The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data (research papers) which could potentially contain malicious instructions targeting the LLM.
Ingestion points: scripts/paper_index.py reads content from files in a user-provided directory (papers/).
Boundary markers: SKILL.md includes grounding instructions for the agent (e.g., "Answer only from retrieved text", "Do not fill gaps from memory").
Capability inventory: The skill is restricted to file reading, local SQLite indexing, and querying; it lacks network access or sensitive file write capabilities.
Sanitization: The Python script implements a query tokenizer (fts_query) and uses parameterized SQL queries to prevent technical injection into the database engine.
[SAFE]: No malicious patterns, obfuscation, or unauthorized data access were detected. The skill operates locally on the user's filesystem and uses standard libraries for its intended purpose.

paper-corpus-rag