The Agent Skills Directory

[SAFE]: The skill exhibits no malicious patterns. It is well-structured with automated tests and validation scripts that prevent common errors such as checklist fabrication or transcription drift.
[INDIRECT_PROMPT_INJECTION]: The skill operates by reading user-supplied manuscripts (manuscript.md) and screening TSV files (round1.tsv, etc.), which constitutes an ingestion surface for untrusted data.
Ingestion points: Manuscript and TSV files are read and processed by audit scripts including check_prisma_figure.py and prisma_cascade_check.py.
Boundary markers: The workflow instructions guide the agent to scan the manuscript for specific sections and items, relying on the model's task-specific instructions rather than explicit boundary delimiters for the audited content.
Capability inventory: The skill can execute internal Python scripts (python3) and write diagnostic artifacts to the qc/ directory.
Sanitization: Audit scripts use regular expressions and structured parsing (CSV/TSV) to extract data, which provides inherent structural validation of the processed content.
[COMMAND_EXECUTION]: The workflow involves executing several internal Python utility scripts and standard shell commands like grep. All commands are static, scoped to the skill's local directory and the project's manuscript files, and do not incorporate unvalidated user input into shell execution paths.
[EXTERNAL_DOWNLOADS]: The maintenance script refresh_from_europepmc.py (located in the tests/ directory) contains functionality to fetch XML data from the Europe PMC API (ebi.ac.uk). Europe PMC is a well-known scientific service; the script is used for developer-led updates of the checklists and is not part of the active manuscript auditing workflow.

check-reporting