The Agent Skills Directory

[SAFE]: The skill is purely informational, providing documentation and examples for configuring a rule-based monitoring system. It contains no executable scripts or instructions that bypass safety protocols.
[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture where the agent reads instructions from local markdown files in the .claude/ directory and injects them into the context based on event triggers.
Ingestion points: Files matching the pattern .claude/hookify.*.local.md in the project root.
Boundary markers: The system uses YAML frontmatter to define specific event triggers and patterns, limiting when the injected content is shown.
Capability inventory: The system can issue warnings or block tool execution for bash, file, stop, and prompt events.
Sanitization: The skill recommends using specific Python regex patterns to accurately target unwanted behaviors and minimize false positives.
[DATA_EXPOSURE]: The skill references sensitive file paths and patterns (e.g., .env, API_KEY, credentials) but only within the context of teaching users how to create protective rules to guard those assets.
[COMMAND_EXECUTION]: The skill lists potentially dangerous bash commands (e.g., sudo, rm -rf) as examples for users to monitor and restrict, promoting safer agent operations.

writing-hookify-rules