contexts (startup vs enterprise)
- Different constraints (time, resources, stakeholder complexity)
- Cultural or industry differences
Quality Validation:
- What does "excellent" output look like?
- What are the most common mistakes to avoid?
- How will we know if the skill is working?
Document these test scenarios — they'll be used after generation to validate and iterate.
Step 10: Generate the Skill
With methodologies confirmed and test scenarios designed,
invoke the skill to generate the final skill with proper format.
HOW TO INVOKE:
Use the Skill tool with: skill: "skill-creator:skill-creator"
This ensures:
- Proper YAML frontmatter (name, description)
- Correct directory structure
- Validation before packaging
- Imperative writing style (not second person)
For non-technical skills, CRITICAL:
- Add in the YAML frontmatter
- This ensures the skill uses Claude Opus, not a weaker model
The generated skill should:
- Credit the methodology sources in a comment (documenting provenance)
- Translate expert wisdom into actionable instructions
- Include concrete examples derived from golden examples found
- Capture quality criteria as explicit checkpoints
- Include "don't do this" anti-patterns from pitfall research
- Match the quality bar of the best human practitioners
- Include the test scenarios as part of the skill's self-validation
Step 11: Test, Review, and Iterate
Don't stop at first generation. Quality emerges through iteration.
- Run Test Scenarios: Apply the skill to each test case designed in Step 9
- Evaluate Results: Compare outputs against quality criteria
- Identify Gaps: Where did the skill fall short?
- Refine Methodology: Do we need additional expert guidance?
- Regenerate: Update the skill based on learnings
- Repeat: Until the skill consistently produces excellent results
Involve the user in this evaluation — they know their domain and can spot nuances.
Quality Checklist
Before finalizing, verify:
Scope & Clarity:
Methodology Depth:
Generation Quality:
Testing & Iteration:
Plan Mode (if applicable):
Example Interaction: Full Narrowing Process
This example shows the complete 5-layer narrowing in action.
User: I want to create a skill for writing product requirement documents.
Assistant: Great! PRDs are an important skill. Let me help narrow this down using a systematic approach.
[Layer 1: Domain Identification]
PRDs can serve different purposes depending on the audience and goal. Which is closest to what you need?
- Engineering-focused specs: Detailed feature requirements for implementation
- Stakeholder alignment docs: High-level documents to get buy-in from executives/leadership
- Design briefs: Product requirements focused on user experience and design direction
- External vendor specs: Technical requirements for third-party development
Which type of PRD do you most often need to write?
User: Engineering-focused specs.
Assistant: Perfect. Now let me understand the context better.
[Layer 2: Context Constraints (5W1H)]
- WHO: What's your role? (PM, Senior PM, Product Lead, or Founder?)
- WHERE: What type of company? (Startup, mid-size, or enterprise? B2B or B2C?)
- WHEN: How often do you write these? (Every sprint? Per quarter? Per major initiative?)
- WHY: Primary goal - is it to get eng buy-in, ensure nothing is missed, or facilitate estimation?
- HOW: Any constraints - do you follow a specific template? Any required sections?
User: I'm a Senior PM at a B2B SaaS company (100 people, growing fast). I write PRDs for major features (usually 3-5 per quarter). Main goal is to ensure engineers have everything they need to start building without constant back-and-forth. No required template, but we try to keep docs under 2 pages.
Assistant: Excellent context. Let me narrow further.
[Layer 3: Comparative Narrowing]
Within "engineering PRDs for major features," which scenario is most common?
A. Net-new features: Building something completely new (new product area, new capability)
B. Major enhancements: Significant upgrades to existing features (new workflow, architecture change)
C. Cross-system features: Features requiring coordination across multiple teams/services
D. Customer-driven builds: Custom features for enterprise customers with tight timelines
Which is most frequent, or do you need to handle multiple types?
User: Mostly A and B. Sometimes C, but rarely D.
Assistant: Got it. Let me confirm the boundaries.
[Layer 4: Boundary Validation]
So this skill should:
- ✅ INCLUDE: Net-new features and major enhancements for engineering teams
- ✅ INCLUDE: B2B SaaS context with 3-10 engineer scope typically
- ✅ INCLUDE: Cross-team coordination when needed
- ❌ EXCLUDE: Small story-level specs (those go directly in Jira/tickets)
- ❌ EXCLUDE: Customer-specific custom builds with tight deadlines
- ❌ EXCLUDE: Executive-facing strategy documents or roadmap presentations
Does this match what you need?
