The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted manuscript files from the inputs/ directory and writes them to a canonical artifact, creating a potential vector for instructions to influence downstream agent pipeline steps.\n
Ingestion points: scripts/run.py reads content from file paths identified by find_workspace_text_source (e.g., inputs/manuscript.md, inputs/manuscript.txt, or inputs/manuscript.pdf).\n
Boundary markers: The skill does not implement delimiters or 'ignore' instructions when writing the extracted text to output/PAPER.md.\n
Capability inventory: The scripts/run.py script has capabilities to read from and write to the filesystem within the workspace.\n
Sanitization: No input validation or instruction filtering is performed on the manuscript content before it is processed by downstream tools like claims-extractor.\n- [DYNAMIC_EXECUTION]: The script dynamically modifies the system path to resolve internal dependencies based on the directory structure.\n
Evidence: scripts/run.py traverses parent directories to find an AGENTS.md marker and inserts the resulting path into sys.path to enable the import of the tooling module.

manuscript-ingest