The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill retrieves study metadata and author affiliations from PubMed's official E-Utilities API (ncbi.nlm.nih.gov) via scripts/cohort_overlap_check.py. This operation targets a well-known scientific service for legitimate metadata enrichment.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests and processes untrusted data from research papers.
Ingestion points: Untrusted content enters the context via research papers (Markdown full-text) scanned by scripts/extract_assist.py and screening decision logs (TSV/CSV) processed by scripts/screening_reconcile.py.
Boundary markers: The skill instructions and scripts explicitly label AI suggestions as AI_SUGGESTED and require a mandatory human confirmation pass before any data is moved into the synthesis pool.
Capability inventory: The skill utilizes R scripts for statistical synthesis and Pandoc for artifact generation.
Sanitization: Data is processed using specific regex patterns, and the mandatory human-verification gate effectively sanitizes the input by ensuring clinical and methodological accuracy before execution.
[COMMAND_EXECUTION]: The skill executes shell commands to run statistical analyses in R and to convert documents using Pandoc. These commands are integral to the meta-analysis workflow and are executed on data that has cleared multiple internal consistency gates.

meta-analysis