Table of Contents

• Overview
• Core Concept
• Key Benefits
• The Iron Law
• Skill Types
• 1. Technique Skills
• 2. Pattern Skills
• 3. Reference Skills
• Quick Start
• Minimal Viable Skill Creation
• File Structure Requirements
• Description Optimization
• Formula
• Requirements
• Example Patterns
• The TDD Cycle for Skills
• RED Phase: Document Baseline Failures
• GREEN Phase: Minimal Skill Implementation
• REFACTOR Phase: Bulletproof Against Rationalization
• Anti-Rationalization Techniques
• Common Rationalization Patterns
• Red Flags for Self-Checking
• Red Flags That You're Rationalizing
• Explicit Exception Handling
• When NOT to Use This Skill
• Module References
• Deployment Checklist
• Quality Gates
• Validation Command
• Common Pitfalls
• 1. Writing Without Testing
• 2. Vague Descriptions
• 3. Monolithic Skills
• 4. Missing Anti-Rationalization
• 5. Theoretical Examples
• Integration with Other Skills
• With modular-skills
• With skills-eval
• Workflow
• Summary

Skill Authoring Guide

Overview

This skill teaches how to write effective Claude Code skills using Test-Driven Development (TDD) methodology, persuasion principles from compliance research, and official Anthropic best practices. It treats skill writing as process documentation requiring empirical validation rather than theoretical instruction writing.

Core Concept

Skills are not essays or documentation—they are behavioral interventions designed to change Claude's behavior in specific, measurable ways. Like software, they must be tested against real failure modes before deployment.

Key Benefits

• Empirical Validation: TDD validates skills address real failure modes, not imagined ones
• Research-Backed Persuasion: Compliance techniques proven to double adherence rates
• Discoverability: Optimized descriptions for Claude's skill selection process
• Progressive Disclosure: Efficient token usage through modular file structures
• Anti-Rationalization: Bulletproofing against Claude's tendency to explain away requirements

The Iron Law

NO SKILL WITHOUT A FAILING TEST FIRST

Every skill must begin with documented evidence of Claude failing without it. This validates you're solving real problems, not building imaginary solutions.

Skill Types

1. Technique Skills

Purpose: Teach specific methods or approaches

Examples: TDD workflow, commit message style, API design patterns

Structure: Step-by-step procedures with decision points

2. Pattern Skills

Purpose: Document recurring solutions to common problems

Examples: Error handling patterns, module organization, testing strategies

Structure: Problem-solution pairs with variations

3. Reference Skills

Purpose: Provide quick lookup information

Examples: Command reference, configuration options, best practice checklists

Structure: Tables, lists, and indexed information

Quick Start

Skill Analysis

```bash

Analyze skill complexity

python scripts/analyze.py

Estimate tokens

python scripts/tokens.py ```

Validation

```bash

Validate skill structure

python scripts/abstract_validator.py --check ```

Verification: Run analysis and review token estimates before proceeding.

Description Optimization

Skill descriptions are critical for Claude's discovery process. They must be optimized for both semantic search and explicit triggering.

Formula

[What it does] + [When to use it] + [Key triggers]

Verification: Run the command with --help flag to verify availability.

Requirements

Always:

• Third person voice ("Teaches...", "Provides...", "Guides...")
• Include "Use when..." clause
• Specific, concrete language
• Key discovery terms

Never:

• First person ("I teach...", "We provide...")
• Vague descriptions ("Helps with coding")
• Marketing language
• Missing use case context

Example Patterns

Good:

description: Guides API design using RESTful principles and best practices. Use when designing new APIs, reviewing API proposals, or standardizing endpoint patterns. Covers resource modeling, HTTP method selection, and versioning strategies.

Verification: Run the command with --help flag to verify availability.

Bad:

description: This skill helps you design better APIs.

Verification: Run the command with --help flag to verify availability.

