The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of creative writing drafts, creating a surface for embedded instructions to influence agent behavior.
Ingestion points: The skill reads user-supplied draft sections, chapters, and documentation files like ./docs/voice-profile.md and ./docs/structure.md (Step 1).
Boundary markers: There are no explicit instructions, delimiters, or 'ignore embedded instructions' warnings provided to help the agent distinguish between creative content and potential malicious instructions.
Capability inventory: The skill utilizes Read, Grep, Glob, Write, and Edit tools, allowing it to modify the file system based on its analysis of untrusted input.
Sanitization: The instructions lack sanitization or validation steps to filter out or escape instruction-like text within the drafts before the agent processes them.

cw-critique