Skill Creator — Complete Guide to Building Copilot Skills

You are an expert at creating high-quality skills for GitHub Copilot. When the user asks you to create a skill, follow this guide to produce a complete, well-structured result.

What is a Skill?

A skill is a SKILL.md file that injects domain-specific knowledge into the AI agent's context. Skills live in directories and can be distributed via Git repositories.

The agent receives the skill's content as system-level instructions — everything you write in the body of SKILL.md becomes part of the agent's working knowledge for that session.

Directory Structure

Every skill is a folder. The minimum is a SKILL.md, but a complete skill includes:

skills/
  <skill-name>/
    SKILL.md              ← REQUIRED: frontmatter + instructions
    FEEDBACK.md           ← RECOMMENDED: feedback protocol
    .skillconfig.json     ← OPTIONAL: per-skill settings (e.g. forceGlobal)
    references/           ← OPTIONAL: on-demand deep-dive docs
      api-reference.md
      examples.md
      troubleshooting.md

Naming Convention

• Folder name = skill identifier (kebab-case, lowercase)
• Must match the name field in frontmatter
• Examples: react-testing, docker-compose, git-workflow

SKILL.md — Anatomy

1. Frontmatter (YAML)

The frontmatter is a YAML block delimited by --- at the top of the file. It controls how the skill is discovered and invoked.

---
name: my-skill
description: "Clear, specific description of what this skill teaches"
argument-hint: What the user should type when invoking this skill
user-invocable: true
disable-model-invocation: false
compatibility:
  - copilot-chat
license: MIT
metadata:
  author: your-name
  tags:
    - category
    - domain
---

Field Reference

Field	Required	Type	Purpose
name	Yes	string	Unique identifier. Must match folder name. kebab-case.
description	Yes	string	What the skill teaches. Used for discovery — the agent reads this to decide if the skill is relevant.
argument-hint	No	string	Hint shown to users when invoking via slash command. Guides what context to provide.
user-invocable	No	boolean	If true, user can invoke directly via /skill-name. Default: true.
disable-model-invocation	No	boolean	If true, the model won't auto-invoke this skill. User must invoke explicitly. Default: false.
compatibility	No	string[]	Which Copilot surfaces support this skill (e.g. copilot-chat).
license	No	string	SPDX license identifier.
metadata	No	object	Free-form metadata: author, tags, version, etc.

2. Body (Markdown)

The body is the actual knowledge injected into the agent. This is where you write the instructions.

Structure Template

# <Skill Name> — <Short Subtitle>

<Opening paragraph: establish the agent's role and the skill's purpose>

## <Section 1: Core Concepts>
...

## <Section 2: Step-by-Step Workflows>
...

## <Section 3: Reference Tables>
...

## <Section N: When the User Asks for Help>
<Map common questions to specific answers>

Writing the Description — The Most Important Field

The description field is the single most critical element for skill discoverability. The agent reads descriptions to decide which skills to activate. A bad description means the skill never gets used.

Pattern: Type Tag + USE FOR + DO NOT USE

Every description should start with a type tag in bold, followed by a one-line summary, use cases, and exclusions:

description: "**<TYPE> SKILL** — <one-line summary>. USE FOR: <comma-separated use cases>. DO NOT USE FOR: <explicit exclusions to prevent false matches>."

Common type tags:

Tag	When to use
**WORKFLOW SKILL**	Guides a multi-step process or workflow
**ENFORCEMENT SKILL**	Enforces rules, checks, or constraints
**CONVERSION SKILL**	Converts between formats or representations
**MEASUREMENT SKILL**	Measures, benchmarks, or evaluates
**GUIDE SKILL**	Teaches how to use a tool or extension
**REFERENCE SKILL**	Provides domain knowledge or lookup tables

Choose the tag that best describes the skill's primary purpose. If none fits, create a descriptive one.

Pattern: Knowledge Skill Description

For skills that provide domain knowledge (alternative to type-tag pattern):

description: "**REFERENCE SKILL** — <Domain> knowledge covering <specific topics>. Includes <key features like examples, tables, commands>."

Rules for Good Descriptions

1. Be specific — "React testing" is vague. "Testing React components with React Testing Library and Jest, including async patterns and custom hooks" is specific.
2. Include action verbs — "Create, configure, debug, troubleshoot" tells the agent when to activate.
3. List explicit exclusions — "DO NOT USE FOR: runtime debugging" prevents false activations.
4. Front-load keywords — The most relevant terms should appear early.
5. Aim for ~300 chars — The hard limit is 1024, but ~300 chars is the sweet spot for LLM attention. Beyond that, the model may not weigh all keywords equally during discovery. If your description needs 400+ chars to cover all use cases, consider whether the skill should be split into two focused skills — each with a tight, discoverable description.

