The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages hooks that process potentially untrusted data from tool inputs and user prompts.\n
Ingestion points: Event schemas in references/input-output-schemas.md include tool_input and prompt.\n
Boundary markers: SKILL.md recommends using jq for validation and keeping blocking rules selective to avoid accidental interruption of work.\n
Capability inventory: Handlers described in references/claude-code.md, references/codex.md, and references/cursor.md allow execution of shell commands and prompt-based reasoning in response to agent events.\n
Sanitization: The skill recommends using trusted absolute paths (e.g., $CLAUDE_PROJECT_DIR) and setting execution timeouts for external commands.\n- [DYNAMIC_EXECUTION]: The skill is primarily designed to configure and manage lifecycle hooks, which are commands or prompts executed dynamically by the agent platform at runtime based on defined matchers and events.\n- [COMMAND_EXECUTION]: The skill provides examples of hooks that run shell commands like jq and prettier for logging tool usage and formatting code.\n- [SAFE]: External documentation links in references/research-sources.md target official domains and repositories for Claude Code, OpenAI Codex, and Cursor, including code.claude.com, github.com/openai, and cursor.com.

hooks-manager