The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes manuscript data from output/PAPER.md, which is a potential vector for indirect prompt injection if the manuscript contains malicious instructions designed to influence downstream AI tasks.
Ingestion points: scripts/run.py ingests the full content of output/PAPER.md using the read_text method.
Boundary markers: The script lacks clear delimiters or protective instructions to prevent the AI from following commands embedded in the manuscript content.
Capability inventory: The script includes functions for reading from and writing to the file system within the workspace (write_jsonl, write_text) via imports from the tooling module.
Sanitization: There is no significant sanitization or filtering of the manuscript text before it is processed into claim candidates, beyond basic quote and whitespace stripping.

claims-extractor