The TDD Cycle for Skills

RED Phase: Document Baseline Failures

Goal: Establish empirical evidence that intervention is needed

Process:

1. Create 3+ pressure scenarios combining:

   • Time pressure ("quickly", "simple task")
   • Ambiguity ("standard approach", "best practices")
   • Multiple requirements
   • Edge cases

2. Run scenarios in fresh Claude instances WITHOUT skill

3. Document failures verbatim:

   ## Baseline Scenario 1: Simple API endpoint
   
   **Prompt**: "Quickly add a user registration endpoint"
   
   **Claude Response** (actual, unedited):
   [paste exact response]
   
   **Failures Observed**:
   - Skipped error handling
   - No input validation
   - Missing rate limiting
   - Didn't consider security
   

   Verification: Run the command with --help flag to verify availability.

4. Identify patterns across failures

GREEN Phase: Minimal Skill Implementation

Goal: Create smallest intervention that addresses documented failures

Process:

1. Write SKILL.md with required frontmatter:

   ---
   name: skill-name
   description: [optimized description]
   version: 1.0.0
   category: [appropriate category]
   tags: [discovery, keywords, here]
   dependencies: []
   estimated_tokens: [realistic estimate]
   
   # Claude Code 2.1.0+ Optional Fields:
   context: fork              # Run in isolated sub-agent context
   agent: agent-name          # Specify agent type for execution
   user-invocable: false      # Hide from slash command menu (default: true)
   model: sonnet              # Model override for this skill
   
   # YAML-style allowed-tools (cleaner syntax)
   allowed-tools:
     - Read
     - Grep
     - Bash(npm *)            # Wildcard patterns supported
   
   # Lifecycle hooks scoped to skill
   hooks:
     PreToolUse:
       - matcher: "Bash"
         command: "./validate.sh"
         once: true           # Run only once per session
     Stop:
       - command: "./cleanup.sh"
   ---
   

   Verification: Run the command with --help flag to verify availability.

2. Add content that directly counters baseline failures

3. Include ONE example showing correct behavior

4. Test with skill present:

   • Run same pressure scenarios
   • Document new behavior
   • Verify improvement over baseline

REFACTOR Phase: Bulletproof Against Rationalization

Goal: Eliminate Claude's ability to explain away requirements

Process:

1. Run new pressure scenarios with skill active

2. Document rationalizations:

   **Scenario**: Add authentication to API
   
   **Claude's Rationalization**:
   "Since this is a simple internal API, basic authentication
   is sufficient for now. We can add OAuth later if needed."
   
   **What Should Happen**:
   Security requirements apply regardless of API scope.
   Internal APIs need proper authentication.
   

   Verification: Run the command with --help flag to verify availability.

3. Add explicit counters:

   • Exception tables with "No Exceptions" rows
   • Red flag lists
   • Decision flowcharts with escape hatches blocked
   • Commitment statements

4. Iterate until rationalizations stop

Anti-Rationalization Techniques

Claude is sophisticated at finding ways to bypass requirements while appearing compliant. Skills must explicitly counter common rationalization patterns.

Common Rationalization Patterns

Excuse	Counter
"This is just a simple task"	Complexity doesn't exempt you from core practices. Use skills anyway.
"I remember the key points"	Skills evolve. Always run current version.
"Spirit vs letter of the law"	Foundational principles come first. No shortcuts.
"User just wants quick answer"	Quality and speed aren't exclusive. Both matter.
"Context is different here"	Skills include context handling. Follow the process.
"I'll add it in next iteration"	Skills apply to current work. No deferral.

Red Flags for Self-Checking

Skills should include explicit red flag lists:

## Red Flags That You're Rationalizing

Stop immediately if you think:
- "This is too simple for the full process"
- "I already know this, no need to review"
- "The user wouldn't want me to do all this"
- "We can skip this step just this once"
- "The principle doesn't apply here because..."

Verification: Run the command with --help flag to verify availability.

