The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for managing the file system to organize the skill repository.
Evidence: Includes instructions for using mkdir -p, ln -s, rm, and cat to create directory structures and manage symlinks between the central repository and project directories.
[PERSISTENCE]: Recommends persisting a required environment variable in the user's shell configuration.
Evidence: Instructions guide the user to add export CLAUDE_METADATA="..." to ~/.zshrc or ~/.bashrc to ensure the skill repository path is available across sessions.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of new AI agent skills and commands, which naturally involves a surface for indirect instructions.
Evidence: Provides templates for SKILL.md files and command markdown files that the agent is instructed to write, which will influence future AI behavior. The skill includes best practices for manual review to mitigate risks.
[DYNAMIC_EXECUTION]: Describes workflows for generating scripts at runtime based on templates.
Evidence: Explains how to use sed to perform string substitution on shell script templates to generate project-specific logic.
[REMOTE_CODE_EXECUTION]: Provides guidance on adopting external commands from GitHub repositories.
Evidence: Documents the use of the gh CLI to fetch file contents from remote repositories for local adaptation and emphasizes that content should be reviewed and tailored rather than executed verbatim.

claude-skill-management