# Agent-Readable Design Files Beat Opaque Canvases for AI Output ## Claim The structure a design tool exposes to an AI agent sets the ceiling on how reusable and build-ready the agent's generated output can be, independent of the model driving it. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds when the agent both reads and writes the design file format — the claim is about round-trip fidelity, not one-way rendering. - Weakens for simple, single-state components where any reasonably capable tool converges on similar output regardless of underlying data structure. - Vendor-sourced (Penpot's own promotional comparison against Figma), with no shown artifacts or independent methodology — treat as a hypothesis worth testing, not a settled finding, until reproduced outside the vendor's own demo. ## Evidence Penpot's "real estate card test" clip: an identical complex prompt handed to Penpot's MCP server, Figma's MCP server, and unnamed competitors allegedly produced "broken layouts and elements you can't reuse" from the competitors, while Penpot's "design as code" approach — a structured, parseable underlying file format rather than an opaque canvas — gave the agent "the rich context they need to read and write design files you can actually build on." ## Links - [[AI & the Future of Knowledge Work MOC]] — the structure of your artifacts sets the ceiling on what an agent can do with them - [[3. Resources/Permanent Notes/Agent-Readable Data Is Becoming the Default Plugin Integration Surface]] — because extends; the WordPress-plugin instance of the same underlying pattern — exposing structured data to an agent, not just a UI to a human, is becoming table stakes across tool categories, not only design tools. - [[3. Resources/Permanent Notes/Tool-Interface Choice Shapes Agent Architecture More Than Model Choice]] — because extends; both claims locate agent-output quality in the interface/data layer rather than the model, from different source domains (CLI/MCP basics vs. design tooling). - [[3. Resources/MOCs/Visual Thinking MOC]] — because applies-in; whether a design tool's data model is agent-legible is now a first-class visual-tooling design decision, not just a nice-to-have. ## Tests / falsifiers - Independently reproduce the real-estate-card prompt across Penpot MCP, Figma MCP, and a third MCP-enabled design tool, capturing and comparing actual output artifacts — if Figma's output is comparably clean and reusable, the vendor's comparative claim is refuted even if the general structure-matters mechanism survives. - Test the same prompt on a simple, single-state component — if all tools converge, that confirms the scope condition that the effect concentrates in complex, multi-state prompts. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/Penpot MCP vs Figma McP The Real Estate Card Test]] — the identical-prompt comparison and the "design as code" thesis line