Skill Structure Explained

What You'll Learn

Precisely understand all SKILL.md field requirements and format specifications
Master the design principles and use cases for references/, scripts/, assets/
Optimize skill token usage and loading performance
Avoid common format errors and path resolution issues
Use progressive loading to improve AI context efficiency

Current Challenges

You've learned to create basic skills, but your understanding of the complete SKILL.md specification is limited. Your skills may encounter these problems:

SKILL.md is too long, causing excessive token consumption
Uncertain about what content belongs in references/ versus SKILL.md
AI agents cannot correctly load resources from scripts/ or assets/
YAML frontmatter format errors causing installation failures

When to Use This

Skill review: Check if existing skills meet Anthropic specifications
Performance optimization: Resolve slow skill loading or token limit issues
Resource refactoring: Split large skills into SKILL.md + bundled resources
Complex skill development: Write complete skills with API docs and executable scripts

🎒 Prerequisites

Prerequisites Check

Before starting, ensure:

✅ Read Create Custom Skills
✅ Installed at least one skill (understand the basic workflow)
✅ Familiar with YAML and Markdown basic syntax

Core Concepts

SKILL.md Design Philosophy

SKILL.md is the core of the Anthropic skill system, designed with progressive loading:

mermaid

graph LR
    A[Metadata<br/>name + description] -->|Always loaded| B[Context]
    B -->|AI judges need| C[SKILL.md<br/>Core instructions]
    C -->|On-demand reference| D[Resources<br/>references/scripts/assets]

Three-layer loading advantages:

Metadata layer: All skills' name and description are always in context, enabling AI to quickly understand available skills
SKILL.md layer: Loaded only when relevant, containing core instructions (< 5000 words)
Resources layer: Detailed documentation and executable files loaded on-demand, avoiding token waste

Bundled Resources Classification

Directory	Loaded to Context	Use Cases	Example Types
`references/`	✅ On-demand	Detailed docs, API specs	API docs, database schemas
`scripts/`	❌ Not loaded	Executable code	Python/Bash scripts
`assets/`	❌ Not loaded	Templates, output files, images	JSON templates, boilerplate code

Follow Along

Step 1: Understand Complete YAML Frontmatter Specification

Why: YAML frontmatter is skill metadata and must meet strict specifications

SKILL.md must start and end with ---:

yaml

---
name: my-skill
description: Use this skill when you need to demonstrate proper format.
---

Required fields:

Field	Type	Format Requirements	Example
`name`	string	kebab-case, no spaces	`pdf-editor`, `api-client`
`description`	string	1-2 sentences, third person	`Use this skill to edit PDF files`

Common Errors

Error Example	Problem	Fix Method
`name: My Skill`	Contains spaces	Change to `name: my-skill`
`name: my_skill`	Underscore format	Change to `name: my-skill`
`description: You should use this when...`	Second person	Change to `description: Use this skill when...`
`description:` Too long	Over 100 words	Simplify to 1-2 sentence overview
Missing ending `---`	YAML not properly closed	Add ending delimiter

Source verification: OpenSkills uses non-greedy regex to validate format

typescript

// src/utils/yaml.ts
export function hasValidFrontmatter(content: string): boolean {
  return content.trim().startsWith('---');
}

export function extractYamlField(content: string, field: string): string {
  const match = content.match(new RegExp(`^${field}:\\s*(.+?)$`, 'm'));
  return match ? match[1].trim() : '';
}

Step 2: Write SKILL.md Body (Imperative Form)

Why: AI agents expect imperative instructions, not conversational descriptions

Correct approach:

markdown

## Instructions

To execute this task:

1. Read the input file
2. Process data using the algorithm
3. Generate output in specified format

Incorrect approach (avoid):

markdown

## Instructions

You should execute this task by:

1. Reading the input file
2. Processing data using the algorithm
3. Generating output in specified format

Comparison table:

✅ Correct (Imperative/Infinitive)	❌ Incorrect (Second Person)
"Load this skill when X"	"If you need Y"
"To accomplish Z, execute A"	"You should do Z"
"See references/guide.md"	"When you want to Z"

Writing guidelines:

