The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process data from various tool operations (Bash commands, file writes, tool outputs). This creates a surface for indirect prompt injection where malicious instructions embedded in the data being processed (e.g., a file being read or a command output) could attempt to influence the logic of the hooks or the agent's next actions.
Ingestion points: Hook input via fs.readFileSync(0, 'utf-8') (stdin) in pre-hook-template.js, post-hook-template.js, and hook-template-generator.js.
Boundary markers: None explicitly enforced in templates to separate data from instructions; the skill relies on the user to implement regex or logic checks.
Capability inventory: Templates demonstrate the ability to block/approve commands, write to log files (fs.appendFileSync), and execute system commands via the registered hooks.
Sanitization: The skill provides a 'Security Considerations' section advising validation of all input, but the templates themselves do not include default sanitization functions beyond basic JSON parsing.
[COMMAND_EXECUTION]: The skill frequently references and generates scripts that execute commands. While these are intended for legitimate automation and validation, the templates include examples of regex-based command blocking (e.g., rm -rf, sudo) which may be bypassed by sophisticated inputs if not carefully implemented.
[SAFE]: The skill includes extensive 'Anti-Patterns' and 'Shell Script Best Practices' sections that correctly identify and mitigate common security and portability issues such as improper variable quoting, silent failures, and dangerous flags.

hook-creator