The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill aggregates data from numerous external sources, including news articles, scientific papers, clinical trial results, and patent databases. This creates a surface for indirect prompt injection if those external sources contain adversarial instructions intended to influence the agent's analysis or report generation.
Ingestion points: External data is ingested in multiple steps defined in references/legacy-report-spec.md (Steps 5, 6, 7, 9, 9.5, and 10).
Boundary markers: The instructions do not explicitly specify delimiters or sanitization protocols for the external content before it is interpolated into the final HTML report.
Capability inventory: The skill has the capability to write files to the session output directory and perform complex data synthesis across 11 different dimensions.
Sanitization: While no specific string sanitization is mentioned, the skill mandates a 'Data Cross-Validation' step (Chapter 11) to check consistency across multiple sources, which serves as a significant mitigation against poisoned data.
[EXTERNAL_DOWNLOADS]: The skill includes a requirement to fetch the ECharts library from the jsDelivr CDN (https://cdn.jsdelivr.net/npm/echarts@5/dist/echarts.min.js) to provide data visualizations within the generated HTML report. jsDelivr is a well-known and trusted content delivery service.
[COMMAND_EXECUTION]: The skill uses a files tool to manage report generation, involving shell-like operations such as begin_write, append, and finish_write to save the final assessment to @session/outputs/. This is standard functionality for producing artifacts.

drug-lifecycle-evaluation