Writing the Body — Effective Instructions

Principle 1: Write for the Agent, Not the Human

The reader is an AI agent. Write clear, unambiguous instructions it can follow mechanically.

Bad:

Docker Compose is a nice tool for multi-container apps.

Good:

You are helping a developer work with Docker Compose. When asked to create a compose file, always use version 3.8+ syntax and include health checks.

Principle 2: Use Imperative Voice

Tell the agent what to DO, not what things ARE.

Bad:

The --watch flag enables file watching.

Good:

When the user wants live reload, add the --watch flag to the command.

Principle 3: Tables for Reference, Prose for Logic

Use tables for lookups (commands, options, mappings). Use prose for conditional logic and workflows.

## Commands

| Command | Purpose |
|---------|---------|
| `npm test` | Run all tests |
| `npm test -- --watch` | Watch mode |

## When to Use Watch Mode

If the user is actively developing and running tests repeatedly,
suggest watch mode. If they're running tests in CI, never suggest watch mode.

Principle 4: Map User Questions to Answers

Include a "When the User Asks..." section that covers common scenarios:

## When the User Asks for Help

- **"How do I X?"** → Step-by-step instructions for X
- **"Why is Y failing?"** → Common causes of Y failure and fixes
- **"What's the best way to Z?"** → Recommended approach with trade-offs

Principle 5: Progressive Disclosure with References

Don't overload the main SKILL.md. Put deep-dive content in references/:

• SKILL.md — Core instructions, workflows, quick-reference tables (aim for < 500 lines)
• references/ — API docs, extended examples, troubleshooting guides, migration paths

The agent will read references on-demand when it needs more detail, keeping the base context lightweight.

Principle 6: One Skill, One Domain

A skill should cover one coherent domain. If you find yourself writing unrelated sections, split into multiple skills.

Good scope: "React Testing" — covers RTL, Jest, async testing, hooks testing, snapshot testing Bad scope: "React" — too broad, mixes routing, state management, testing, styling

Advanced Body Techniques

These patterns improve agent navigation and scannability. Use them when the skill's complexity warrants it — not every skill needs all of them.

Technique: Decision Trees

For skills that involve triage, routing, or conditional workflows, provide an ASCII decision tree the agent can follow mechanically. This reduces reasoning errors by making the decision path explicit.

#### Triage Flow

\```
New request arrives
  │
  ├─ Condition A?
  │   └─ ✅ Action A
  │
  ├─ Condition B?
  │   └─ ⚠️ Action B (with caveat)
  │
  └─ Condition C?
      └─ 🔴 Escalate / ask the user
\```

When to use: Escalation logic, complexity triage, tool selection cascades, error diagnosis flows.

Technique: ❌/✅ Pitfall Markers

Instead of plain text rules, use visual markers for common mistakes. This format is more scannable than prose and communicates do/don't at a glance.

### Common Pitfalls

| | Pitfall | Consequence |
|---|---------|-------------|
| ❌ | Doing X without checking Y | Z breaks silently |
| ❌ | Assuming A is true | B fails in production |
| ✅ | Checking Y before doing X | Z works reliably |
| ✅ | Verifying A with tool | B is built on solid ground |

When to use: Rules sections, anti-patterns, common mistakes. Especially effective when the skill has 5+ rules — visual markers prevent "wall of NEVER" fatigue.

Technique: Quick Reference Table

A task-oriented lookup table at the end of the skill. Maps common user intents to the right section and key decision. Gives the agent a fast-path before reading the full skill.

## Quick Reference

| I need to... | Go to section | Key decision |
|-------------|---------------|-------------|
| Do X | Section Y | Choose between A or B |
| Debug Z | Troubleshooting | Check W first |

When to use: Skills with many sections (5+ headings), tool/command skills, reference-heavy skills.

Technique: Black-Box Script Instructions

For skills that bundle executable scripts in scripts/, instruct the agent to treat them as black boxes — run --help to learn usage, never read the source code.

> **⚠️ IMPORTANT: Treat scripts as black boxes.** Run `python scripts/<script>.py --help`
> to learn usage. Do **NOT** read the source code — it pollutes the context window with
> implementation details you don't need.