Start with verbs: Create → Use → Return
Omit "You": Don't say "You should"
Clear paths: Use references/, scripts/, assets/ prefixes when referencing resources

Step 3: Use references/ for Detailed Documentation

Why: Keep SKILL.md concise, load detailed documentation on-demand

Applicable scenarios:

API documentation (endpoint descriptions over 500 words)
Database schemas (table structures, field definitions)
Detailed guides (configuration options, FAQ)
Code examples (large code snippets)

Directory structure:

my-skill/
├── SKILL.md              (~2,000 words, core instructions)
└── references/
    ├── api-docs.md       (Detailed API documentation)
    ├── database-schema.md (Database structure)
    └── troubleshooting.md (Troubleshooting guide)

Reference method in SKILL.md:

markdown

## Instructions

To interact with the API:

1. Read the request parameters
2. Call the API endpoint
3. For detailed response format, see `references/api-docs.md`
4. Parse the response
5. Handle errors (see `references/troubleshooting.md`)

references/api-docs.md example:

markdown

# API Documentation

## Overview

This API provides endpoints for data processing.

## Endpoints

### POST /api/process

**Request:**
```json
{
  "input": "data to process",
  "options": {
    "format": "json"
  }
}

Response:

json

{
  "status": "success",
  "result": {
    "output": "processed data"
  }
}

Error Codes:

400: Invalid input format
500: Server error


::: tip Best Practice

**references/ file size recommendations**:
- Single file: Recommend < 10,000 words
- Total size: Recommend < 50,000 words (split across multiple files)
- Naming: Use kebab-case (`api-docs.md` not `API_Docs.md`)

:::

---

### Step 4: Use scripts/ for Deterministic Tasks

**Why**: Executable scripts don't need to be loaded into context, suitable for repetitive tasks

**Applicable scenarios**:

- Data transformation (JSON → CSV, format conversion)
- File processing (compression, decompression, renaming)
- Code generation (generate code from templates)
- Test execution (unit tests, integration tests)

**Directory structure**:

my-skill/ ├── SKILL.md └── scripts/ ├── process.py (Python script) ├── transform.sh (Bash script) └── validate.js (Node.js script)


**Reference method in SKILL.md**:

```markdown
## Instructions

To process the input data:

1. Validate the input file format
2. Execute the processing script:
   ```bash
   python scripts/process.py --input data.json --output result.json

Verify the output file
If validation fails, see scripts/validate.py for error messages


**scripts/process.py example**:

```python
#!/usr/bin/env python3
import json
import sys

def main():
    input_file = sys.argv[1]
    output_file = sys.argv[2]

    with open(input_file, 'r') as f:
        data = json.load(f)

    # Processing logic
    result = transform_data(data)

    with open(output_file, 'w') as f:
        json.dump(result, f, indent=2)

    print(f"✅ Processed {input_file} → {output_file}")

if __name__ == "__main__":
    main()

Advantages of scripts/

Compared to inlining code in SKILL.md:

Feature	Inline Code	scripts/
Token usage	✅ High	❌ Low
Reusability	❌ Poor	✅ Good
Testability	❌ Difficult	✅ Easy
Complexity	❌ Token limit	✅ No limit

Step 5: Use assets/ for Templates and Output Files

Why: Templates and output files don't need to be loaded into context, saving tokens

Applicable scenarios:

Output templates (JSON, XML, Markdown templates)
Boilerplate code ( project scaffolding, config files)
Images and diagrams (flowcharts, architecture diagrams)
Test data (sample inputs, expected outputs)

Directory structure:

my-skill/
├── SKILL.md
└── assets/
    ├── template.json    (JSON template)
    ├── boilerplate.js   (Boilerplate code)
    └── diagram.png     (Flowchart)

Reference method in SKILL.md:

markdown

## Instructions

To generate the output file:

1. Load the template: `assets/template.json`
2. Replace placeholders with actual data
3. Write to output file
4. For boilerplate code, see `assets/boilerplate.js`

assets/template.json example:

json

{
  "title": "{{ title }}",
  "description": "{{ description }}",
  "version": "{{ version }}",
  "author": "{{ author }}",
  "created_at": "{{ timestamp }}"
}

