The Agent Skills Directory

Indirect Prompt Injection Surface: The skill facilitates the creation of hooks that process tool inputs, results, and user prompts, creating a surface for potential indirect prompt injection. (1) Ingestion points include scripts that read JSON from stdin, such as examples/validate-bash.sh and examples/validate-write.sh. (2) Boundary markers are supported through the use of prompt-based hooks which allow for explicit validation instructions. (3) The capability inventory includes bash command execution, file system access, and documented network patterns. (4) Sanitization guidance is provided through the scripts/hook-linter.sh utility and example scripts that check for path traversal and dangerous command patterns.
Local Script Execution for Testing: The scripts/test-hook.sh utility is designed to execute local bash scripts to verify their logic before deployment. While this is a functional requirement for the skill's purpose, it facilitates local code execution, and users should ensure the integrity of scripts before testing.
Network Integration Examples: Reference documentation in references/advanced.md provides patterns for integrating hooks with external services like Slack or databases using curl and psql. These examples follow secure practices by utilizing environment variables for credentials rather than hardcoding them.
Command Hook Capabilities: The skill allows for the implementation of bash-based hooks to perform deterministic checks. The provided documentation emphasizes secure coding practices, such as quoting all variables and validating tool input fields, to mitigate risks associated with command execution.

hook-development