The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted data from user-provided ideas and project files, creating a potential surface for indirect prompt injection.\n- Ingestion points: Processes hunches directly from user chat and "project files" during the claim restatement phase (SKILL.md).\n- Boundary markers: Lacks explicit boundary markers or delimiters to isolate untrusted user data from instructional logic.\n- Capability inventory: The skill is designed to read file contents, invoke external search skills (literature-research), and append data to local record files.\n- Sanitization: No specific instructions are provided for sanitizing or validating the contents of user inputs or project files before they are incorporated into the prompt context.\n- [EXTERNAL_DOWNLOADS]: References a research paper on the arXiv repository. This is a well-known and trusted service for academic preprints.

idea-critique