The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents prompt-based hooks that interpolate untrusted data (such as tool inputs, results, and user prompts) directly into LLM prompts for validation logic. This creates a surface for indirect prompt injection where malicious content in tool outputs could potentially influence the validation process or agent behavior.\n- Ingestion points: $TOOL_INPUT, $TOOL_RESULT, and $USER_PROMPT placeholders in SKILL.md and references/patterns.md.\n- Boundary markers: The provided prompt templates lack explicit delimiters (e.g., triple backticks or XML tags) or specific instructions to treat the interpolated content as untrusted data.\n- Capability inventory: The agent environment allows for file system operations, command execution via bash, and various tool interactions across the development lifecycle.\n- Sanitization: No sanitization, escaping, or strict schema validation mechanisms are described for the content being interpolated into the prompts.

Hook Development