Claude Skills Overview 2026

Claude Code Skills System - Complete Reference (March 2026)

Skills extend what Claude can do. Create a SKILL.md file with instructions, and Claude adds it to its toolkit. Claude uses skills when relevant, or you can invoke one directly with /skill-name.

To create or improve skills using a guided workflow, load the /plugin-creator:skill-creator skill.

Skills and slash commands are now unified - they are the same system. A file at .claude/commands/review.md and a skill at .claude/skills/review/SKILL.md both create /review and work identically. If a skill and a command share the same name, the skill takes precedence. Skills are the recommended approach as they support additional features like supporting files and advanced frontmatter options.

Portable skills? This reference covers Claude Code-specific features (hooks, context fork, model selection, invocation control). If you need to create skills that work across Claude Code, Cursor, Gemini CLI, OpenAI Codex, VS Code, and 20+ other agents, see the ../agentskills/SKILL.md skill instead — it covers the portable subset of the format defined at agentskills.io.

Bundled Skills

Bundled skills ship with Claude Code and are available in every session. Unlike built-in commands (which execute fixed logic), bundled skills are prompt-based: they give Claude a detailed playbook and let it orchestrate work using its tools (spawn parallel agents, read files, adapt to your codebase).

• /simplify: Reviews recently changed files for code reuse, quality, and efficiency issues, then fixes them. Spawns three review agents in parallel (code reuse, code quality, efficiency), aggregates findings, and applies fixes. Pass optional text to focus: /simplify focus on memory efficiency.

• /batch <instruction>: Orchestrates large-scale changes across a codebase in parallel. Researches the codebase, decomposes work into 5–30 independent units, presents a plan for approval. Once approved, spawns one background agent per unit in an isolated git worktree, each implementing its unit, running tests, and opening a pull request. Requires a git repository.

• /debug [description]: Troubleshoots your current Claude Code session by reading the session debug log. Optionally describe the issue to focus the analysis.

• /loop [interval] <prompt>: Schedules a recurring prompt at an interval. See resources/scheduled-tasks.md.

• /claude-api: Loads Claude API reference material for your project's language (Python, TypeScript, Java, Go, Ruby, C#, PHP, or cURL) and Agent SDK reference. Covers tool use, streaming, batches, structured outputs, and common pitfalls. Also activates automatically when your code imports anthropic, @anthropic-ai/sdk, or claude_agent_sdk.

Source: Official documentation at https://code.claude.com/docs/en/skills.md (section: "Bundled skills", accessed 2026-03-17)

SKILL.md Complete Format

---
name: my-skill
description: What this Skill does and when to use it
argument-hint: "[optional-arg]"
allowed-tools: Read, Grep, Glob
model: claude-sonnet-4-20250514
context: fork
agent: general-purpose
user-invocable: true
disable-model-invocation: false
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate.sh"
          once: true
  PostToolUse:
    - matcher: "Write|Edit"
      hooks:
        - type: command
          command: "./scripts/lint.sh"
  Stop:
    - hooks:
        - type: command
          command: "./scripts/cleanup.sh"
---

# Skill Title

Your instructions here...

Validation: Use claude plugin validate to validate plugin structure. For skills bundled in plugins, see ../claude-plugins-reference-2026/SKILL.md for plugin.json schema requirements.

All Frontmatter Fields

All fields are optional. Only description is recommended so Claude knows when to use the skill.

The fields name, description, license, compatibility, metadata, and allowed-tools are part of the ../agentskills/SKILL.md and are portable across all compatible agents. The remaining fields below are Claude Code extensions.

[!IMPORTANT]

When an allowed-tools field is not specified, the skill inherits the tool capabilities of the parent agent. This is a common pattern for skills that need to use tools from the parent agent. Such as when a skill is used by the orchestrator agent for knowledge or task information, no allowed-tools field is needed.

The allowed-tools field grants permission for the listed tools while the skill is active, so Claude can use them without prompting you for approval. It does not restrict which tools are available: every tool remains callable, and your permission settings still govern tools that are not listed. To block a skill from using certain tools, add deny rules in your permission settings instead.

SOURCE: Extend Claude with skills section "Pre-approve tools for a skill" (accessed 2026-04-23)

[!NOTE] name: is required per the Agent Skills open standard. A bug in Claude Code v2.1.23 caused plugin skills with name: to not appear as slash commands; that bug was resolved 2026-02-20. Include name: in all SKILL.md files. The value must match the directory name.

Skill Tokenomics

Skills use progressive disclosure - only frontmatter loads initially (~100 tokens), full content loads on activation.

The context window is a public good shared with everything else Claude needs to know. Only add context Claude does not already have. Before writing each section, challenge it:

• Does Claude really need this information to complete the task?
• Can I assume Claude already knows this from training?
• Does this paragraph justify its token cost?

