Claude Code is already one of the most capable AI coding agents, and its custom skills system is what makes it genuinely extensible. Where most AI coding tools limit you to passive instruction files, Claude Code skills are executable, scriptable and isolatable: a first-class extension mechanism built on the open Agent Skills standard.
What Are Claude Code Skills?
A skill is a directory holding a SKILL.md file, where YAML frontmatter controls how it activates and Markdown instructions tell Claude what to do. Skills live in .claude/skills/ (project-wide) or ~/.claude/skills/ (personal), and there are three ways to invoke them:
/skill-namecommand — Direct slash invocation.- Automatic invocation — Claude reads the
descriptionfield and activates the skill based on conversation context. - On mention — Triggered when Claude detects relevant keywords.
Skills replaced the older custom commands system. A file at .claude/skills/deploy/SKILL.md does everything .claude/commands/deploy.md did, and on top of that adds directories for supporting files, rich frontmatter for invocation control, and automatic loading.
Creating Your First Skill
The layout is simple:
my-skill/
├── SKILL.md # Main instructions (required)
├── template.md # Template for Claude to fill in
├── examples/
│ └── sample.md
└── scripts/
└── validate.shA minimal skill:
---
name: summarize-changes
description: Summarizes uncommitted changes and flags risks.
allowed-tools: Bash(git *)
---
!`git diff HEAD`
Summarize the changes above in two or three bullet points, then list
any risks: missing error handling, hardcoded values, tests needing update.The frontmatter fields that drive most of the behaviour: disable-model-invocation restricts a skill to user-only invocation, the right setting for deploys; user-invocable: false hides a skill from the slash menu and keeps it as background knowledge; paths globs scope activation to the files being touched; and hooks wire the skill to lifecycle events.
Advanced Features
Dynamic Context Injection. The !`command` syntax runs shell commands before the skill content reaches Claude, injecting the output inline. Multi-line commands use ```! fenced blocks. A skill can fetch live data — git diffs, PR comments, build status — without external tooling.
Subagent Isolation. Setting context: fork runs the skill in an isolated subagent, separate from the main conversation, so its work stays out of the session’s context and several skills can run in parallel. The agent field picks the subagent type: Explore, Plan, general-purpose, or a custom one.
Argument Substitution. Skills take positional and named arguments. $0, $ARGUMENTS, $ARGUMENTS[1], and $name for declared parameters are all available, and meta-variables like ${CLAUDE_SESSION_ID}, ${CLAUDE_EFFORT}, and ${CLAUDE_SKILL_DIR} give a skill access to its own environment.
Tool Pre-Approval. The allowed-tools and disallowed-tools fields pre-approve or restrict tools for one skill invocation, which removes the repetitive permission prompts. A PR review skill can pre-approve Bash(gh *) and keep the rest of the session under normal permission checks.
Comparison with Cursor Rules and Copilot Instructions
Claude Code skills work differently from the passive instruction files of the other tools. Cursor’s .cursorrules is always loaded into context, sits there as one monolithic block, and has no argument system, no dynamic shell injection and no subagent isolation. Copilot’s .github/copilot-instructions.md is a single passive document that is read and never executed.
Skills are on-demand execution units in place of always-on context, which frees token budget for the conversation. The description budget system manages how many skill descriptions are loaded at once, dropping the least-used entries when capacity runs low. Because skills sit on the Agent Skills open standard (agentskills.io), they can potentially work across compatible tools; Cursor and Copilot extensions stay locked to their platform.
Practical Examples
A codebase visualizer skill bundles a Python script alongside its SKILL.md, locates the script through ${CLAUDE_SKILL_DIR}, pre-approves Bash(python3 *), and generates an interactive HTML tree of the project with no manual steps. A PR review skill uses context: fork to work in isolation, injects the PR diff via !`gh pr diff`, and pre-approves Bash(gh \*): a complete security-aware review in seconds.
Other use cases include deploy workflows, batch file transformations, security checklists, and technical writing linting. Claude Code ships with bundled skills like /code-review, /simplify, /debug, and /batch, all built with the same system that is available to every user.
Community and Plugins
Skills can be packaged into plugins: bundles that carry hooks, MCP servers, agents and multiple skills in a single directory with a plugin.json manifest. Anthropic maintains an official curated plugin repository with hundreds of community-contributed skills. The skill-creator plugin provides evaluation tooling: test case authoring, isolated subagent runs, pass/fail grading, and blind A/B comparison between skill versions.
For teams, project skills committed to version control are shared automatically with every developer. Enterprise users can push skills organization-wide through managed settings.
The Fable 5 Impact
The US government shutdown of Claude Fable 5 in June 2026 sent ripples through the Claude Code ecosystem. Users who had built skills relying on Fable 5’s advanced reasoning — complex code review, vulnerability research, multi-step analysis — found those skills suddenly underpowered. Skills may be portable through the Agent Skills standard, but model access remains a fragile dependency.
Skills written with model-agnostic instructions and robust error handling fared better during the transition, and the shutdown accelerated interest in fallback-aware patterns that degrade gracefully when frontier models become unavailable.
Claude Code skills are the most thoughtfully designed extension system in AI coding tools today. Dynamic context injection, subagent isolation, argument substitution, and tool pre-approval, all on an open standard, give it capabilities that Cursor Rules and Copilot instructions simply cannot match. For teams willing to invest in their skill library, the payoff is a coding assistant that genuinely adapts to how they work.