Claude Skills Overview 2026

Claude Code Skills System - Complete Reference (January 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.

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. Skills are the recommended approach as they support additional features like supporting files and advanced frontmatter options.

SKILL.md Complete Format

---
name: skill-identifier
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.

[!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 is a capability scoping mechanism. The allowed-tools field is not an automatic approval mechanism. Making a tool available does not imply permission to use it; approval and availability are distinct concerns handled by the runtime.

The allowed-tools field exists primarily to scope the tool surface exposed to the skill, reducing prompt and context size by including only the tool definitions the skill may need.

Skill Tokenomics

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

Budget Constraints

YAML Multiline Bug

Do NOT use YAML multiline indicators (>-, |, |-) for descriptions. Claude Code's skill indexer does not parse them correctly - the description appears as ">-" instead of actual content.

# WRONG - will show ">-" as description
description: 'This is a multiline description that breaks.  # WRONG - same problem'
description: 'This breaks too.  # CORRECT - single quoted string'
description: 'This works correctly. Use single quotes for descriptions with special characters or keep on one line.'

Truncation Behavior

When total skill metadata exceeds ~15,000 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 increase 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-thinking` skill
- For delegation: use `/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:

---
name: migrate-component
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

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.

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. See ./claude-hooks-reference-2026/SKILL.md 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, see ./claude-hooks-reference-2026/SKILL.md.

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.

WARNING: context: fork only makes sense for skills with explicit instructions. If your skill contains guidelines like "use these API conventions" without a task, the subagent receives the guidelines but no actionable prompt, and returns 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 Task 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.

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.

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

---
name: commit-messages
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

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

# Safe Reader

ONLY read files. Never modify.

Forked Context

---
name: deep-research
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

---
name: secure-ops
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

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

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

User-Only (No Auto-Invoke)

---
name: deploy-production
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+)

• 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
• 15,000 character budget for skill metadata
• context: fork with agent selection
• Hot reload - immediate updates without restart

Sources

• Primary: Claude Code Skills Documentation (accessed 2026-01-28)
• Standards: Agent Skills Open Standard
• Examples: anthropics/skills
• Blog: Anthropic Engineering Blog - Agent Skills
• Experimental: Context fork tool restrictions verified 2026-01-22