The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads documentation files from https://code.claude.com/docs/en/ using curl. These sources are official product documentation and the skill is restricted via allowed-tools to only fetch from this specific domain.\n- [COMMAND_EXECUTION]: The skill uses Bash to execute commands like curl, python3 -m json.tool, and strings. These tools are used for legitimate configuration auditing tasks, such as fetching docs, validating JSON syntax, and searching the claude binary for configuration keys. Access is restricted to specific command patterns in the YAML frontmatter.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external markdown files and local configuration files. While this presents a surface for indirect prompt injection, the skill targets official documentation and local files that the user owns. The instructions emphasize human-in-the-loop confirmation before applying any changes.\n
Ingestion points: Documentation from code.claude.com and local files like ~/.claude/settings.json and CLAUDE.md (specified in SKILL.md).\n
Boundary markers: The agent is instructed to use the files as 'ground truth' for cross-referencing specific configuration keys.\n
Capability inventory: Shell command execution via Bash (restricted to curl, python3, strings) and file reading/writing (restricted to ~/.claude/).\n
Sanitization: Edits to configuration files are validated using python3 -m json.tool before being finalized.

audit-claude-settings