The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses shell commands via uv run and Rscript to execute pipeline automation scripts (e.g., tooling/python/extract_pdf_text.py, 01_setup.R). These commands are the primary mechanism for the skill's functionality.
[EXTERNAL_DOWNLOADS]: The fulltext-manager role is instructed to retrieve research PDF documents from external services like Unpaywall and DOI-based providers. Additionally, the statistician role uses R's renv to initialize and download package dependencies from CRAN/R-project repositories.
[PROMPT_INJECTION]: An indirect prompt injection surface is present in roles that process untrusted external data, specifically the screening-reviewer (bibliography entries) and data-extractor (PDF text). Maliciously crafted content within academic papers could theoretically attempt to influence agent decisions or extraction accuracy.
Ingestion points: 02_search/round-01/dedupe.bib, 04_fulltext/*.pdf, and bibliography data processed in Stage 03.
Boundary markers: None explicitly implemented in the provided role prompts to separate untrusted data from instructions.
Capability inventory: All roles possess the capability to execute shell commands (uv run, Rscript) and perform file system operations within the project directory.
Sanitization: The instructions do not specify sanitization or validation of external text before it is analyzed by the AI model.

ma-agent-teams