SOURCE: Anthropic skill-authoring best practices (docs.anthropic.com, accessed 2026-03-23)

Budget Constraints

YAML Multiline Bug

Use single-line strings for descriptions — Claude Code's skill indexer does not parse YAML multiline indicators (>-, |, |-) correctly; the description may appear as ">-" instead of actual content. Quote the value only when YAML syntax requires it (colons, leading special characters, boolean literals).

# Correct — single quoted string
description: 'Extract text from PDFs. Use when the user mentions PDFs or document extraction.'

# Wrong — multiline indicators break the indexer
description: >-
  This appears as ">-" in the index

Truncation Behavior

When total skill metadata exceeds the budget (2% of context window, fallback 16,000 characters):

1. Skills are truncated from the <available_skills> block
2. Truncated skills cannot be auto-invoked by Claude
3. User can still invoke truncated skills explicitly with /skill-name
4. Run /context to check for a warning about excluded skills

To override the limit, set the SLASH_COMMAND_TOOL_CHAR_BUDGET environment variable.

Source: Official documentation at https://code.claude.com/docs/en/skills.md (section: "Troubleshooting")

Fallback Strategy

If you have many skills, embed pointers in CLAUDE.md as a safeguard:

## Skills Available
- For debugging: use `/scientific-method:scientific-thinking` skill
- For delegation: use `/agent-orchestration:delegate` skill

This ensures Claude can find skills even if truncated from <available_skills>.

String Substitutions

Skills support string substitution for dynamic values in the skill content:

Example with positional arguments:

---
description: Migrate a component from one framework to another
---

Migrate the $0 component from $1 to $2.
Preserve all existing behavior and tests.

Running /migrate-component SearchBar React Vue replaces $0 with SearchBar, $1 with React, and $2 with Vue.

Dynamic Context Injection

The exclamation then backtick syntax runs shell commands before the skill content is sent to Claude. The command output replaces the placeholder, so Claude receives actual data, not the command itself.

Example: This skill summarizes a pull request by fetching live PR data with the GitHub CLI: @resources/pr-summary-example.md

How it works:

1. Each `!`command`` executes immediately (before Claude sees anything)
2. The output replaces the placeholder in the skill content
3. Claude receives the fully-rendered prompt with actual PR data

This is preprocessing, not something Claude executes. Claude only sees the final result.

Extended Thinking: To enable extended thinking mode in a skill, include the word "ultrathink" anywhere in your skill content.

Source: Official documentation at https://code.claude.com/docs/en/skills.md (sections: "Inject dynamic context", "Advanced patterns")

Directory Structure

skill-name/
├── SKILL.md              # Required
├── references/           # Optional: docs loaded on demand
├── scripts/              # Optional: executed, not loaded into context
├── templates/            # Optional: reusable content
└── examples/             # Optional: sample output showing expected format

Supporting files can also live at the skill root (e.g., reference.md, forms.md). Reference them from SKILL.md so Claude knows what each file contains and when to load it.

Subdirectory Warning — Skills Must Be One Level Deep

Skill subdirectory warning — nested skill directories silently fail to register, with a wrong/right example; read before organizing related skills into grouping subdirectories. Unrelated to the "Automatic Discovery from Nested Directories" behavior below, which discovers separate .claude/skills/ roots in subdirectories of your project, not grouping folders inside a single skills/ directory.

Location Priority (Highest to Lowest)

1. Managed/Enterprise - System-level (see managed settings)
2. Personal - ~/.claude/skills/
3. Project - .claude/skills/
4. Plugin - Bundled with plugins (see ../claude-plugins-reference-2026/SKILL.md)

When skills share the same name across levels, higher-priority locations win: enterprise > personal > project. Plugin skills use a plugin-name:skill-name namespace, so they cannot conflict with other levels.

Automatic Discovery from Nested Directories

When you work with files in subdirectories, Claude Code automatically discovers skills from nested .claude/skills/ directories. For example, if you're editing a file in packages/frontend/, Claude Code also looks for skills in packages/frontend/.claude/skills/. This supports monorepo setups where packages have their own skills.

Skills from Additional Directories

Skills defined in .claude/skills/ within directories added via --add-dir are loaded automatically and picked up by live change detection, so you can edit them during a session without restarting.

Note: CLAUDE.md files from --add-dir directories are not loaded by default. Set CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 to load them. See memory documentation for details.

Source: Official documentation at https://code.claude.com/docs/en/skills.md (section: "Where skills live")

Hooks in Skills

Skills can define hooks in frontmatter to respond to events during the skill's lifecycle. Use Skill(skill: "plugin-creator:hooks-guide") for complete hook documentation.

