Claude Code Skill Expert

Guide for creating effective, well-structured Claude Code skills that follow Anthropic's official best practices.

When to Use This Skill

• Creating a new SKILL.md file
• Improving an existing skill's structure or effectiveness
• Auditing a skill for token efficiency
• Troubleshooting why a skill isn't being triggered
• Optimizing skill context usage

---

Core Principles

1. Progressive Disclosure (CRITICAL)

Structure skills as a table of contents where Claude loads content on-demand:

• Metadata (name + description) loads at startup
• SKILL.md loads when triggered
• Reference files load only as needed

Pattern: Keep SKILL.md under 500 lines. Move detailed content to separate reference files.

Example:

## Step 1: Analyze Code

Run the analyzer: `bun run analyzer.ts`

**Output**: See [output-templates.md#analysis](references/output-templates.md#analysis) for format

2. Name and Description Quality

The most critical fields - Claude uses these to decide when to trigger the skill.

Good description:

description: Books cinema tickets at Classic Cinemas. Use when Nathan asks to book tickets, see movies showing, or get cinema seats.

Bad description:

description: A helpful skill for booking things

Rules:

• Include what the skill does
• Include when to use it (trigger phrases)
• Use third person ("Books tickets" not "I book tickets")
• Max 1024 characters
• Include key terms users would mention

3. File Organization

One-level-deep references from SKILL.md:

skill/
├── SKILL.md              # Overview, workflow steps
└── references/
    ├── templates.md      # Output formats
    ├── commands.md       # CLI reference
    └── errors.md         # Error recovery

Don't do:

skill/
├── SKILL.md
└── references/
    ├── advanced/
    │   └── deep-file.md  # Too deep!

---

SKILL.md Structure

Recommended Format

---
name: my-skill
description: What it does and when to use it
allowed-tools: Bash, Read, Write
model: claude-sonnet-4-5
---

# Skill Title

Brief introduction explaining purpose.

## Workflow Overview

[Optional: Visual diagram of steps]

## Step 1: First Action

**Action**: What to do

**Command/Tool**: Specific tool or bash command

**Output**: Template reference or inline format

**Wait for**: User response or validation

---

## Step 2: Second Action

[Continue pattern...]

---

## References

- [Templates](references/templates.md) - Exact output formats
- [Commands](references/commands.md) - CLI reference
- [Errors](references/errors.md) - Recovery patterns

Key Elements

1. Workflow clarity: Numbered steps with clear actions
2. Template references: Link to exact formats
3. Validation checkpoints: Where to wait for user confirmation
4. Progressive disclosure: "See X for details" instead of inline detail

---

Output Templates Best Practice

CRITICAL: Create a separate references/output-templates.md file with exact copy-paste formats.

Template File Structure

# Output Templates

**CRITICAL**: Use these exact templates. Copy the structure precisely.

---

## Template Name

**ALWAYS use this exact format**:

\`\`\`
[Exact format with placeholders]
\`\`\`

**Rules**:
- Specific formatting requirements
- What to show/hide
- How to handle edge cases

**Mapping from data**:
\`\`\`
field_name → [PLACEHOLDER]
nested.field → [OTHER_PLACEHOLDER]
\`\`\`

**Example**:
\`\`\`
[Concrete example with real data]
\`\`\`

Why This Works

• Strictness levels: "ALWAYS use this exact format" vs "use your best judgment"
• Visual examples: Claude sees the desired output structure
• Mapping clarity: Links data sources to placeholders
• Progressive disclosure: Loaded only when needed for that step

---

Common Patterns

1. Conditional Workflow Pattern

## Step 2: Choose Path

**If** creating new content → Go to Step 3
**If** editing existing → Go to Step 5

2. Validation Loop Pattern

## Step 4: Validate

Run validator → identify errors → fix → repeat until passing

**See**: [error-handling.md#validation](references/error-handling.md#validation)

3. Tool Selection Pattern

## Step 1: Analyze Request

**Use**:
- `kit_grep` for literal text search
- `kit_semantic` for "find where we handle X"
- `kit_symbols` for function definitions

**See**: [tool-selection.md](references/tool-selection.md) for full decision tree

4. High-Level Guide with References

# My Skill

Keep SKILL.md as overview.

**Resources**:
- [Forms](references/forms.md) - Advanced form handling
- [API](references/api.md) - API details
- [Examples](references/examples.md) - Concrete examples

---

Token Efficiency Strategies

1. Split Mutually Exclusive Content

If certain contexts are rarely used together, keep paths separate:

references/
├── create-workflow.md    # Only for creating new items
└── edit-workflow.md      # Only for editing existing

2. Use Code for Deterministic Operations

Scripts don't load into context - only their output does.

Efficient:

bun run extract-fields.ts  # Code not in context

Inefficient:

Parse the JSON manually and extract these 20 fields...
[20 lines of parsing instructions]

3. Table of Contents for Long References

For files >100 lines:

# Long Reference File

## Contents
- [Section 1](#section-1)
- [Section 2](#section-2)

## Section 1
...

---

Troubleshooting

Skill Not Triggering

Likely causes:

1. Description doesn't match user's language
2. Name is too generic or doesn't convey purpose
3. Missing key trigger phrases

Fix: Add trigger phrases to description: "Use when users ask about X, need to Y, or mention Z"

Skill Over-Triggers

Likely causes:

1. Description too broad
2. Missing constraints

Fix: Narrow description with specific scenarios

Token Bloat

Likely causes:

1. Everything in SKILL.md
2. Templates inline instead of referenced
3. No progressive disclosure

Fix: Move details to reference files, use "See X" pattern

Inconsistent Output

Likely causes:

1. No explicit templates
2. Vague formatting instructions
3. Missing examples

Fix: Create output-templates.md with exact formats and examples

---

Quality Checklist

Before considering a skill complete:

• Name is gerund form (e.g., processing-pdfs)
• Description includes what + when to use
• SKILL.md under 500 lines
• Templates in separate file with exact formats
• One-level-deep references
• Clear numbered workflow steps
• Validation checkpoints defined
• Error handling documented
• Examples provided
• No nested subdirectories

---

Official Resources

• Skill Authoring Best Practices
• Claude 4 Prompting Guide
• Equipping Agents for the Real World

---

Example: Well-Structured Skill

See the-cinema-bandit plugin's ticket-booking skill for a production example:

skills/ticket-booking/
├── SKILL.md (127 lines)          # Lean workflow
└── references/
    ├── output-templates.md        # Strict formats
    ├── cli-commands.md            # CLI reference
    └── error-handling.md          # Recovery patterns

Key features:

• Progressive disclosure (47% smaller than original)
• Exact output templates with examples
• Clear action → output → wait pattern
• Visual workflow diagram
• Reference files loaded on-demand