The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted research paper data, which can serve as a vector for indirect prompt injection. \n
Ingestion points: Paper metadata and full text are ingested from papers/core_set.csv, papers/fulltext_index.jsonl, and papers/fulltext/*.txt files via the scripts/run.py processing script. \n
Boundary markers: None identified; extracted evidence snippets are placed directly into papers/evidence_bank.jsonl without the use of delimiters or 'ignore instructions' warnings. \n
Capability inventory: The skill's primary script scripts/run.py is limited to local file system reads and writes using standard library functions and a local helper module. It does not possess network access or arbitrary command execution capabilities. \n
Sanitization: The extraction logic in scripts/run.py uses regex-based sentence splitting but does not include any sanitization, validation, or filtering of the content to prevent embedded instructions from influencing downstream agent behavior.

paper-notes