The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/extract_pdf_figures.py invokes local PDF processing tools such as pdfimages and pdftoppm using subprocess.run. The implementation uses list-based arguments to prevent shell injection and verifies the presence of binaries via shutil.which before execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from academic manuscripts and code repositories, creating a surface for indirect prompt injection.\n
Ingestion points: External content is ingested via references/paper.md (manuscript text) and references/repo.md (repository source code).\n
Boundary markers: The references/missing-info-policy.md file defines a policy for handling incomplete evidence, instructing the agent to use placeholders (e.g., [module_name]) and label inferred claims, which helps prevent the blind execution of embedded instructions.\n
Capability inventory: The skill has the capability to read local files and write extracted images and JSON reports to the local filesystem via the extract_pdf_figures.py script.\n
Sanitization: The skill instructions explicitly forbid the invention of modules or architecture layers not present in the source, serving as a logic-level check against malicious data manipulation.

fig1-analyze