Claude Code Skill

Display name for the skill using lowercase letters, numbers, and hyphens only (max 64 characters). This becomes the /slash-command name (e.g. name: review-code creates /review-code). If omitted, defaults to the directory or file name. Plugin skills are namespaced as /plugin-name:skill-name.

Describes what the skill does and when to use it. Claude uses this to decide when to automatically load the skill, so write it as a triggering guide — include concrete examples of user requests that should activate it (e.g. "Use when the user asks to review code"). Skill descriptions are always loaded into Claude's context so it knows they exist, but full skill content only loads when invoked. If omitted, defaults to the first paragraph of the markdown body. Note: descriptions share a context budget (2% of the context window, minimum 16KB). Override with the SLASH_COMMAND_TOOL_CHAR_BUDGET environment variable. Run /context to check for a warning about excluded skills.

Hint shown during / autocomplete to indicate expected arguments. Purely cosmetic — does not enforce argument structure. Arguments are accessed via $ARGUMENTS (all args), $ARGUMENTS[N] or $N (positional, 0-based) in the markdown body. If $ARGUMENTS is not present in the content, arguments are appended as ARGUMENTS: <value>.

Which Claude model to use when this skill is active. If omitted, inherits from the current session.

Tools Claude can use without asking permission when this skill is active. Built-in tools: Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, Task, Skill. Supports patterns like Bash(git *) to allow only specific subcommands, Bash(npm run *) to scope shell access, or Skill(commit) to allow invoking a specific skill. MCP tools use the pattern mcp__<server>__<tool>. If omitted, all tools are available with normal permission prompts. This field only grants auto-approval — it does not override denials set in the user's permission settings.

When true, prevents Claude from automatically loading this skill based on its description — the skill's description is not included in Claude's context, so Claude will never know it exists unless the user invokes it. The skill can only be triggered manually via its /slash-command. Use for manual-only workflows like /deploy or /commit where automatic triggering would be disruptive. See control who invokes a skill.

When false, hides the skill from the / autocomplete menu so users cannot invoke it directly via slash command. The skill can still be loaded automatically by Claude based on its description. Use for background knowledge that Claude should apply silently — for example, coding conventions or project-specific patterns. Note: this only controls menu visibility, not Skill tool access. Use disable-model-invocation: true to block programmatic invocation. See control who invokes a skill.

Set to fork to run the skill in an isolated subagent context instead of inline. When forked, the skill's markdown body becomes the subagent's task prompt (not reference material). The subagent runs in a separate conversation fork with its own context window — it does not have access to the main conversation history. Tool calls and output from the forked skill do not appear in the main session. Use for heavy workflows (running test suites, large refactors) where inline execution would clutter the conversation. See run skills in a subagent.

Which subagent type to use when context is set to fork. Built-in types: Explore (read-only, optimized for codebase exploration using Glob, Grep, Read), Plan (read-only, for designing implementation plans), general-purpose (full-capability agent with all tools). Can also be the name of a custom agent defined in .claude/agents/. Ignored if context is not fork. Defaults to general-purpose. See built-in subagents.

Lifecycle hooks scoped to this skill. Hooks defined here only fire when this skill is active and are cleaned up when the skill is unloaded. Each key is an event name and its value is an array of matcher groups. See hooks in skills and agents.