The Agent Skills Directory

[SAFE]: No malicious code patterns, such as unauthorized network access, data exfiltration, or persistence mechanisms, were detected in the skill's scripts or instructions.
[PROMPT_INJECTION]: The skill analyzes untrusted manuscript data (abstracts and figure captions), which constitutes an indirect prompt injection attack surface. This is addressed with the following security considerations: 1. Ingestion points: Untrusted data is processed through the derive_paper_brief function in kernel.py. 2. Boundary markers: The LLM prompts utilize Markdown headers (e.g., ## Abstract) to isolate user-provided content from instructions. 3. Capability inventory: The skill is restricted to narrative analysis and does not possess capabilities for file modification, network communication, or code execution. 4. Sanitization/Verification: The skill documentation and code explicitly mandate a human-in-the-loop review process, requiring the user to confirm the accuracy of derived briefs before they are utilized in subsequent analysis steps.

paper-narrative