Hook Events

• PreToolUse: Before tool executes
• PostToolUse: After successful execution
• Stop: When Skill finishes

Hook Structure

hooks:
  PreToolUse:
    - matcher: "Bash"           # Regex pattern
      hooks:
        - type: command
          command: "./scripts/check.sh"
          once: true            # Run only once per session

Hook I/O

• Receives JSON via stdin (session info, tool name, parameters)
• Exit 0: Success
• Exit 2: Blocking error (prevents tool, shows stderr)
• Other: Non-blocking error

For complete hook configuration including all events, matchers, JSON output control, and examples, use Skill(skill: "plugin-creator:hooks-guide").

Context Fork Behavior

Add context: fork to your frontmatter when you want a skill to run in isolation. The skill content becomes the prompt that drives the subagent. It won't have access to your conversation history.

Use context: fork only when the skill includes explicit instructions and a clear task. If the skill contains guidelines like "use these API conventions" without a task, the subagent receives the guidelines but no actionable prompt and may return without meaningful output.

Agent Types

context: fork
agent: Explore

Tool Restrictions in Forked Contexts

VERIFIED BEHAVIOR (experimentally confirmed 2026-01-22):

When context: fork is set, the forked subagent has access to:

• File operations: Read, Write, Edit, Grep, Glob
• Web operations: WebSearch, WebFetch
• MCP tools (if configured)
• Bash and other system tools (depending on agent type)

The Agent tool is NOT available in forked contexts. This means forked skills cannot delegate to other subagents. If you need hierarchical delegation (subagent delegates to another subagent), the parent must run in the main context (no context: fork), not in a forked context.

Source: Experimental verification on 2026-01-22. Official documentation at https://code.claude.com/docs/en/skills.md does not explicitly document this restriction.

Skills vs Subagents

Skills and subagents work together in two directions:

With context: fork, you write the task in your skill and pick an agent type to execute it. For the inverse (defining a custom subagent that uses skills as reference material), see the Sub-Agents documentation.

Skills vs Agent Teams

Agent teams coordinate multiple independent Claude Code instances with inter-agent messaging and a shared task list. Unlike subagents (which report back to caller only), teammates communicate directly with each other.

Use agent teams when 3+ workers need to share findings, challenge each other, and coordinate independently. Use subagents when workers report back without needing cross-communication.

For complete agent teams reference including architecture, display modes, lifecycle management, use case patterns, and quality gate hooks (TeammateIdle, TaskCompleted), see resources/agent-teams.md.

Source: Agent Teams Documentation (accessed 2026-02-06)

Invocation Control

Control who invokes a skill

By default, both you and Claude can invoke any skill. You can type /skill-name to invoke it directly, and Claude can load it automatically when relevant to your conversation. Two frontmatter fields let you restrict this:

• disable-model-invocation: true: Only you can invoke the skill. Use this for workflows with side effects or that you want to control timing, like /commit, /deploy, or /send-slack-message. You don't want Claude deciding to deploy because your code looks ready. This field also removes the skill description from Claude's context entirely.

• user-invocable: false: Only Claude can invoke the skill. Use this for background knowledge that isn't actionable as a command. A legacy-system-context skill explains how an old system works. Claude should know this when relevant, but /legacy-system-context isn't a meaningful action for users to take.

How invocation control affects context loading:

In a regular session, skill descriptions are loaded into context so Claude knows what's available, but full skill content only loads when invoked. Subagents with preloaded skills work differently: the full skill content is injected at startup. Subagents do not inherit skills from the parent conversation — they must be listed explicitly in the skills frontmatter field. Skills that set disable-model-invocation: true cannot be preloaded. See resources/claude-code-skills-official.md for the subagent skills field syntax and model resolution priority.

SOURCE: Create custom subagents sections "Preload skills into subagents" and "Choose a model" (accessed 2026-04-23)

Source: Official documentation at https://code.claude.com/docs/en/skills.md (section: "Control who invokes a skill")

Restrict Claude's skill access

By default, Claude can invoke any skill that doesn't have disable-model-invocation: true set. Skills that define allowed-tools grant Claude access to those tools without per-use approval when the skill is active. Your permission settings still govern baseline approval behavior for all other tools. Built-in commands like /compact and /init are not available through the Skill tool.

Three ways to control which skills Claude can invoke:

Disable all skills by denying the Skill tool in /permissions:

# Add to deny rules:
Skill

Allow or deny specific skills using permission rules:

# Allow only specific skills
Skill(commit)
Skill(review-pr *)

# Deny specific skills
Skill(deploy *)

Permission syntax: Skill(name) for exact match, Skill(name *) for prefix match with any arguments.

Hide individual skills by adding disable-model-invocation: true to their frontmatter. This removes the skill from Claude's context entirely.

Note: The user-invocable field only controls menu visibility, not Skill tool access. Use disable-model-invocation: true to block programmatic invocation.

