The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides instructions for creating prompt-based hooks that interpolate untrusted data into LLM prompts for validation. • Ingestion points: Data enters the agent context via variables like $TOOL_INPUT, $USER_PROMPT, and $TRANSCRIPT_PATH as described in SKILL.md and references/patterns.md. • Boundary markers: Most provided prompt templates lack explicit delimiters or instructions to ignore embedded commands. • Capability inventory: The skill documents capabilities including bash command execution, network requests via curl, and file system operations. • Sanitization: While some bash-based examples show regex validation, the prompt-based hook guidance relies on LLM reasoning without recommending specific sanitization for interpolated content.
[COMMAND_EXECUTION]: The skill includes several bash utility scripts in the scripts/ and examples/ directories. For example, scripts/test-hook.sh is designed to execute other local scripts for testing purposes using subprocess calls. These tools are intended for developer use within the Claude Code environment.
[DATA_EXFILTRATION]: Documentation in references/advanced.md provides patterns for sending data to external services (such as Slack notifications or database logging) using curl and psql. These patterns involve transmitting data to network endpoints, which is a common but sensitive capability for automated workflows.

Hook Development