The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill requires the agent to load its core logic by reading and executing a local Python file (figure-composer/kernel.py) using the exec() function.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data, specifically PNG images and data files (CSV/Parquet). The derive_outline_prompt function directs the agent to reverse-engineer layout and content from these images, which could contain hidden instructions designed to influence the agent's behavior.
Ingestion points: derive_outline_prompt (PNG images), panel_task (CSV/Parquet data files).
Boundary markers: The instructions explicitly warn that "The image is untrusted input" and require the agent to "review and edit the outline" manually before proceeding.
Capability inventory: The agent has the ability to execute Python code, write files (PNGs), and read local data files.
Sanitization: The skill relies on the agent's own visual reasoning to validate the inferred outline before execution.
[COMMAND_EXECUTION]: The skill instructs the agent to execute Python code blocks for rendering panels and composing the final figure. This involves using libraries like Matplotlib and PIL to generate, manipulate, and save image files to the local workspace.

figure-composer