Using templates in scripts:

python

import json
from string import Template

def generate_output(data, template_path):
    with open(template_path, 'r') as f:
        template_str = f.read()

    template = Template(template_str)
    output = template.safe_substitute(data)

    return output

assets/ Considerations

Not loaded to context: AI agents cannot directly read content, must load via scripts
Path resolution: Use relative paths, like assets/template.json
File size: Recommend single file < 10MB (avoid transfer delays)

Step 6: Optimize File Size and Performance

Why: File size directly affects AI context token consumption and loading speed

File size guidelines (official recommendations):

Directory	Size Limit	Loading Behavior
SKILL.md	< 5,000 words	Always loaded (when needed)
references/	No strict limit	Loaded on-demand
scripts/	Not counted	Not loaded, only executed
assets/	Not loaded	Not loaded, only copied

Performance optimization tips:

Split references/:

bash

# ❌ Single large file (20,000 words)
references/all-docs.md

# ✅ Split into multiple small files (each < 5,000 words)
references/
├── api-docs.md
├── database-schema.md
└── troubleshooting.md

Use scripts/ for data processing:
markdown
```
# ❌ Inline large code blocks in SKILL.md (consumes tokens)
## Instructions
Execute this code:
```python
# 500 lines of code...
```
✅ Reference scripts/ (no token consumption)
Instructions
Execute: python scripts/processor.py
Streamline SKILL.md:
- Keep only core instructions and steps
- Move detailed explanations to references/
- Use concise imperative language

Verify file size:

bash

# Count SKILL.md words
wc -w my-skill/SKILL.md

# Count references/ total words
find my-skill/references -name "*.md" -exec wc -w {} + | tail -1

# Check scripts/ file size
du -sh my-skill/scripts/

Step 7: Understand Resource Resolution Mechanism

Why: Understand path resolution rules to avoid reference errors

Base directory concept:

When AI agents load a skill, openskills read outputs the base directory:

Reading: my-skill
Base directory: /path/to/project/.claude/skills/my-skill

Relative path resolution rules:

Reference Path	Resolved Result
`references/api.md`	`/base/directory/references/api.md`
`scripts/process.py`	`/base/directory/scripts/process.py`
`assets/template.json`	`/base/directory/assets/template.json`

Source verification:

typescript

// src/commands/read.ts
export function readSkill(skillNames: string[] | string): void {
  const skill = findSkill(name);
  const content = readFileSync(skill.path, 'utf-8');

  // Output base directory for AI to resolve relative paths
  console.log(`Base directory: ${skill.baseDir}`);
  console.log(content);
}

Path Error Examples

❌ Incorrect Method	Problem	✅ Correct Method
`/absolute/path/to/api.md`	Uses absolute path	`references/api.md`
`../other-skill/references/api.md`	Cross-skill reference	`references/api.md`
`~/references/api.md`	Uses tilde expansion	`references/api.md`

Step 8: Validate Skill Format

Why: Validate format before installation to avoid runtime errors

Use openskills to validate:

bash

npx openskills install ./my-skill

You should see:

✔ Found skill: my-skill
  Description: Use this skill when you need to demonstrate proper format.
  Size: 2.1 KB

? Select skills to install: (Use arrow keys)
❯ ☑ my-skill

Validation checklist:

[ ] SKILL.md starts with ---
[ ] Contains name field (kebab-case)
[ ] Contains description field (1-2 sentences)
[ ] YAML ends with ---
[ ] Body uses imperative/infinitive form
[ ] All references/, scripts/, assets/ references use relative paths
[ ] SKILL.md word count < 5,000 words
[ ] references/ file naming uses kebab-case

Manual YAML frontmatter validation:

bash

# Check if starts with ---
head -1 my-skill/SKILL.md

# Verify YAML fields (using yq or other tools)
yq eval '.name' my-skill/SKILL.md

Step 9: Test Skill Loading

Why: Ensure skills can be correctly loaded into AI context

Use openskills read to test:

bash

npx openskills read my-skill

