The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, including research papers (extracted PDF text) and BibTeX reference files. This creates a surface for indirect prompt injection where instructions could theoretically be embedded in academic content.
Ingestion points: Full-text paper extraction (pdfs/*.md), reference metadata (manuscript/_src/refs.bib), and statistical data (analysis/tables/*.csv).
Boundary markers: The skill uses strict section boundaries and IMRAD structure models to compartmentalize processed data.
Capability inventory: Executes local Python scripts for structural verification and interacts with sibling skills for search and reference management.
Sanitization: Implements deterministic regex-based checks and numerical audits to verify data consistency against primary sources.
[SAFE]: The skill makes frequent network references to well-known technology and academic services (e.g., Springer, Elsevier, Nature, Wiley, OUP, and BMJ) to retrieve journal profiles and reporting guidelines. These are established domains and do not contribute to risk.
[SAFE]: All executable components are local Python scripts (scripts/check_placeholders.py, scripts/build_title_page_affiliations.py, etc.) designed to perform deterministic linting and formatting tasks without arbitrary command execution risk.

write-paper