Skill Creator

Skill Creator Guide

This guide helps you create effective skills for Apollo GraphQL and GraphQL development following the Agent Skills specification.

What is a Skill?

A skill is a directory containing instructions that extend an AI agent's capabilities with specialized knowledge, workflows, or tool integrations. Skills activate automatically when agents detect relevant tasks.

Directory Structure

A skill requires at minimum a SKILL.md file:

skill-name/
├── SKILL.md              # Required - main instructions
├── references/           # Optional - detailed documentation
│   ├── topic-a.md
│   └── topic-b.md
├── scripts/              # Optional - executable helpers
│   └── validate.sh
├── templates/            # Optional - config/code templates
│   └── config.yaml
└── assets/               # Optional - static resources (images, schemas, data files)

SKILL.md Format

Frontmatter (Required)

---
name: skill-name
description: >
  A clear description of what this skill does and when to use it.
  Include trigger conditions: (1) first condition, (2) second condition.
license: MIT
compatibility: Works with Claude Code and similar AI coding assistants.
metadata:
  author: your-org
  version: "1.0.0"
allowed-tools: Read Write Edit Glob Grep
---

Frontmatter Fields

Name Rules

• Use lowercase letters, numbers, and hyphens only
• Do not start or end with a hyphen
• Do not use consecutive hyphens (--)
• Must match the parent directory name

Good: apollo-client, graphql-schema, rover Bad: Apollo-Client, -apollo, apollo--client

Description Best Practices

Write descriptions that help agents identify when to activate the skill:

# Good - specific triggers and use cases
description: >
  Guide for designing GraphQL schemas following industry best practices. Use this skill when:
  (1) designing a new GraphQL schema or API,
  (2) reviewing existing schema for improvements,
  (3) deciding on type structures or nullability,
  (4) implementing pagination or error patterns.

# Bad - vague and unhelpful
description: Helps with GraphQL stuff.

Body Content

The markdown body contains instructions the agent follows. Structure it for clarity:

Recommended Sections

1. Overview - Brief explanation of the skill's purpose
2. Process - Step-by-step workflow (use checkboxes for multi-step processes)
3. Quick Reference - Common patterns and syntax
4. Security - Risks, mitigations, and validation (if the skill touches anything security-sensitive)
5. Reference Files - Links to detailed documentation
6. Key Rules - Important guidelines organized by topic
7. Ground Rules - Critical do's and don'ts

Example Structure

# Skill Title

Brief overview of what this skill helps with.

## Process

Follow this process when working on [task]:

- [ ] Step 1: Research and understand requirements
- [ ] Step 2: Implement the solution
- [ ] Step 3: Validate the result

## Quick Reference

### Common Pattern

\`\`\`graphql
type Example {
  id: ID!
  name: String
}
\`\`\`

## Security

> **Risk: [brief description of what can go wrong].**
> [What the user MUST do to prevent it.]

- ALWAYS [secure default behavior]
- NEVER [dangerous configuration] in production

## Reference Files

- [Topic A](references/topic-a.md) - Detailed guide for topic A
- [Topic B](references/topic-b.md) - Detailed guide for topic B

## Key Rules

### Category One

- Rule about this category
- Another rule

### Category Two

- Rule about this category

## Ground Rules

- ALWAYS do this important thing
- NEVER do this problematic thing
- PREFER this approach over that approach

Security-Sensitive Content

When a skill generates configuration, code, or guidance that could cause security issues if misused, the skill MUST make those risks explicit and visible to the LLM. An LLM cannot infer security implications from context alone — it needs clearly labeled signals.

When does a skill need security guidance?

If any of these apply, the skill is security-sensitive:

• Generates config that controls access to data (caching, auth, CORS, permissions)
• Handles secrets, credentials, or tokens
• Produces code that runs with elevated privileges
• Controls what data is shared, public, or exposed to users
• Configures network bindings, endpoints, or external access

How to surface security in a skill

1. Dedicated Security section in SKILL.md or a reference file, labeled ## Security. Not "Private data" or "Customization" — use the word "Security" so the LLM recognizes the category.

2. Explicit warnings at the point of risk — place security guidance next to the config or code that creates the risk, not in a separate file the LLM may not load:

   ### Response caching scope
   
   > **Security: data leakage risk.** All cached data is PUBLIC by default.
   > User-specific fields MUST use `scope: PRIVATE` with a `private_id`
   > configured, or they will be shared across all users.

3. Validation checklist items — every security-sensitive feature must have corresponding checks in the validation checklist. Group them under a ## Security heading.

4. Ground rules — add ALWAYS/NEVER rules for security-critical behavior. These are the strongest signal to the LLM.

5. Require the data model — if correct security configuration depends on understanding the user's data model (e.g., which fields are user-specific), the skill must instruct the LLM to ask the user before generating config. Do not let the LLM guess.

Anti-patterns

• Describing a security-sensitive default (like "public by default") without labeling it as a security concern
• Placing security guidance only in reference files that load on demand — the SKILL.md itself must contain the key warnings
• Using soft language ("you may want to consider") for hard security requirements — use "MUST" and "NEVER"
• Assuming the LLM understands which fields in a schema are private — require explicit user input

Progressive Disclosure

Structure skills to minimize context usage:

1. Metadata (~100 tokens): name and description load at startup for all skills
2. Instructions (< 5000 tokens): Full SKILL.md loads when skill activates
3. References (as needed): Files in references/ load only when required

Keep SKILL.md under 500 lines. Move detailed documentation to reference files.

Reference Files

Use references/ for detailed documentation:

references/
├── setup.md          # Installation and configuration
├── patterns.md       # Common patterns and examples
├── troubleshooting.md # Error solutions
└── api.md            # API reference

Reference files should be:

• Focused on a single topic
• Self-contained (readable without other files)
• Under 300 lines each

Link to references from SKILL.md:

## Reference Files

- [Setup](references/setup.md) - Installation and configuration
- [Patterns](references/patterns.md) - Common patterns and examples