Source: Official documentation at https://code.claude.com/docs/en/skills.md (section: "Restrict Claude's skill access")

Description Best Practices

Good:

description: Extract text and tables from PDFs, fill forms, merge documents. Use when working with PDF files or when user mentions PDFs, forms, extraction.

Bad:

description: Helps with documents

Template:

[Action 1], [Action 2], [Action 3]. Use when [situation 1], [situation 2],
or when the user mentions [keywords].

Examples

Simple Skill

---
description: Generate conventional commit messages from git diffs. Use when writing commits.
---

# Commit Messages

1. Run `git diff --staged`
2. Determine type (feat, fix, docs)
3. Write message under 50 chars
4. Use imperative mood

Tool-Restricted

---
description: Read files without changes. Use for code review.
allowed-tools: Read, Grep, Glob
---

# Safe Reader

ONLY read files. Never modify.

Forked Context

---
description: Thorough codebase research. Use for complex investigations.
context: fork
agent: Explore
---

# Deep Research

Runs in isolated context to avoid polluting main conversation.

With Hooks

---
description: Audit-logged file operations.
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/security-check.sh"
          once: true
  PostToolUse:
    - matcher: "Write|Edit"
      hooks:
        - type: command
          command: "./scripts/lint.sh"
---

# Secure Operations

All modifications logged and validated.

Hidden from Users

---
description: Auto-apply code review standards.
user-invocable: false
---

Claude invokes this, users don't see it in menu.

User-Only (No Auto-Invoke)

---
description: Deploy to production.
disable-model-invocation: true
---

Only runs when user types `/deploy-production`.

Skills vs Other Features

Note: Skills and slash commands are now the same system. Files in .claude/commands/ still work but skills are recommended.

Installation

Via Plugins: Skills can be bundled in plugins. See ../claude-plugins-reference-2026/SKILL.md for plugin creation and distribution.

Marketplace:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

Manual: Copy to ~/.claude/skills/ or .claude/skills/

Hot Reload: Changes apply immediately without restart.

Recent Updates (2.1+)

• Dynamic token budget — 2% of context window (fallback 16,000 chars) for skill metadata
• Skills from --add-dir — Loaded with live change detection
• Unified skills and commands — .claude/commands/ files now work as skills, skills recommended
• Dynamic context injection - !`command` syntax for preprocessing shell command output
• argument-hint field - Show autocomplete hints for expected arguments
• Optional name/description - If omitted, uses directory name and first paragraph
• once: true for hooks - Run only once per session
• ${CLAUDE_SESSION_ID} - Session-scoped operations
• context: fork with agent selection
• Hot reload - immediate updates without restart

Troubleshooting

Skill not triggering? Check the description includes keywords users would naturally say; verify the skill appears in "What skills are available?"; try rephrasing the request; invoke directly with /skill-name if user-invocable.

Skill triggers too often? Make the description more specific; add disable-model-invocation: true if you only want manual invocation.

Claude doesn't see all skills? Run /context to check for excluded-skills warnings. The budget scales at 2% of context (fallback 16,000 chars). Set SLASH_COMMAND_TOOL_CHAR_BUDGET to override.

Related Skills

• ../agentskills/SKILL.md — The portable specification (agentskills.io). Use when creating skills for cross-agent compatibility. Covers the subset of frontmatter fields (name, description, license, compatibility, metadata, allowed-tools) recognized by all 25+ compatible agents.
• ../claude-plugins-reference-2026/SKILL.md — Plugin creation, distribution, and plugin.json schema.
• ../hooks-guide/SKILL.md — Complete hook events, matchers, and configuration. Cross-platform coverage with Node.js CJS and Python guides.
• resources/claude-code-skills-official.md — Authoritative specification from code.claude.com: frontmatter fields, discovery rules, invocation control, budget limits, and bundled skills.
• resources/scheduled-tasks.md — CronCreate, CronList, CronDelete, /loop syntax, cron expressions, plugin integration patterns.
• resources/headless-agent-sdk.md — Running Claude Code programmatically via claude -p. Covers the critical constraint that skills are unavailable in -p mode, all CLI flags, output formats (json/stream-json/schema), automation patterns, and design checklist for plugin authors.
• resources/output-styles.md — Custom output styles, plugin.json outputStyles field, comparison with CLAUDE.md and skills.
• resources/agent-teams.md — TeammateIdle/TaskCompleted hooks, team architecture, display modes, lifecycle.

Sources

• Primary: Claude Code Skills Documentation (accessed 2026-03-01) — full reference at resources/claude-code-skills-official.md
• Standards: Agent Skills Open Standard — see also the ../agentskills/SKILL.md for the full portable specification reference
• Examples: anthropics/skills
• Blog: Anthropic Engineering Blog - Agent Skills
• Experimental: Context fork tool restrictions verified 2026-01-22