The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a specific, complex ffmpeg command in SKILL.md to stack video files and apply an overlay. While the filenames used (a.mp4, b.mp4, c.mp4, overlay.png) are static in the example, the agent is instructed to run this shell command to produce the final output.
[INDIRECT_PROMPT_INJECTION]: The skill's core function is to ingest "the brief" (user input) and process HTML code generated by three external AI models. This creates a surface where malicious instructions or scripts could be embedded in the model-generated HTML.
Ingestion points: The "brief" provided by the user and the index.html files generated by the compared models (Step 1 and Step 2).
Boundary markers: The skill provides a prompt block for the models ("Deliver ONE self-contained index.html...") which includes a directive: "No external requests." This serves as a boundary marker to restrict the models' behavior.
Capability inventory: The agent uses ffmpeg (Step 3) and a "shared renderer" (Step 2) to execute and export the HTML content.
Sanitization: The skill relies on the prompt instructions to the models to prevent external requests but does not specify automated sanitization or sandboxing for the rendering process itself.

model-showdown