The Agent Skills Directory

PROMPT_INJECTION (LOW): Indirect Prompt Injection Surface. The skill derives file modifications from untrusted conversation data. * Ingestion points: Processes conversation context, invocation messages, and external documentation (Step 1 and 2) to identify errors. * Boundary markers: Lacks explicit delimiters or instructions to prevent the agent from obeying instructions embedded within the conversation history being analyzed. * Capability inventory: Utilizes the Edit tool for writing to the filesystem and Bash(git:*) for repo management, allowing it to persist changes to the agent's behavior. * Sanitization: Includes a mandatory human review and approval step (Step 5) which acts as a primary defense against unauthorized or malicious modifications.

heal-skill