The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text from research paper metadata (titles and abstracts) stored in papers/papers_dedup.jsonl. If these records contain maliciously crafted instructions, they could potentially influence the agent's behavior during subsequent evidence-first writing or synthesis stages where these mappings are utilized.
Ingestion points: The skill reads from papers/core_set.csv, outline/outline.yml, papers/papers_dedup.jsonl, GOAL.md, and queries.md within the workspace.
Boundary markers: No explicit delimiters or boundary markers (e.g., XML tags or clear "ignore instructions" warnings) are applied to the text data before it is processed.
Capability inventory: The scripts/run.py script has the capability to write TSV and Markdown files to the outline/ directory in the workspace.
Sanitization: While the script tokenizes and filters text for scoring purposes, it does not perform sanitization or filtering to detect or neutralize embedded natural language instructions within the paper abstracts.

section-mapper