The Agent Skills Directory

[COMMAND_EXECUTION]: The documentation provides instructions for testing hooks using command-line tools like npx tsx, cat, and echo. These are standard developer operations for verifying the behavior of local scripts within the .claude/hooks directory.
[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture where hooks ingest user prompts and file contents to trigger agent behavior, creating a potential attack surface for indirect prompt injection.
Ingestion points: The UserPromptSubmit hook processes user prompts, and the PreToolUse hook processes the content of files being edited.
Boundary markers: The documentation examples use visual delimiters (e.g., ━━━━━━━━━━━━━━) to separate injected context from the user prompt.
Capability inventory: The hooks can block critical tools like Edit and Write by exiting with specific codes.
Sanitization: The guides do not explicitly detail sanitization or escaping of the ingested data, which is an inherent risk factor of the documented platform architecture.
[DYNAMIC_EXECUTION]: The hook system described in the guides relies on executing local TypeScript files via npx tsx. This is a core component of the Claude Code architecture and is documented here for legitimate integration and development purposes.

skill-developer