The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill utilizes bundled Python scripts (prepare_paper.py and audit_paper_card.py) to process input and audit the generated output. The instructions in SKILL.md explicitly forbid the agent from writing inline Python or using replacement scripts, requiring the use of these fixed, bundled tools only. This approach significantly reduces the risk of arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The prepare_paper.py script suggests the installation of pymupdf (PyMuPDF) if it is missing from the environment. PyMuPDF is a well-known and reputable library for PDF processing. No other external dependencies or remote code downloads were identified.
[COMMAND_EXECUTION]: The regression tests located in tests/test_prepare_paper.py use subprocess.run to verify the CLI functionality of the preparation script. These calls are confined to testing the skill's own logic within a controlled environment and do not involve shell execution or the processing of untrusted system commands.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from research papers, which is a potential surface for indirect prompt injection. However, it implements comprehensive mitigations: a mandatory fixed output schema (Sections 01-16), strict provenance labeling ([Paper], [Analysis], [External]), and a dedicated auditor script that verifies that all claims are grounded in the source text and that no out-of-scope content (like quizzes or social media copy) is generated.
[DATA_EXFILTRATION]: The skill does not contain any network operations or hardcoded credentials. It focuses on local file processing (reading PDFs/JSON and writing Markdown/JSON) within the user's workspace.

nature-paper-card