Writing Skills

Overview

A skill is a reference guide for proven techniques, patterns, or tools. Skills help future Claude instances find and apply effective approaches.

Skills are: Reusable techniques, patterns, tools, reference guides
Skills are NOT: Narratives about how you solved a problem once, project-specific conventions (those go in CLAUDE.md)

SKILL.md Structure

Frontmatter (YAML):

• Required fields: name, description
• Max 1024 characters total
• name: Letters, numbers, hyphens only — no parentheses or special chars
• description: Third-person, describes ONLY when to use (NOT what it does) — see CSO section

---
name: skill-name
description: Use when [specific triggering conditions and symptoms]
user-invocable: true # only if user should invoke with /skill-name
---

# Skill Name

## Overview

Core principle in 1-2 sentences.

## When to Use

Bullet list with symptoms and use cases. When NOT to use.

## Core Pattern

Before/after comparison or workflow steps.

## Quick Reference

Table or bullets for scanning.

## Common Mistakes

What goes wrong + fixes.

Claude Search Optimization (CSO)

The description field is the discovery mechanism. Claude reads it to decide which skills to load.

Description = Triggering Conditions ONLY

NEVER summarize the skill's workflow in the description.

Why this matters: Testing revealed that when a description summarizes the workflow, Claude follows the description instead of reading the full skill body. A description saying "code review between tasks" caused Claude to do ONE review, even though the skill's flowchart showed TWO reviews. Changing the description to triggering conditions only fixed it.

The trap: Descriptions that summarize workflow create a shortcut Claude will take. The skill body becomes documentation Claude skips.

# BAD: Summarizes workflow — Claude may follow this instead of reading skill
description: Use when executing plans - dispatches subagent per task with code review between tasks

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

Content rules:

• Start with "Use when..."
• Include specific symptoms, situations, contexts
• Technology-agnostic unless skill is technology-specific
• Third person (injected into system prompt)
• Keep under 500 characters

Keyword Coverage

Include words Claude would search for:

• Error messages: "race condition", "ENOTEMPTY", "hook timed out"
• Symptoms: "flaky", "hanging", "zombie"
• Tools: actual command names, library names

Token Efficiency

Skills that load frequently must be concise.

Skill type	Target word count
Frequently-loaded workflows	< 200 words
Other skills	< 500 words

Techniques:

• Reference --help instead of documenting all flags inline
• Cross-reference other skills instead of repeating their content
• One excellent code example beats five mediocre ones

Skill Types

Type	Description	Examples
Technique	Concrete method with steps	root-cause-trace, cross-layer-check
Pattern	Way of thinking about problems	careful, tdd
Reference	API docs, tool documentation	playwright, cocoindex-code

Directory Structure

.claude/skills/
  skill-name/
    SKILL.md              # Main reference (required)
    supporting-file.*     # Only if needed for heavy reference or reusable tools

Keep inline: principles, concepts, code patterns < 50 lines.
Separate file for: heavy reference (100+ lines), reusable scripts/templates.

Quality Checklist

Before deploying any skill:

• name uses only letters, numbers, hyphens
• description starts with "Use when..." — no workflow summary
• Description is under 500 characters
• Keywords throughout for discoverability
• Clear overview with core principle
• No narrative storytelling ("in session 2025-10-03 we found...")
• One excellent example (not multi-language)
• Entry added to claude/README.md skills table (required by self-maintenance rule)

Common Mistakes

Mistake	Fix
Description summarizes workflow	Rewrite as triggering conditions only
Multiple language examples	Pick the most relevant language, one good example
Narrative storytelling	Convert to reusable technique/pattern
Project-specific rules in skill	Move to CLAUDE.md
Generic labels (step1, helper2)	Use semantic names
Flowchart for linear instructions	Use numbered list instead
No entry in README	Add to skills table immediately

When to Create a Skill

Create when:

• Technique wasn't intuitively obvious
• You'd reference this again across projects
• Pattern applies broadly (not project-specific)

Don't create for:

• One-off solutions
• Standard practices well-documented elsewhere
• Project-specific conventions → CLAUDE.md instead
• Mechanical constraints enforceable with automation