Explicit Exception Handling

When exceptions truly exist, document them explicitly:

## When NOT to Use This Skill

 **Don't use when:**
- User explicitly requests prototype/draft quality
- Exploring multiple approaches quickly (note for follow-up)
- Working in non-production environment (document shortcut)

 **Still use for:**
- "Quick" production changes
- "Simple" fixes to production code
- Internal tools and scripts

Verification: Run the command with --help flag to verify availability.

Module References

For detailed implementation guidance:

• TDD Methodology: See modules/tdd-methodology.md for RED-GREEN-REFACTOR cycle details
• Persuasion Principles: See modules/persuasion-principles.md for compliance research and techniques
• Description Writing: See modules/description-writing.md for discovery optimization
• Progressive Disclosure: See modules/progressive-disclosure.md for file structure patterns
• Anti-Rationalization: See modules/anti-rationalization.md for bulletproofing techniques
• Graphviz Conventions: See modules/graphviz-conventions.md for process diagram standards
• Testing with Subagents: See modules/testing-with-subagents.md for pressure testing methodology
• Deployment Checklist: See modules/deployment-checklist.md for final validation

Deployment Checklist

Before deploying a new skill:

Quality Gates

• RED Phase Complete: 3+ baseline scenarios documented with actual failures
• GREEN Phase Complete: Skill tested and shows measurable improvement
• REFACTOR Phase Complete: Rationalizations identified and countered
• Frontmatter Valid: All required YAML fields present and correct
• Description Optimized: Third person, includes "Use when", has key terms
• Line Count: SKILL.md under 500 lines (move extras to modules)
• Module References: All referenced files exist and are linked correctly
• Examples Present: At least one concrete example included
• Scripts Executable: Any tools tested and working
• No Orphans: No dead links or missing dependencies

Validation Command

python scripts/skill_validator.py

Verification: Run python --version to verify Python environment.

Exit codes:

• 0 = Success, ready to deploy
• 1 = Warnings, should fix but can deploy
• 2 = Errors, must fix before deploying

Common Pitfalls

1. Writing Without Testing

Problem: Creating skills based on what "should" work

Solution: Always start with documented failures (RED phase)

2. Vague Descriptions

Problem: "Helps with testing" - not discoverable or actionable

Solution: "Guides TDD workflow with RED-GREEN-REFACTOR cycle. Use when writing new tests, refactoring existing code, or ensuring test coverage."

3. Monolithic Skills

Problem: Everything in SKILL.md, 1000+ lines

Solution: Keep main file under 500 lines, use progressive disclosure with modules

4. Missing Anti-Rationalization

Problem: Claude finds creative ways to bypass requirements

Solution: Add explicit exception tables, red flags, and commitment statements

5. Theoretical Examples

Problem: Examples show ideal scenarios, not real challenges

Solution: Use actual failure cases from RED phase as examples

Integration with Other Skills

With modular-skills

• Use skill-authoring for creating individual skills
• Use modular-skills for architecting skill structure

With skills-eval

• Use skill-authoring for initial creation and testing
• Use skills-eval for ongoing quality assessment

Workflow

1. Create new skill using skill-authoring (TDD approach)
2. Validate structure using modular-skills (architecture check)
3. Assess quality using skills-eval (compliance and metrics)
4. Iterate based on feedback

Summary

Effective skill authoring requires:

1. Empirical Testing: Start with real failures (RED)
2. Minimal Intervention: Solve actual problems (GREEN)
3. Bulletproofing: Counter rationalizations (REFACTOR)
4. Optimized Discovery: Write descriptions for activation
5. Progressive Disclosure: Structure for token efficiency
6. Persuasive Design: Apply compliance research
7. Validation: Test before deploying

Remember: Skills are behavioral interventions, not documentation. If you haven't tested it against real failure modes, you haven't validated that it works.

Troubleshooting

Common Issues

Skill not loading Check YAML frontmatter syntax and required fields

Token limits exceeded Use progressive disclosure - move details to modules

Modules not found Verify module paths in SKILL.md are correct