SkillsIntermediateClaude Code
A skill with context: fork keeps your session clean
If a skill produces a lot of output (analysis, exploration), run it in an isolated subagent and only the summary lands in your session.
In the SKILL.md frontmatter, context: fork runs the skill in an isolated subagent. With allowed-tools you limit which tools it can use while it runs (for example, read-only). argument-hint tells whoever invokes it which parameter is missing. Use it for one-off tasks; whatever must always hold goes in CLAUDE.md.
Example
--- name: analyze-structure description: Analyzes the dependencies and patterns of a directory. Use it when asked to map a module. context: fork allowed-tools: Read, Grep, Glob argument-hint: Path of the directory to analyze --- Analyze the given directory and return a short report of dependencies and patterns.