The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes structured data and natural language instructions from an external annotation widget via the get_renoise_whiteboard_revision_intent tool. This content is used to build prompts for image and video generation tasks.
Ingestion points: The skill reads input from the RevisionIntent.instruction field and visual annotation data in SKILL.md (Handle an annotation submission).
Capability inventory: The skill uses create_tasks for remote generation and upload_files for material management.
Mitigation: A strong human-in-the-loop constraint is enforced. The skill must present the final prompt, model selection, and estimated cost to the user in a fenced block, requiring explicit confirmation before any task execution.
Sanitization: The skill relies on structured annotationBindings to map user marks to authoritative sources rather than interpreting raw visual data alone.
[COMMAND_EXECUTION]: The skill interacts with the local filesystem to initialize the whiteboard session.
Evidence: The render_renoise_whiteboard_widget tool accepts an absolute projectDir path.
Security control: The skill includes explicit instructions in the "Safety rules" section to never substitute the project root or home directory for the user-approved directory and to only resolve directories within the current project scope.

canvas