Writing Skills

Writing skills IS Test-Driven Development applied to process documentation.

Personal skills live in agent-specific directories (

~/.claude/skills

~/.codex/skills

You write test cases (pressure scenarios with subagents), watch them fail (baseline behavior), write the skill (documentation), watch tests pass (agents comply), and refactor (close loopholes).

Core principle: If you didn't watch an agent fail without the skill, you don't know if the skill teaches the right thing.

When to Create a Skill

Create when: Technique wasn't intuitively obvious, reusable across projects, pattern applies broadly, others would benefit.

Don't create for: One-off solutions, standard well-documented practices, project-specific conventions (use CLAUDE.md), mechanically enforceable constraints (automate instead).

Skill Types

• Technique: Concrete method with steps (condition-based-waiting)
• Pattern: Way of thinking about problems (flatten-with-flags)
• Reference: API docs, syntax guides, tool documentation

Directory Structure

skills/
  skill-name/
    SKILL.md              # Main reference (required)
    supporting-file.*     # Only if needed

Separate files for: heavy reference (100+ lines), reusable tools. Keep everything else inline.

SKILL.md Structure

Frontmatter: Only

name

description

CRITICAL: Description = triggering conditions ONLY. Never summarize the skill's workflow in description. Testing showed Claude follows description shortcuts instead of reading skill body.

# BAD: Summarizes workflow
description: Use when executing plans - dispatches subagent per task with code review between tasks

# GOOD: Just triggers
description: Use when executing implementation plans with independent tasks in the current session

Body structure:

# Skill Name
## Overview (1-2 sentences, core principle)
## When to Use (flowchart IF decision non-obvious, bullets with symptoms)
## Core Pattern (before/after code)
## Quick Reference (table/bullets)
## Common Mistakes (what goes wrong + fixes)

Claude Search Optimization (CSO)

• Use concrete triggers/symptoms in description, not language-specific symptoms
• Include keywords Claude would search: error messages, symptoms, tool names
• Name by what you DO:

  condition-based-waiting

  not

  async-test-helpers

• Gerunds work well:

  creating-skills

  ,

  debugging-with-logs

Token Efficiency:

• Getting-started workflows: <150 words
• Frequently-loaded skills: <200 words
• Other skills: <500 words
• Move details to

  --help

  , use cross-references, compress examples

Cross-References: Use

**REQUIRED SUB-SKILL:** Use superpowers:skill-name

@

Flowchart Usage

Use ONLY for: non-obvious decisions, process loops, "when to use A vs B". Never for: reference material, code examples, linear instructions.

Conventions and rendering: see

graphviz-conventions.dot

render-graphs.js

The Iron Law (Same as TDD)

NO SKILL WITHOUT A FAILING TEST FIRST

Applies to new skills AND edits. Write skill before testing? Delete it. Start over. No exceptions.

RED-GREEN-REFACTOR for Skills

RED: Run pressure scenario WITHOUT skill. Document exact failures and rationalizations.

GREEN: Write minimal skill addressing those specific failures. Re-test with skill.

REFACTOR: Agent found new rationalization? Add explicit counter. Re-test until bulletproof.

Testing Skills With Subagents

For detailed testing methodology, see

references/CLAUDE_MD_TESTING.md

Run scenarios without skill (RED), write skill addressing failures (GREEN), close loopholes (REFACTOR).

RED Phase: Baseline Testing

Run pressure scenario WITHOUT skill. Document exact failures.

Process:

• Create pressure scenarios (3+ combined pressures)
• Run WITHOUT skill
• Document choices and rationalizations verbatim
• Identify patterns

Pressure Types

Best tests combine 3+ pressures.

Key Elements

1. Concrete A/B/C choices (not open-ended)
2. Real constraints (specific times, consequences)
3. Real file paths
4. Make agent act ("What do you do?" not "What should you do?")

REFACTOR Phase: Close Loopholes

Capture new rationalizations verbatim. For each, add:

1. Explicit negation in rules
2. Rationalization table entry
3. Red flag entry
4. Updated description with violation symptoms

Re-test after each refactor. Continue until no new rationalizations.

Meta-Testing

After agent chooses wrong: "You read the skill and chose wrong. How could it be clearer?"

Three responses:

1. "Skill WAS clear, I chose to ignore" -> Need stronger foundational principle
2. "Skill should have said X" -> Add their suggestion
3. "I didn't see section Y" -> Make it more prominent

Persuasion Principles for Skill Design

LLMs respond to the same persuasion principles as humans. Meincke et al. (2025) tested 7 principles with N=28,000 AI conversations. Persuasion techniques doubled compliance rates (33% -> 72%, p < .001).

Effective Principles

• Authority: Imperative language ("YOU MUST", "Never", "Always"), non-negotiable framing. For discipline-enforcing skills.
• Commitment: Require announcements, force explicit choices. For ensuring skills are followed.
• Scarcity: Time-bound requirements ("Before proceeding"), sequential dependencies. For immediate verification.
• Social Proof: Universal patterns ("Every time", "Always"), failure modes. For documenting universal practices.
• Unity: Collaborative language ("our codebase", "we're colleagues"). For collaborative workflows.
• Reciprocity & Liking: Use sparingly or avoid.

Anthropic Best Practices

Skill Categories (from Anthropic's Guide)

Success Criteria Methodology

1. Define what "good output" looks like before writing the skill
2. Create 3+ concrete evaluation scenarios with expected outcomes
3. Test against scenarios, measure pass rate
4. Iterate until pass rate meets threshold (aim for >80%)

Core Principles

• Concise is key: Only add what Claude doesn't already know. Challenge each piece: "Does Claude need this?"
• Degrees of freedom: High (text instructions) for multiple valid approaches; Medium (pseudocode) for preferred patterns; Low (exact scripts) for fragile operations

Anti-Patterns

• Offering too many library options (provide a default with escape hatch)
• Assuming packages installed (list dependencies)
• Deeply nested references (keep one level deep)
• Time-sensitive information (use "old patterns" section)

Evaluation-Driven Development

1. Run Claude on tasks without Skill, document failures
2. Create 3+ evaluation scenarios
3. Establish baseline performance
4. Write minimal instructions to pass evaluations
5. Iterate: evaluate, compare baseline, refine

Quick Checklist (from Anthropic's Reference A)

• Frontmatter:

  name

  ,

  description

  (with trigger phrases),

  allowed-tools

• Description starts with "Use when..." and includes negative triggers
• Body is concise (under 500 words for most skills)
• At least one code example or concrete output
• No redundant information Claude already knows
• Cross-references use relative paths, not

  @

  links

Skill Creation Checklist

RED Phase:

• Create pressure scenarios (3+ combined pressures for discipline skills)
• Run WITHOUT skill - document baseline failures verbatim
• Identify rationalization patterns

GREEN Phase:

• YAML frontmatter: name (letters/numbers/hyphens), description (Use when..., third-person)
• Address specific baseline failures
• Keywords throughout for search
• One excellent code example (not multi-language)
• Run WITH skill - verify compliance

REFACTOR Phase:

• Add counters for new rationalizations
• Build rationalization table and red flags list
• Re-test until bulletproof

Deployment:

• Commit and push