Claude skills vs GitHub Copilot skills: directories, surfaces, and the instructions layer

GitHub Copilot adopted the Agent Skills specification and did something none of the other implementations did: it reads Claude Code’s directory natively. A repository with .claude/skills/ committed is already a repository with Copilot skills. That makes “Claude skills vs Copilot skills” less a question of format and more a question of surfaces, always-on instruction layers, and which tool’s optional features you are leaning on.

This article covers where Copilot looks, which of its many surfaces run skills, how skills relate to Copilot’s older customization files, and every place the Copilot implementation differs from Claude Code’s.

Shared format

Both tools read the same SKILL.md:

---
name: my-skill
description: "Description of what the skill does and when to use it."
---
# My Skill
Detailed instructions for the skill...

That example is lifted almost verbatim from GitHub’s own documentation, and it is a valid Claude Code skill as written. Both tools inject the file into the agent’s context when the skill is chosen, both choose based on the description, and both recommend keeping the main file under 500 lines with detailed reference material split into separate files.

Directories

Scope	Claude Code	GitHub Copilot
Project	.claude/skills/<name>/SKILL.md	.github/skills/, .claude/skills/, or .agents/skills/
Personal	~/.claude/skills/<name>/SKILL.md	~/.copilot/skills/ or ~/.agents/skills/
Plugin	<plugin>/skills/ as /plugin:skill	Not a separate concept
Enterprise	Managed settings directory	Organization-level configuration through GitHub

The overlap on .claude/skills is the headline. A team where some people use Claude Code and some use Copilot can commit skills once to .claude/skills/ and both tools pick them up. If the team also includes Codex or Cursor users, .agents/skills/ is read by Copilot, Codex, and Cursor but not Claude Code, so .claude/skills/ remains the single path with the widest reach, and a CLI install to every detected agent covers the rest:

npm i -g skillmds
skillmd add github/review-and-refactor                    # all detected agents
skillmd add github/review-and-refactor -a github-copilot  # Copilot only

The GitHub Copilot install guide on SkillMD lists the paths and the manual copy steps.

Surfaces: where the skill runs

This is the biggest structural difference. Claude Code is one product with a terminal UI, an IDE extension, a desktop app, a web app, and cloud sessions, all reading the same skill directories. Copilot is a family of products, and skills reach all of them:

Copilot surface	Reads repo skills	Notes
Copilot CLI	Yes	Also reads personal skills
Copilot cloud coding agent	Yes	Runs in GitHub’s environment; repo skills only
Copilot code review	Yes	A review skill in the repo shapes automated PR reviews
Copilot app	Yes
VS Code agent mode	Yes	Personal skills too
JetBrains agent mode	Yes
Visual Studio	Yes

The code-review row is worth a second look. A skill named, say, review-and-refactor committed to the repository influences how Copilot reviews pull requests, not just how it writes code in the editor. Claude Code’s equivalent is the /code-review bundled skill, which a project skill of the same name replaces, and which runs as a forked subagent locally or through Claude’s GitHub integration.

Skills vs the older Copilot customization files

Copilot had a customization system before skills, and it still exists. Each file type answers a different question.

File	Loads	Use it for
.github/copilot-instructions.md	Every request in the repo	Facts and constraints: stack, conventions, do-nots
*.instructions.md with applyTo globs	Requests touching matching files	Area-specific constraints
*.prompt.md prompt files	When you run them	Reusable prompts you trigger by hand
*.agent.md custom agents	When you select the agent	A persona with its own tools and instructions
SKILL.md skills	When the task matches the description	Procedures and reference material, loaded on demand

The mapping to Claude Code is close:

Copilot	Claude Code
copilot-instructions.md	CLAUDE.md
*.instructions.md with applyTo	Nested CLAUDE.md in subfolders, or a paths-style approach via skill descriptions
Prompt files	Manual-only skills (disable-model-invocation: true)
Custom agents	Subagents in .claude/agents/
Skills	Skills

The rule for splitting content is the same in both ecosystems: always-on files hold constraints, skills hold procedures. If your copilot-instructions.md has grown a “how to add a migration” section with eight numbered steps, that section is a skill.

