Writing Skills

Core Principle

Skills are reusable process documentation. Write them like code: test-first, iterative, validated.

When to Use This Skill

•Discovered a useful process
•Want to codify knowledge
•Pattern worth sharing
•Repeatedly solving same problem
•Need consistent approach
•Team needs standard practices
•Contributing to community

The Iron Law

NO SKILL WITHOUT A FAILING TEST FIRST.

If you can't demonstrate the problem the skill solves, you don't need the skill yet.

Why Write Skills?

Benefits: ✅ Codifies tacit knowledge ✅ Creates consistency ✅ Accelerates learning ✅ Reduces mistakes ✅ Enables scaling ✅ Builds institutional memory

Without documented skills: ❌ Knowledge stays in heads ❌ Inconsistent approaches ❌ Repeat mistakes ❌ Slow onboarding ❌ Can't scale

TDD for Skill Writing

Step 1: Baseline (RED Phase)

code

🔴 RED: Establish baseline without skill

Observation: Debugging takes too long, focuses on symptoms

Evidence gathering:
1. Observe current behavior (without skill)
2. Document specific failures
3. Collect multiple examples
4. Identify patterns

Example baseline observation:
---
Scenario: Debug slow API endpoint

Current approach (no skill):
1. User adds caching ← Treats symptom
2. Performance improves temporarily
3. Real issue (N+1 queries) remains
4. Problem returns as data grows
5. Multiple "fixes" accumulate
6. Root cause never found

Failures observed:
- ❌ Fixed symptom, not cause
- ❌ No systematic investigation
- ❌ Didn't trace to origin
- ❌ Quick fix instead of proper solution
- ❌ Problem will recur

Baseline documented ✅
Need: Systematic debugging process
---

This baseline PROVES we need the skill.

Baseline requirements:

•Real scenarios (not hypothetical)
•Specific failures (not vague problems)
•Multiple examples (not one-offs)
•Clear pattern (not random)

Step 2: Document Failures (Still RED)

code

Failure catalog from baseline:

Failure 1: Stopping at symptoms
Example: Added caching instead of fixing N+1 queries
Impact: Problem returns later
Frequency: 8/10 debugging sessions

Failure 2: No systematic process
Example: Random changes until something works
Impact: Wastes time, creates new bugs
Frequency: 7/10 sessions

Failure 3: Not tracing to origin
Example: Found slow query but not why it's slow
Impact: Incomplete understanding
Frequency: 9/10 sessions

Failure 4: Quick fixes under pressure
Example: "Just make it work for demo"
Impact: Technical debt
Frequency: 5/10 sessions

Failures documented ✅
These define what skill must address

Step 3: Write Skill (GREEN Phase)

code

🟢 GREEN: Write skill to address failures

Skill name: systematic-debugging

Must address documented failures:
- Failure 1 → Need process that finds root causes
- Failure 2 → Need step-by-step methodology
- Failure 3 → Need techniques for tracing origins
- Failure 4 → Need guidance for handling pressure

Skill structure:

1. Core Principle
   "Debug methodically, not randomly."
   ← Addresses Failure 2

2. When to Use This Skill
   - Encountering bugs
   - Unexpected behavior
   - Tests failing
   ← Clear trigger conditions

3. The Iron Law
   "NEVER try random things until it works."
   ← Directly counters Failure 2

4. The 5-Step Process
   1. Reproduce
   2. Isolate
   3. Identify (root cause)
   4. Fix
   5. Verify
   ← Addresses Failures 1, 2, 3

5. Examples
   - Performance degradation
   - Data corruption
   - Intermittent failures
   ← Real scenarios from baseline

6. Handling Pressure
   "Time pressure is when you MOST need systematic approach"
   ← Addresses Failure 4

7. Common Mistakes
   - Trying random things
   - Fixing symptoms
   - Not verifying
   ← Directly from failure catalog

8. Integration with Other Skills
   - Use with TDD
   - Use with code-review
   ← Shows how it fits

9. Authority
   - Scientific method
   - Professional practices
   - Research evidence
   ← Builds credibility

10. Your Commitment
    Checklist of commitments
    ← Creates accountability

11. Bottom Line
    One-sentence summary
    ← Memorable takeaway

Skill written ✅

Step 4: Test the Skill (GREEN Validation)

code

Test skill with baseline scenarios:

