claude-skills_skill

Overview

This skill provides guidance and configuration patterns for creating Claude Code skills with Claude-specific frontmatter like allowed-tools, context modes, argument hints, and model overrides. It documents syntax for tool restrictions, dynamic context injection, and fork vs inherit execution so you can author predictable, safe skills. Use it to design skills that run tools without prompts, isolate execution, or expose concise slash-command arguments.

How this skill works

The skill inspects and documents frontmatter fields that extend base agent skills: allowed-tools, user-invocable, context, agent, model, hooks, and argument-hint. It explains how to declare tool permissions (including Bash patterns and MCP tool names), how to inject dynamic command output into skill text, and how forked contexts delegate to subagents. It also includes testing, troubleshooting tips, and integration patterns for chaining or delegating skills.

When to use it

• When authoring Claude-specific skills that need tool permission control
• When you want an isolated subagent using context: fork for specialized processing
• When you need to expose a concise command picker hint or slash command arguments
• When overriding the model for a skill (haiku, sonnet, opus)
• When you need dynamic environment information injected at load time

Best practices

• List only necessary tools in allowed-tools and prefer least privilege
• Use Bash(...) patterns with wildcards to cover expected command shapes
• Use context: fork for side tasks to avoid polluting the main conversation
• Provide clear argument-hint strings to improve slash-command discoverability
• Test with claude --debug and real data to validate tool permissions and activation rules

Example use cases

• Code review skill that runs Read, Grep, and Bash(git *) to evaluate diffs
• Security-audit skill run in a forked agent with a different model for focused analysis
• PDF processor that injects pdfinfo output at load time for dynamic context
• Internal validator with user-invocable: false to run only via automation hooks
• Master-clone orchestration where the master delegates a specialized task to a forked subagent