You should see:

Reading: my-skill
Base directory: /path/to/project/.claude/skills/my-skill

---
name: my-skill
description: Use this skill when you need to demonstrate proper format.
---

# My Skill

## Instructions

To execute this task...

## Bundled Resources

For detailed information: see `references/skill-format.md`

Skill read: my-skill

Checkpoints:

✅ Output includes Base directory (for path resolution)
✅ SKILL.md content is complete (including YAML and body)
✅ No "Invalid SKILL.md" error
✅ All reference paths display correctly

Checkpoint ✅

After completing the above steps, you should:

✅ Understand the complete SKILL.md field specifications
✅ Master the use cases for references/, scripts/, assets/
✅ Be able to optimize skill file size and loading performance
✅ Know how to validate skill format and test loading
✅ Understand resource resolution mechanism and base directory

Common Pitfalls

Issue 1: SKILL.md Exceeds 5000 Words Causing Token Limit

Cause: SKILL.md contains too much detailed documentation

Solution:

Move detailed content to references/ directory
Reference in SKILL.md: See references/guide.md for details
Use wc -w SKILL.md to check word count

Issue 2: scripts/ Scripts Cannot Execute

Cause:

Scripts lack execute permissions
Using absolute paths instead of relative paths

Solution:

bash

# Add execute permissions
chmod +x my-skill/scripts/*.sh

# Use relative paths in SKILL.md
## Instructions
Execute: `python scripts/process.py`  # ✅ Correct
Execute: `/path/to/my-skill/scripts/process.py`  # ❌ Incorrect

Issue 3: references/ Files Loaded On-Demand But AI Cannot Read

Cause: AI agents did not correctly parse references/ paths

Solution:

Confirm openskills read outputs Base directory
Clearly state when referencing: See references/api-docs.md in base directory
Avoid using absolute paths or cross-skill references

Issue 4: assets/ Files Too Large Causing Transfer Delays

Cause: assets/ stores large binary files (> 10MB)

Solution:

Compress images: Use PNG instead of BMP, optimize JPEG quality
Split data: Split large datasets into multiple small files
Use external storage: For extra-large files, provide download links instead of including directly

Issue 5: YAML Frontmatter Format Errors

Cause:

Missing ending ---
Field values contain special characters (colons, hashes) without quotes

Solution:

yaml

# ❌ Incorrect: Missing ending ---
---
name: my-skill
description: Use this skill: for testing
# Missing ---

# ✅ Correct: Complete closure
---
name: my-skill
description: "Use this skill: for testing"
---

Issue 6: Instructions Use Second Person

Cause: Habitual use of "You should", "When you want"

Solution:

Use imperative language starting with verbs
Use "To do X, execute Y" instead of "You should do Y"
Use "Load this skill when Z" instead of "If you need Z"

Comparison table:

Second Person (❌ Avoid)	Imperative (✅ Recommend)
"You should execute..."	"To execute X, run..."
"When you want to..."	"Load this skill when..."
"If you need..."	"Use X to accomplish Y"

Summary

Key points of skill structure:

YAML frontmatter: Required fields name (kebab-case) and description (1-2 sentences)
Body format: Use imperative/infinitive form, avoid second person
references/: Store detailed documentation, loaded on-demand to context (< 10,000 words/file)
scripts/: Store executable scripts, not loaded to context, suitable for deterministic tasks
assets/: Store templates and output files, not loaded to context
File size: SKILL.md < 5,000 words, references/ can be split, scripts/ no limit
Path resolution: Use relative paths (references/, scripts/, assets/), resolved based on base directory
Validation methods: Use openskills install to validate format, openskills read to test loading

Next Lesson Preview

In the next lesson, we'll learn CI/CD Integration.
You'll learn:
How to use the -y/--yes flag in CI/CD environments
Automate skill installation and sync workflows
Integrate OpenSkills in GitHub Actions, GitLab CI

