Skills (Commands)
Skills are reusable instruction sets that extend what an AI assistant can do. You invoke them with /skill-name (in Claude Code), $skill-name (in Codex) or the assistant loads them automatically when they match the current task.
Built-in vs custom skills
Claude Code includes built-in slash commands such as /init and /review, and you can create your own skills for workflows you repeat often.
~/.claude/skills/explain-code/SKILL.md (available in all projects):
---
name: explain-code
description: Explains code with diagrams and analogies
argument-hint: "[file-path]"
user-invocable: true
allowed-tools: Read, Grep, Glob
---
When explaining code:
1. **Analogy first**: Compare the code to something from everyday life
2. **Draw a diagram**: Use ASCII art to show flow or structure
3. **Walk through**: Step-by-step explanation of what happens
4. **Highlight a gotcha**: Common mistakes or misconceptions
Then invoke it:
/explain-code src/auth/middleware.ts
Key frontmatter fields
| Field | Purpose |
|---|---|
name | Identifier (lowercase, hyphens) |
description | When to auto-load; also shown in / menu |
argument-hint | Autocomplete hint, e.g. [file-path] |
user-invocable | true = appears in / menu; false = only auto-loaded |
allowed-tools / tools | Tool access control field (exact key can vary by feature/version) |
model | sonnet, opus, haiku, or inherit |
context | default (inline) or fork (isolated subagent) |
Dynamic content
Skills support string substitutions and inline command execution:
---
name: pr-summary
description: Summarize a pull request
---
Summary of PR $ARGUMENTS:
Changes: !`gh pr diff $ARGUMENTS`
Comments: !`gh pr view $ARGUMENTS --comments`
| Variable | Expands to |
|---|---|
$ARGUMENTS | All arguments passed to the skill |
$0, $1, … | Individual arguments (0-based) |
!`command` | Output of the shell command |
Organizing skill files
For complex skills, add supporting files alongside SKILL.md:
.claude/skills/my-skill/
├── SKILL.md # Main instructions (loaded first)
├── reference.md # API docs (loaded when needed)
├── examples.md # Usage examples
└── scripts/helper.sh # Executed, not loaded as context
Analogous features in other tools
Cursor uses rules files (.cursor/rules/) that inject instructions into conversations, serving a similar purpose to skills but without the invocation model or argument passing. GitHub Copilot has instructions files (.github/copilot-instructions.md) that set project-level context.
Sources: Anthropic Claude Code slash commands, Cursor rules documentation, GitHub Copilot custom instructions