Scenario 1: Slow API endpoint (same as baseline)
- Give fresh subagent task + skill
- Observe behavior

Result:
✅ Follows 5-step process
✅ Reproduces issue
✅ Isolates to N+1 queries
✅ Identifies root cause
✅ Makes proper fix
✅ Verifies with tests

Improvement: Fixed root cause, not symptom ✅

Scenario 2: Intermittent test failure
- Give fresh subagent task + skill
- Observe behavior

Result:
✅ Reproduces reliably
✅ Isolates to timing issue
✅ Identifies race condition
✅ Fixes architecture
✅ Verifies fix

Improvement: Found and fixed root cause ✅

Skill solves documented failures ✅

Step 5: Iterate and Refine (REFACTOR)

code

🔵 REFACTOR: Test edge cases, find loopholes

Edge case 1: Time pressure
Test: Give skill + "Fix this in 10 minutes"
Result: ❌ Subagent skips process, makes quick fix

LOOPHOLE! Skill doesn't hold under time pressure.

Fix: Strengthen time pressure guidance:
"Time pressure is when you MOST need systematic approach.
Quick fixes create 10x more work later."

Updated skill ✅

Retest: ✅ Now handles time pressure

Edge case 2: Exhaustion
Test: Give skill + "You've been debugging for 6 hours"
Result: ❌ Subagent shortcuts process

LOOPHOLE! Skill doesn't address exhaustion.

Fix: Add exhaustion guidance:
"Exhaustion impairs judgment. The process protects you
when judgment fails. Follow it systematically."

Updated skill ✅

Retest: ✅ Now handles exhaustion

Edge case 3: "Works on my machine"
Test: Give skill + environment-specific bug
Result: ✅ Process finds environment differences

No loophole ✅

All edge cases tested ✅
Loopholes closed ✅
Skill refined ✅

The Four Types of Skills

Type 1: Discipline

What it is: Commitment to a practice

Structure:

•Iron Law (the discipline)
•Why it matters
•How to maintain it
•What to avoid

Examples:

•test-driven-development (Test first, code second)
•database-backup (Always backup before tests)
•code-review (Never merge without review)

Writing guide:

code

Focus on:
- The commitment required
- Why shortcuts fail
- How to stay disciplined
- Handling pressure to skip it

Type 2: Technique

What it is: Step-by-step process

Structure:

•When to use
•Step-by-step instructions
•Examples for each step
•Common mistakes

Examples:

•systematic-debugging (5-step process)
•root-cause-tracing (Backward tracing)
•git-bisect (Binary search for bugs)

Writing guide:

code

Focus on:
- Clear steps
- Decision points
- What to do at each step
- How to know you're done

Type 3: Pattern

What it is: Reusable solution

Structure:

•The problem
•The solution
•When to use it
•Variations

Examples:

•condition-based-waiting (Wait for condition, not timeout)
•dependency-injection (Pass dependencies in)
•repository-pattern (Abstract data access)

Writing guide:

code

Focus on:
- Problem it solves
- How the pattern works
- Implementation examples
- When NOT to use it

Type 4: Reference

What it is: Conceptual knowledge

Structure:

•Core concepts
•Principles
•Guidelines
•Best practices

Examples:

•testing-anti-patterns (What NOT to do)
•anthropic-best-practices (Prompt engineering)
•persuasion-principles (Influence techniques)

Writing guide:

code

Focus on:
- Key concepts
- Principles to follow
- Examples of each
- Common misunderstandings

Skill Document Structure

Required Sections

markdown

---
name: skill-name
description: "One-line description: when to use and what it does"
---

