The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes a visualization_agent and a VLM Figure Verification Protocol that involve generating and executing code.
Evidence: In agents/visualization_agent.md and references/vlm_figure_verification.md, the skill describes generating Python (matplotlib, seaborn) and R (ggplot2) code to visualize results. The VLM verification step explicitly includes an 'Execute code to render figure image' phase where the generated code is run.
[INDIRECT_PROMPT_INJECTION]: The skill has a large surface for ingesting untrusted data which could influence agent logic.
Ingestion points: agents/revision_coach_agent.md is designed to parse unstructured reviewer comments from any source (email, PDF pastes), and agents/intake_agent.md processes user-supplied past writing samples for style calibration.
Capability inventory: The skill has the ability to generate and execute code, perform file operations, and produce complex reasoning outputs based on these untrusted inputs.
Sanitization: There is no documented mechanism for filtering or sanitizing the content of these external inputs before they are used to influence the drafting and code generation phases.

academic-paper