The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied review text provided in review_scoring_table.csv and interpolates these strings into a generated report.md as evidence quotes. While this creates a surface for indirect prompt injection (where instructions embedded in reviews might influence the agent reading the report), the skill includes several mitigations:
Ingestion points: Raw text is ingested via the review_text column in review_scoring_table.csv (handled in stp_runner/io.py).
Boundary markers: The instructions in SKILL.md and references/01-router-and-gates.md enforce a strict separation where scripts only process already-scored artifacts, reducing the chance of an agent being tricked into executing raw review content as instructions.
Capability inventory: The skill performs file writing (CSV, JSON, Markdown) and image generation (PNG), but does not have network access or dynamic code execution capabilities.
Sanitization: stp_runner/io.py performs extensive validation on the structure and data types of inputs to ensure they meet the canonical contract before processing.

review-mining-stp