User: Yes, exactly!
Assistant: Perfect. One final question to anchor the methodology research.
[Layer 5: Concrete Case Anchoring]
Can you describe a recent PRD you wrote that was either particularly challenging or time-consuming? For example:
- What feature were you speccing?
- What made it hard to write the PRD?
- How long did it take you?
- Did engineers come back with lots of questions, or did it work well?
User: Last month I wrote a PRD for an API rate limiting feature. Took me about 6 hours spread over 3 days. The hard part was explaining the business logic (what counts toward limits, how to handle bursts) clearly enough that engineers understood the intent but could still design the technical solution. We did 2 rounds of clarification before they felt ready to estimate. Would love to get that down to 3 hours and one round of Q&A.
Assistant: Perfect! Now I have a crystal-clear picture.
📌 Final Skill Definition:
"Create engineering-focused PRDs for B2B SaaS senior PMs writing specs for net-new features and major enhancements. PRDs should be 1-2 pages, clearly communicate business logic and user needs without over-specifying technical implementation, and enable engineering teams to estimate and start building with minimal back-and-forth."
This passes all 5 stop conditions:
- ✅ Unique methodology - Yes (balancing business logic vs technical freedom)
- ✅ Clear quality bar - Yes (minimize eng questions, enable estimation)
- ✅ Specific constraints - Yes (B2B SaaS, 1-2 pages, major features)
- ✅ Concrete example - Yes (API rate limiting feature, 6→3 hours)
- ✅ Excludes alternatives - Yes (not small stories, not exec docs)
Ready to move to Step 2: Methodology Research. Since this is a non-technical skill involving communication and judgment, I'd recommend using plan mode to think through the methodology research carefully. Should I proceed in plan mode?
User: Yes, go ahead.
Assistant: Perfect, that's specific enough. Since this is a non-technical skill involving communication and judgment, I'd recommend using plan mode to think through the methodology research carefully. Should I proceed in plan mode?
User: Yes, go ahead.
Assistant: [Enters plan mode and structures the research approach]
Let me search systematically for the best methodologies on engineering-focused PRD writing.
[Searches: "PRD best practices expert", "product requirements document methodology for engineers", "best PRD examples B2B SaaS"]
I found several strong frameworks:
Product Management:
- Marty Cagan (SVPG): Emphasizes problem-first PRDs, separating problem space from solution space
- Teresa Torres: Continuous discovery, outcome-focused requirements
Business Writing:
- Barbara Minto: Pyramid Principle - lead with conclusion, MECE structure
- Amazon: Working backwards from press release, 6-pager narrative format
Golden Examples Found:
- Intercom's public PRD template
- Linear's product spec format
Common Mistakes (from my search):
- Jumping to solutions before defining the problem
- Writing for engineers instead of for alignment
- Missing success metrics
Let me dig deeper into primary sources to get very clear conclusions...
[Continues searching until methodology is crystal clear]
Which of these approaches resonates? Or should I search deeper into any of them?
User: I like Amazon's approach combined with Cagan's problem-first thinking. Can you find more details on Amazon's method?
Assistant: [Searches: "Amazon 6-pager template", "Amazon working backwards methodology", "Amazon PR FAQ process"]
[Fetches relevant articles and synthesizes key principles]
Here's what I found from primary sources...
[After thorough methodology research]
Assistant: Before I generate the skill, let me design test scenarios. Let's think through diverse cases:
Test Scenarios:
- Simple feature: Adding a filter to existing dashboard
- Complex feature: New AI-powered recommendation engine
- Edge case: Feature requiring significant backend architecture changes
- Failure mode: Vague problem statement that needs refinement
Does this cover the range of PRDs you typically write? Any other scenarios we should test?
User: Also add: Feature that needs cross-team coordination.
Assistant: Great addition. Now I'll invoke skill-creator to generate the skill with proper format.
[Invokes: Skill tool with skill: "skill-creator:skill-creator"]
The skill-creator will guide me through:
- Creating proper YAML frontmatter (name, description, model: opus)
- Writing in imperative form
- Structuring the skill directory
- Validating the output
[After skill-creator completes]
I've generated the skill. Let me test it against our scenarios...
[Reviews results, identifies gaps, iterates]
[Continues to Step 10: refinement based on test results]