When to use: Any skill that includes bundled scripts, CLI tools, or executable utilities. Prevents the agent from wasting context on implementation details.

FEEDBACK.md — Enabling the Improvement Loop

A FEEDBACK.md enables users to submit improvements to the skill via PR.

Template

---
threshold: 5
---

## Feedback Protocol — <skill-name>

### When to Log a Review

Log a review whenever you help a user and:
- The instructions in SKILL.md were insufficient or unclear
- You had to improvise an answer not covered by the skill
- A command, API, or tool changed and the skill is outdated

### Review Format

Create a JSON file in `.vscode/skill-reviews/<skill-name>/`:

\```json
{
  "date": "YYYY-MM-DD",
  "author": "dev-name",
  "type": "improvement | correction | addition",
  "section": "Section Name",
  "suggestion": "What should change",
  "context": "What prompted this feedback"
}
\```

### Consolidation

When threshold reviews accumulate, summarize them into a single
actionable improvement for the skill maintainer.

The threshold in frontmatter defines how many reviews trigger consolidation.

.skillconfig.json — Per-Skill Settings

{
  "forceGlobal": false
}

Field	Type	Purpose
forceGlobal	boolean	If true, this skill is always installed globally (~/.copilot/skills/) regardless of context. Use for universal skills.

Complete Example — Creating a "Docker Compose" Skill

Step 1: Create the directory

skills/docker-compose/

Step 2: Write SKILL.md

---
name: docker-compose
description: "Docker Compose workflows — create, debug, and optimize multi-container setups. Covers v2 CLI, health checks, networking, volumes, and production profiles."
argument-hint: Describe what you want to do with Docker Compose
---

# Docker Compose — Agent Guide

You are helping a developer work with Docker Compose. Follow these
instructions to provide accurate, production-ready guidance.

## Creating a Compose File

When asked to create a new compose file:
1. Use the `compose.yaml` filename (not `docker-compose.yml`)
2. Target Compose Specification (no `version` field needed)
3. Always include `restart: unless-stopped` for services
4. Add health checks for services that accept connections

## Common Services Table

| Service | Image | Default Port |
|---------|-------|-------------|
| PostgreSQL | postgres:16-alpine | 5432 |
| Redis | redis:7-alpine | 6379 |
| Nginx | nginx:alpine | 80/443 |

## When the User Asks for Help

- **"Create a compose file for X"** → Scaffold with best practices above
- **"My container won't start"** → Check logs: `docker compose logs <service>`
- **"How do I add a database?"** → Add service from table, include volume + healthcheck

Step 3: Add FEEDBACK.md

Use the template from the FEEDBACK.md section above.

Step 4: Add references/ (if needed)

references/
  networking.md       ← Deep dive on Docker networking
  production.md       ← Production deployment patterns
  troubleshooting.md  ← Common errors and fixes

Quality Checklist

Before considering a skill complete, verify:

• name matches folder name (kebab-case)
• description is specific, keyword-rich, ideally ~300 chars (max 1024; if >400, consider splitting the skill)
• license field present in frontmatter
• Body opens with role context — "You are helping a developer with..."
• Imperative voice — instructs the agent what to do
• Tables for reference data, prose for workflows
• "When the User Asks" section covers 3+ common scenarios
• SKILL.md under 500 lines — heavy content goes to references/
• FEEDBACK.md included with review format
• No unrelated topics — one skill, one domain
• Decision tree included if the skill has triage/routing logic (optional)
• ❌/✅ pitfalls used instead of plain text rules when 5+ rules exist (optional)
• Quick Reference table at the end for complex skills with 5+ sections (optional)
• Black-box script instruction present if skill bundles scripts in scripts/ (optional)
• Tested — invoke the skill in Copilot Chat and verify it guides correctly

Anti-Patterns to Avoid

Anti-Pattern	Why It's Bad	Fix
Vague description	Agent can't match it to user intent	Use specific keywords + USE FOR/DO NOT USE FOR
Wall of text	Agent loses focus	Break into sections with headers
Passive voice	Agent doesn't know what to do	Use imperative: "When X, do Y"
Too broad scope	Diluted knowledge, conflicts with other skills	Split into focused skills
Everything in SKILL.md	Bloats context window	Move deep-dives to references/
No examples	Agent generates inconsistent output	Add input/output examples
Human-oriented prose	Agent processes instructions, not explanations	Write for the agent as the reader