The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of prompt-based hooks that analyze potentially untrusted file content, creating an attack surface for indirect prompt injection.
Ingestion points: Hook input received via stdin during tool execution events.
Boundary markers: The skill recommends using JSON schema validation to constrain hook outputs.
Capability inventory: Hooks can trigger shell commands, modify files, and block tool operations using specific exit codes.
Sanitization: The documentation provides extensive advice on validating JSON fields and sanitizing file paths using jq and bash patterns.
[COMMAND_EXECUTION]: The skill provides templates for executing shell commands and local CLI utilities within a hook environment.
Evidence: Templates in references/code-templates.md demonstrate the use of prettier, black, gofmt, and git.
Safety context: The skill explicitly instructs on the use of variable quoting and absolute paths to prevent common execution vulnerabilities.
[DYNAMIC_EXECUTION]: The skill provides a pattern for using the eval command to process multiple JSON fields simultaneously.
Evidence: The references/quick-reference.md file contains a snippet using eval with jq's @sh filter.
Safety context: The use of the @sh filter ensures that values are properly escaped for shell evaluation, mitigating injection risks.

claude-hook-writer