Appendix: Source Reference

Click to expand source code locations

Last updated: 2026-01-24

Feature	File Path	Line
YAML frontmatter validation	`src/utils/yaml.ts`	12-14
YAML field extraction	`src/utils/yaml.ts`	4-7
Skill read command	`src/commands/read.ts`	1-49
Base directory output	`src/commands/read.ts`	42
Format validation on install	`src/commands/install.ts`	242, 291, 340

Example skill files:

examples/my-first-skill/SKILL.md - Complete structure example
examples/my-first-skill/references/skill-format.md - Format specification reference

Key functions:

hasValidFrontmatter(content: string): boolean - Validate if SKILL.md starts with ---
extractYamlField(content: string, field: string): string - Extract YAML field value (non-greedy matching)
readSkill(skillNames: string[] | string): void - Read skill to standard output (for AI use)

Getting Started

Platform Features

Advanced Features

FAQ

Appendix

Changelog

Skill Structure Explained

What You'll Learn

Current Challenges

When to Use This

🎒 Prerequisites

Core Concepts

SKILL.md Design Philosophy

Bundled Resources Classification

Follow Along

Step 1: Understand Complete YAML Frontmatter Specification

Step 2: Write SKILL.md Body (Imperative Form)

Step 3: Use references/ for Detailed Documentation

Step 5: Use assets/ for Templates and Output Files

Step 6: Optimize File Size and Performance

✅ Reference scripts/ (no token consumption)

Instructions

Step 7: Understand Resource Resolution Mechanism

Step 8: Validate Skill Format

Step 9: Test Skill Loading

Checkpoint ✅

Common Pitfalls

Issue 1: SKILL.md Exceeds 5000 Words Causing Token Limit

Issue 2: scripts/ Scripts Cannot Execute

Issue 3: references/ Files Loaded On-Demand But AI Cannot Read

Issue 4: assets/ Files Too Large Causing Transfer Delays

Issue 5: YAML Frontmatter Format Errors

Issue 6: Instructions Use Second Person

Summary

Next Lesson Preview

Appendix: Source Reference

Skill Structure Explained ​

What You'll Learn ​

Current Challenges ​

When to Use This ​

🎒 Prerequisites ​

Core Concepts ​

SKILL.md Design Philosophy ​

Bundled Resources Classification ​

Follow Along ​

Step 1: Understand Complete YAML Frontmatter Specification ​

Step 2: Write SKILL.md Body (Imperative Form) ​

Step 3: Use references/ for Detailed Documentation ​

Step 5: Use assets/ for Templates and Output Files ​

Step 6: Optimize File Size and Performance ​

✅ Reference scripts/ (no token consumption) ​

Instructions ​

Step 7: Understand Resource Resolution Mechanism ​

Step 8: Validate Skill Format ​

Step 9: Test Skill Loading ​

Checkpoint ✅ ​

Common Pitfalls ​

Issue 1: SKILL.md Exceeds 5000 Words Causing Token Limit ​

Issue 2: scripts/ Scripts Cannot Execute ​

Issue 3: references/ Files Loaded On-Demand But AI Cannot Read ​

Issue 4: assets/ Files Too Large Causing Transfer Delays ​

Issue 5: YAML Frontmatter Format Errors ​

Issue 6: Instructions Use Second Person ​

Summary ​

Next Lesson Preview ​

Appendix: Source Reference ​

Skill Structure Explained

What You'll Learn

Current Challenges

When to Use This

🎒 Prerequisites

Core Concepts

SKILL.md Design Philosophy

Bundled Resources Classification

Follow Along

Step 1: Understand Complete YAML Frontmatter Specification

Step 2: Write SKILL.md Body (Imperative Form)

Step 3: Use references/ for Detailed Documentation

Step 5: Use assets/ for Templates and Output Files

Step 6: Optimize File Size and Performance

✅ Reference scripts/ (no token consumption)

Instructions

Step 7: Understand Resource Resolution Mechanism

Step 8: Validate Skill Format

Step 9: Test Skill Loading

Checkpoint ✅

Common Pitfalls

Issue 1: SKILL.md Exceeds 5000 Words Causing Token Limit

Issue 2: scripts/ Scripts Cannot Execute

Issue 3: references/ Files Loaded On-Demand But AI Cannot Read

Issue 4: assets/ Files Too Large Causing Transfer Delays

Issue 5: YAML Frontmatter Format Errors

Issue 6: Instructions Use Second Person

Summary

Next Lesson Preview

Appendix: Source Reference