Creating Skills

Overview

Skills are reference guides for proven techniques, patterns, or tools. Write them to help future Claude instances quickly find and apply effective approaches.

Skills must be discoverable (Claude can find them), scannable (quick to evaluate), and actionable (clear examples).

Core principle: Default assumption is Claude is already very smart. Only add context Claude doesn't already have.

When to Use

Create a skill when:

• Technique wasn't intuitively obvious
• Pattern applies broadly across projects
• You'd reference this again
• Others would benefit

Don't create for:

• One-off solutions specific to single project
• Standard practices well-documented elsewhere
• Project conventions (put those in .claude/CLAUDE.md)

Required Structure

Frontmatter (YAML)

yaml
1---
2name: skill-name-with-hyphens
3description: Use when [triggers/symptoms] - [what it does and how it helps]
4tags: relevant-tags
5---

Rules:

• Only name and description fields supported (max 1024 chars total)
• Name: letters, numbers, hyphens only (max 64 chars). Use gerund form (verb + -ing)
• Avoid reserved words: "anthropic", "claude" in names
• Description: Third person, starts with "Use when..." (max 1024 chars)
• Include BOTH triggering conditions AND what skill does
• Match specificity to task complexity (degrees of freedom)

Document Structure

markdown
1# Skill Name
2
3## Overview
4Core principle in 1-2 sentences. What is this?
5
6## When to Use
7- Bullet list with symptoms and use cases
8- When NOT to use
9
10## Quick Reference
11Table or bullets for common operations
12
13## Implementation
14Inline code for simple patterns
15Link to separate file for heavy reference (100+ lines)
16
17## Common Mistakes
18What goes wrong + how to fix
19
20## Real-World Impact (optional)
21Concrete results from using this technique

Degrees of Freedom

Match specificity to task complexity:

• High freedom: Flexible tasks requiring judgment

  • Use broad guidance, principles, examples
  • Let Claude adapt approach to context
  • Example: "Use when designing APIs - provides REST principles and patterns"

• Low freedom: Fragile or critical operations

  • Be explicit about exact steps
  • Include validation checks
  • Example: "Use when deploying to production - follow exact deployment checklist with rollback procedures"

Red flag: If skill tries to constrain Claude too much on creative tasks, reduce specificity. If skill is too vague on critical operations, add explicit steps.

Claude Search Optimization (CSO)

Critical: Future Claude reads the description to decide if skill is relevant. Optimize for discovery.

Description Best Practices

yaml
1# ❌ BAD - Too vague, doesn't mention when to use
2description: For async testing
3
4# ❌ BAD - First person (injected into system prompt)
5description: I help you with flaky tests
6
7# ✅ GOOD - Triggers + what it does
8description: Use when tests have race conditions or pass/fail inconsistently - replaces arbitrary timeouts with condition polling for reliable async tests
9
10# ✅ GOOD - Technology-specific with explicit trigger
11description: Use when using React Router and handling auth redirects - provides patterns for protected routes and auth state management

Keyword Coverage

Use words Claude would search for:

• Error messages: "ENOENT", "Cannot read property", "Timeout"
• Symptoms: "flaky", "hanging", "race condition", "memory leak"
• Synonyms: "cleanup/teardown/afterEach", "timeout/hang/freeze"
• Tools: Actual command names, library names, file types

Naming Conventions

Use gerund form (verb + -ing):

• ✅ creating-skills not skill-creation
• ✅ testing-with-subagents not subagent-testing
• ✅ debugging-memory-leaks not memory-leak-debugging
• ✅ processing-pdfs not pdf-processor
• ✅ analyzing-spreadsheets not spreadsheet-analysis

Why gerunds work:

• Describes the action you're taking
• Active and clear
• Consistent with Anthropic conventions

Avoid:

• ❌ Vague names like "Helper" or "Utils"
• ❌ Passive voice constructions

Code Examples

One excellent example beats many mediocre ones.