# Skill Name

## Core Principle
One-sentence essence of the skill

## When to Use This Skill
- Bullet list of trigger conditions
- Specific scenarios
- Clear indicators

## The Iron Law
**BOLD STATEMENT OF CORE RULE**

Clear, unambiguous, memorable

## Why [Skill Name]?
**Benefits:**
✅ List of benefits

**Without [skill]:**
❌ List of problems

## [Main Content]
Detailed explanation, process, or technique
- Step-by-step if technique
- Principles if reference
- Examples throughout

## Examples
Real-world examples showing:
- Before (without skill)
- After (with skill)
- Improvement demonstrated

## Common Mistakes
What NOT to do
Why mistakes happen
How to avoid them

## Integration with Skills
Which skills this works with
When to combine them
How they complement each other

## Authority
Based on:
- Research
- Industry standards
- Expert recommendations
- Proven practices

## Your Commitment
Checklist format:
- [ ] I will...
- [ ] I will...

## Bottom Line
---
**Bottom Line**: One sentence summary that captures essence

Optional Sections

markdown

## Red Flags
Warning signs this skill is needed

## Checklist
Step-by-step checklist for applying skill

## Advanced Techniques
Beyond basics, for experienced users

## Troubleshooting
Common issues and solutions

## Real-World Case Studies
Detailed examples from actual usage

Writing Style Guidelines

Guideline 1: Be Specific

code

❌ BAD (vague):
"Write good tests"

✅ GOOD (specific):
"Write tests that verify behavior, not implementation.
Test through public API only."

Guideline 2: Use Examples

code

❌ BAD (abstract):
"Follow the debugging process"

✅ GOOD (concrete):
"Example debugging session:
1. Bug: Login fails
2. Reproduce: Happens every time
3. Isolate: Error in AuthController
4. Identify: Missing database table
5. Fix: Run migrations
6. Verify: Login now works"

Guideline 3: Show Before/After

code

❌ BAD (only after):
"Here's how to do it right..."

✅ GOOD (both):
"Before: Random debugging, symptoms fixed
After: Systematic process, root causes found"

Guideline 4: Address Resistance

code

❌ BAD (ignore objections):
"Just follow the process"

✅ GOOD (handle objections):
"Common objection: 'No time for TDD'
Reality: TDD is faster because bugs caught immediately,
not after deployment. Follow the process."

Guideline 5: Make It Scannable

code

❌ BAD (wall of text):
Long paragraphs with no structure, hard to scan,
can't find key points quickly...

✅ GOOD (structured):
- Use bullets
- Use headings
- Use code blocks
- Use bold for key points
- Use checkboxes for processes

Skill Testing Protocol

Test 1: Baseline Without Skill

code

1. Create 3-5 realistic scenarios
2. Fresh subagent for each
3. NO skill provided
4. Observe and document failures
5. Catalog specific problems

Pass criteria: Clear failures documented

Test 2: Improvement With Skill

code

1. Same scenarios as baseline
2. Fresh subagent for each
3. Include skill
4. Observe and document improvements
5. Verify failures addressed

Pass criteria: Measurable improvement in all scenarios

Test 3: Pressure Testing

code

1. Test with time constraints
2. Test with sunk cost
3. Test with exhaustion
4. Test with authority pressure

Pass criteria: Skill holds up under all pressures

Test 4: Integration Testing

code

1. Test skill with related skills
2. Verify compatibility
3. Check for conflicts
4. Document integrations

Pass criteria: Works well with other skills

Skill Maintenance

When to Update

code

Update when:
- Loophole discovered
- Better technique found
- Feedback received
- Edge case encountered
- Integration issue found

Update process:
1. Document the issue
2. Update skill content
3. Re-test all scenarios
4. Verify improvement
5. Update version/date

Versioning

code

Track changes:
- Major update: Core process changed
- Minor update: Examples added, refinements
- Patch: Typos, clarifications

Document:
- What changed
- Why changed
- Test results

Common Skill Writing Mistakes

Mistake 1: No Baseline

code

❌ BAD:
Write skill based on theory
No proof of need
Assume problem exists

Result: Skill solves non-existent problem

✅ GOOD:
Establish baseline first
Document actual failures
Prove skill is needed
Create from reality

