The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill guides the creation of hooks that process data from the environment, creating a surface for indirect prompt injection findings.
Ingestion points: Hooks read platform-provided JSON via stdin, containing command strings and tool inputs, as shown in references/hook-events.md.
Capability inventory: Hooks can execute shell scripts, modify tool inputs, and perform file system operations.
Sanitization: Provided examples use jq to parse JSON fields safely, which is a defensive best practice.
Boundary markers: No specific delimiters or instruction warnings are recommended for the hook logic itself to separate data from instructions.
[DYNAMIC_EXECUTION]: The skill facilitates the generation and execution of shell scripts and LLM-based prompt hooks at runtime.
Scripts in references/examples/ are designed to be executed by the host platform to perform linting, testing, and formatting.
The documentation describes a feature where hooks can dynamically modify tool inputs before they are executed.
[COMMAND_EXECUTION]: The skill includes numerous templates and examples of scripts that execute system commands for project maintenance, such as git push, npm test, and prettier.

claude-code-hook-development