Frontmatter: what each tool honors

Copilot’s documentation describes name and description and defers to the open spec for the rest. Claude Code adds a long list of optional fields. What happens to each in Copilot:

Field	Claude Code	GitHub Copilot
name, description	Read	Read
license, compatibility, metadata	Read, passed through	Tolerated (spec fields)
allowed-tools	Pre-approves tools for the invoking turn	Not documented as honored
disable-model-invocation	Manual-only via /name	Not documented; Copilot triggers by description
user-invocable	Hide from the / menu	Not applicable
context: fork, agent, model, hooks, argument-hint	Read	Ignored

Unknown fields do not break loading in Copilot, so a Claude skill with context: fork still works there; it runs inline. The practical gap is manual-only: Copilot does not offer a frontmatter switch to stop a skill from auto-triggering. If a skill has side effects and you need to guarantee a human triggers it, in Copilot that is a prompt file rather than a skill.

Invocation

Action	Claude Code	GitHub Copilot
Automatic	By description	By description
Explicit	/skill-name [args], body reads $ARGUMENTS	Describe the task; Copilot selects the skill. Prompt files for explicit triggers
Manage installed skills	/skills menu, live reload	gh skill commands in the GitHub CLI
Discover community skills	Registries, plugins, claude.ai sync	gh skill, github/awesome-copilot, anthropics/skills, registries

Cloud behavior

Both tools have the same rule for cloud execution: personal skills stay on your machine.

• Copilot’s cloud coding agent and code review run in GitHub’s environment and see only skills committed to the repository.
• Claude Code’s cloud sessions and Cowork sessions do not read ~/.claude/skills/; they load skills enabled on your claude.ai account plus skills committed to .claude/skills/ in the cloned repo.

If it has to work in CI, in a cloud agent, or in automated review, it has to be in the repo.

Distribution and safety

Copilot points users at gh skill, the github/awesome-copilot collection, and Anthropic’s anthropics/skills repository. Claude Code points at plugins and claude.ai. Both accept anything spec-compliant from anywhere, which puts the safety question on you.

A skill is instructions the agent will follow, and Copilot’s code-review surface means a malicious skill committed to a repository could shape what automated review says about a pull request. Read skills before committing them, check bundled scripts, and prefer sources that review submissions. Every public skill on SkillMD passes a safety review before listing and shows a capabilities summary on its page; the registry lists 25,000+ skills and tags which agents each author tested with.

Writing one skill for both

1. Frontmatter: name, description, license. Nothing else unless you accept it will only take effect in Claude Code.
2. Folder name equals name, lowercase with hyphens, under 64 characters.
3. Body under 500 lines; long reference material in references/.
4. Do not rely on allowed-tools for correctness. In Copilot the skill must work with normal permission prompts.
5. If the skill must never auto-trigger, use disable-model-invocation: true for Claude Code and ship a prompt file for Copilot, or accept description-based triggering in Copilot.
6. Commit to .claude/skills/. Both tools read it. Add .github/skills/ only if your organization standardizes on it.
7. skillmd lint before you commit.

Summary table

Dimension	Claude Code	GitHub Copilot
Spec	Agent Skills SKILL.md	Agent Skills SKILL.md
Project dirs	.claude/skills	.github/skills, .claude/skills, .agents/skills
Personal dirs	~/.claude/skills	~/.copilot/skills, ~/.agents/skills
Surfaces	CLI, IDE extension, desktop, web, cloud sessions	CLI, cloud coding agent, code review, app, VS Code, JetBrains, Visual Studio
Always-on layer	CLAUDE.md	copilot-instructions.md, *.instructions.md
Explicit trigger	/name	Prompt files; description-based selection
Manual-only switch	disable-model-invocation	Not in frontmatter
Tool pre-approval	allowed-tools	Not documented
Subagent execution	context: fork	Custom agents (*.agent.md)
Skill management	/skills, live reload	gh skill
Body length guidance	Under 500 lines	Under 500 lines

Further reading

• Claude skills vs Cursor skills
• Claude skills vs Codex skills
• AI agent skills: the complete guide
• One SKILL.md for every agent: the compatibility matrix
• GitHub Copilot install guide