Choose Language by Use Case

• Testing techniques → TypeScript/JavaScript
• System debugging → Shell/Python
• Data processing → Python
• API calls → TypeScript/JavaScript

Good Example Checklist

• Complete and runnable
• Well-commented explaining WHY not just what
• From real scenario (not contrived)
• Shows pattern clearly
• Ready to adapt (not generic template)
• Shows both BAD (❌) and GOOD (✅) approaches
• Includes realistic context/setup code

Example Template

typescript
1// ✅ GOOD - Clear, complete, ready to adapt
2interface RetryOptions {
3  maxAttempts: number;
4  delayMs: number;
5  backoff?: 'linear' | 'exponential';
6}
7
8async function retryOperation<T>(
9  operation: () => Promise<T>,
10  options: RetryOptions
11): Promise<T> {
12  const { maxAttempts, delayMs, backoff = 'linear' } = options;
13
14  for (let attempt = 1; attempt <= maxAttempts; attempt++) {
15    try {
16      return await operation();
17    } catch (error) {
18      if (attempt === maxAttempts) throw error;
19
20      const delay = backoff === 'exponential'
21        ? delayMs * Math.pow(2, attempt - 1)
22        : delayMs * attempt;
23
24      await new Promise(resolve => setTimeout(resolve, delay));
25    }
26  }
27
28  throw new Error('Unreachable');
29}
30
31// Usage
32const data = await retryOperation(
33  () => fetchUserData(userId),
34  { maxAttempts: 3, delayMs: 1000, backoff: 'exponential' }
35);

Don't

• ❌ Implement in 5+ languages (you're good at porting)
• ❌ Create fill-in-the-blank templates
• ❌ Write contrived examples
• ❌ Show only code without comments

File Organization

Self-Contained (Preferred)

typescript-type-safety/
  SKILL.md    # Everything inline

When: All content fits in ~500 words, no heavy reference needed

With Supporting Files

api-integration/
  SKILL.md           # Overview + patterns
  retry-helpers.ts   # Reusable code
  examples/
    auth-example.ts
    pagination-example.ts

When: Reusable tools or multiple complete examples needed

With Heavy Reference

aws-sdk/
  SKILL.md       # Overview + workflows
  s3-api.md      # 600 lines API reference
  lambda-api.md  # 500 lines API reference

When: Reference material > 100 lines

Token Efficiency

Skills load into every conversation. Keep them concise.

Target Limits

• SKILL.md: Keep under 500 lines
• Getting-started workflows: <150 words
• Frequently-loaded skills: <200 words total
• Other skills: <500 words
• Files > 100 lines: Include table of contents

Challenge each piece of information: "Does Claude really need this explanation?"

Compression Techniques

markdown
1# ❌ BAD - Verbose (42 words)
2Your human partner asks: "How did we handle authentication errors in React Router before?"
3You should respond: "I'll search past conversations for React Router authentication patterns."
4Then dispatch a subagent with the search query: "React Router authentication error handling 401"
5
6# ✅ GOOD - Concise (20 words)
7Partner: "How did we handle auth errors in React Router?"
8You: Searching...
9[Dispatch subagent → synthesis]

Techniques:

• Reference tool --help instead of documenting all flags
• Cross-reference other skills instead of repeating content
• Show minimal example of pattern
• Eliminate redundancy
• Use progressive disclosure (reference additional files as needed)
• Organize content by domain for focused context

Workflow Recommendations

For multi-step processes, include:

1. Clear sequential steps: Break complex tasks into numbered operations
2. Feedback loops: Build in verification/validation steps
3. Error handling: What to check when things go wrong
4. Checklists: For processes with many steps or easy-to-miss details

Example structure:

markdown
1## Workflow
2
31. **Preparation**
4   - Check prerequisites
5   - Validate environment
6
72. **Execution**
8   - Step 1: [action + expected result]
9   - Step 2: [action + expected result]
10
113. **Verification**
12   - [ ] Check 1 passes
13   - [ ] Check 2 passes
14
154. **Rollback** (if needed)
16   - Steps to undo changes

Common Mistakes

Flowchart Usage

Only use flowcharts for:

• Non-obvious decision points
• Process loops where you might stop too early
• "When to use A vs B" decisions

Never use for:

• Reference material → Use tables/lists
• Code examples → Use markdown blocks
• Linear instructions → Use numbered lists

Cross-Referencing Skills

markdown
1# ✅ GOOD - Name only with clear requirement
2**REQUIRED:** Use superpowers:test-driven-development before proceeding
3
4**RECOMMENDED:** See typescript-type-safety for proper type guards
5
6# ❌ BAD - Unclear if required
7See skills/testing/test-driven-development
8
9# ❌ BAD - Force-loads file, wastes context
10@skills/testing/test-driven-development/SKILL.md

Advanced Practices

Iterative Development

Best approach: Develop skills iteratively with Claude

1. Start with minimal viable skill
2. Test with real use cases
3. Refine based on what works
4. Remove what doesn't add value

Build Evaluations First

Before extensive documentation:

1. Create test scenarios
2. Identify what good looks like
3. Document proven patterns
4. Skip theoretical improvements

Utility Scripts

For reliability, provide:

• Scripts with explicit error handling (don't defer errors to Claude)
• Exit codes for success/failure
• Clear error messages
• Examples of usage
• List required dependencies explicitly

Example:

bash
1#!/bin/bash
2set -e  # Exit on error
3
4if [ ! -f "config.json" ]; then
5  echo "Error: config.json not found" >&2
6  exit 1
7fi
8
9# Script logic here
10echo "Success"
11exit 0

Verifiable Intermediate Outputs

For complex operations, create validation checkpoints:

1. Have Claude produce a structured plan file
2. Validate the plan with a script
3. Execute only after validation passes

This catches errors before they compound.

Templates for Structured Output

When skills produce consistent formats:

markdown
1## Output Template
2
3\`\`\`typescript
4interface ExpectedOutput {
5  status: 'success' | 'error';
6  data: YourDataType;
7  errors?: string[];
8}
9\`\`\`
10
11**Usage**: Copy and adapt for your context

Skill Creation Checklist

Before writing:

• Technique isn't obvious or well-documented elsewhere
• Pattern applies broadly (not project-specific)
• I would reference this across multiple projects

Frontmatter:

• Name uses only letters, numbers, hyphens
• Description starts with "Use when..."
• Description includes triggers AND what skill does
• Description is third person
• Total frontmatter < 1024 characters

Content:

• Overview states core principle (1-2 sentences)
• "When to Use" section with symptoms
• Quick reference table for common operations
• One excellent code example (if technique skill)
• Common mistakes section
• Keywords throughout for searchability

Quality:

• Word count appropriate for frequency (see targets above)
• SKILL.md under 500 lines
• No narrative storytelling
• Flowcharts only for non-obvious decisions
• Supporting files only if needed (100+ lines reference)
• Cross-references use skill name, not file paths
• No time-sensitive information
• Consistent terminology throughout
• Concrete examples (not templates)
• Degrees of freedom match task complexity

Testing:

• Tested with Claude Haiku, Sonnet, and Opus (instructions effective for Opus may need more detail for Haiku)
• Tested with subagent scenarios (if discipline-enforcing skill)
• Addresses common rationalizations
• Includes red flags list

Directory Structure

skills/
  skill-name/
    SKILL.md              # Required
    supporting-file.*     # Optional
    examples/             # Optional
      example1.ts
    scripts/              # Optional
      helper.py

Flat namespace - all skills in one searchable directory

Real-World Impact

Good skills:

• Future Claude finds them quickly (CSO optimization)
• Can be scanned in seconds (quick reference)
• Provide clear actionable examples
• Prevent repeating same research
• Stay under 500 lines (token efficient)
• Match specificity to task needs (right degrees of freedom)

Bad skills:

• Get ignored (vague description)
• Take too long to evaluate (no quick reference)
• Leave gaps in understanding (no examples)
• Waste token budget (verbose explanations of obvious things)
• Over-constrain creative tasks or under-specify critical operations
• Include time-sensitive or obsolete information

Remember: Skills are for future Claude, not current you. Optimize for discovery, scanning, and action.

Golden rule: Default assumption is Claude is already very smart. Only add context Claude doesn't already have.