Mistake 2: Too Abstract

code

❌ BAD:
"Follow best practices"
No specifics
No examples
Just platitudes

Result: No one knows how to apply it

✅ GOOD:
Specific steps
Concrete examples
Clear instructions
Actionable guidance

Mistake 3: Not Tested

code

❌ BAD:
Write skill
Ship it
Hope it works

Result: Loopholes in production

✅ GOOD:
Write skill
Test thoroughly
Find loopholes
Fix them
Then ship

Mistake 4: Missing Authority

code

❌ BAD:
"Trust me, this works"
No sources
No evidence
Just opinion

Result: No credibility

✅ GOOD:
Based on research
Industry standards
Expert recommendations
Proven practices

Example: Complete Skill Writing Process

code

🔴 RED PHASE

Problem observation:
Tests failing when run together, passing alone

Baseline testing (no skill):
Scenario 1: Test pollution
- Observe: Developers add retries
- Result: Masks problem
- Failure: Pollution not fixed

Scenario 2: Test isolation
- Observe: Developers can't find polluter
- Result: Comment out tests
- Failure: Lost test coverage

Scenario 3: Database state
- Observe: Manual cleanup added
- Result: Brittle, incomplete
- Failure: Still has pollution

Failures documented:
- Can't find polluter
- Mask instead of fix
- Manual cleanup instead of isolation

Baseline proves need for skill ✅

---

🟢 GREEN PHASE

Write skill: test-pollution-detection

Sections:
1. Core Principle: "Tests must be independent"

2. Iron Law: "Find and fix polluter, don't mask"

3. Detection process:
   - Binary search for polluter
   - Isolate problematic test
   - Fix root cause

4. Examples from baseline:
   - Database state pollution
   - Global variable pollution
   - File system pollution

5. Tools:
   - find-polluter.sh script
   - Test isolation patterns
   - Proper cleanup

6. Common mistakes (from baseline):
   - Adding retries (masking)
   - Manual cleanup (brittle)
   - Commenting out tests (lost coverage)

Skill written ✅

---

🔵 REFACTOR PHASE

Test with skill:
Scenario 1: ✅ Uses binary search, finds polluter
Scenario 2: ✅ Fixes root cause, not symptoms
Scenario 3: ✅ Implements proper isolation

Improvement verified ✅

Pressure testing:
Time pressure: ❌ Skips proper fix, adds retry

LOOPHOLE! Add to skill:
"Masking pollution takes 10x longer to fix later.
Find and fix polluter, even under pressure."

Updated ✅

Retest: ✅ All scenarios pass

Integration test:
- Works with TDD ✅
- Works with systematic-debugging ✅
- Works with CI/CD ✅

Skill validated ✅
Ready to use ✅

Integration with Skills

Required:

•testing-skills-with-subagents - Test every skill

Use with:

•test-driven-development - Apply TDD to skills
•code-review - Review skills like code
•sharing-skills - Share validated skills

Enables:

•Consistent practices
•Knowledge transfer
•Quality processes

Skill Writing Checklist

Before writing:

• Problem clearly identified
• Baseline established (RED)
• Failures documented
• Multiple examples collected

While writing:

After writing:

Authority

This skill is based on:

•Technical writing best practices
•Test-Driven Development (Kent Beck)
•Process documentation standards
•Knowledge management research
•Anthropic best practices for prompt engineering

Research: Studies show documented processes are followed 5x more consistently than undocumented.

Social Proof: All professional organizations maintain process documentation.

Your Commitment

When writing skills:

• I will establish baseline first (RED)
• I will document specific failures
• I will write skill to address failures (GREEN)
• I will test thoroughly (REFACTOR)
• I will find and close loopholes
• I will include real examples
• I will cite authority
• I will only share tested skills

Bottom Line: Skills are code for humans. Write them test-first: baseline (RED) → write skill (GREEN) → test and refine (REFACTOR). No skill without a failing test first.