The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted visual data (whiteboard photos, napkin drawings, sketches) through an image-recognition MCP and converts them into structured code/diagrams (Mermaid, Excalidraw, tldraw).
Ingestion points: SKILL.md (Phase 1, Step 1) allows ingestion of whiteboard photos, napkin drawings, paper wireframes, and other external images.
Boundary markers: The skill includes a 'Security & Guardrails' section requiring mandatory human verification after recognition (Step 6) and a manual Design Readiness Gate (Step 14).
Capability inventory: The skill uses image-recognition to process images and calls output tools like excalidraw, tldraw, and drawio to generate files. It also updates local PRD and SPEC markdown files (Steps 12, 13).
Sanitization: No explicit sanitization of text recognized from images is mentioned before it is interpolated into Mermaid diagrams or PRD/SPEC templates, creating a risk where text in a sketch could be interpreted as instructions by the agent during the structuring phase.

visual-specs