The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process bibliographic records (titles, abstracts, and metadata) from external academic databases such as PubMed, Scopus, and Embase, which could contain malicious instructions designed to influence the agent during the screening process.
Ingestion points: Data is read from user-provided RIS, CSV, and BibTeX files in scripts/dedupe_records.py and formatted for subsequent human or AI-assisted screening.
Boundary markers: The scripts and instructions do not define specific boundary markers or 'ignore' instructions to encapsulate the untrusted bibliographic content when it is presented to the agent for analysis.
Capability inventory: The skill includes scripts capable of reading from and writing to the file system (scripts/dedupe_records.py, scripts/prisma_flow.py).
Sanitization: scripts/dedupe_records.py employs NFKD normalization and ASCII encoding/decoding with error ignoring to strip non-standard characters from titles, providing a layer of protection against certain obfuscation techniques, though it does not explicitly filter for natural language instructions in abstracts.

systematic-review-prisma