The pattern with product teams is not failure at building — they fail at explaining. A designer articulates a vision in Figma. An engineer interprets it. Product is on a timeline. By the time anything ships, the original intent has passed through three people’s working memory and come out a little blurrier each time. That’s the expensive part. Not the pixels, not the code — the gap between minds.
There’s a term from information theory: lossy compression. It describes formats that throw away data when they encode something. JPEG images are lossy — they discard detail the human eye probably won’t notice, and once gone, that detail can’t be recovered. Most creative handoffs in software teams are also lossy. The design decision lives in a Figma comment. The comment doesn’t make it into the ticket. The ticket doesn’t fully explain the spacing logic. The engineer makes a judgment call, reasonable given what they had.
The week three tools converged — Google’s Stitch (a design tool), the Remotion video framework, and Blender MCP (a 3D modeling interface) — the most interesting detail in all the coverage got buried. It was a markdown file.
The part everyone missed
Stitch generates multiple high-fidelity UI directions from a plain-language description of what you’re building. Remotion renders video from code. Blender MCP lets you describe a 3D scene in conversation and watch it assemble in real time. Each of these is impressive in its own right, and the design-to-minutes story is real and worth paying attention to.
But the key is DESIGN.md: a portable markdown file that captures a project’s design system — its colors, typography, spacing, component patterns. You can export it from Stitch, import it into a new project, or hand it to any MCP-compatible tool. MCP, short for Model Context Protocol, is a standard that lets AI-powered tools pass context to each other, the way a shared drive lets a team work off the same documents. DESIGN.md is readable by humans and machines both.
That might sound mundane. It’s not.
Why lossless matters
The current workflow has lossy compression at every step. Figma is the source of truth, until it isn’t — until the developer needs to make a call the design didn’t anticipate, or the design updates and the code doesn’t follow. DESIGN.md doesn’t eliminate those gaps entirely, and that is not the claim being made here. But it’s the first specification format designed from the ground up to be readable by AI agents, exportable across tools, and usable as a constraint file that keeps multiple agents aligned.
The pipeline this enables is worth sketching out: a PM writes a brief in plain language. A design agent generates UI directions with DESIGN.md as a constraint. A coding agent reads the same file and builds against the same colors, spacing, and component patterns — not an interpretation of them. A video agent generates launch content that matches the visual identity. One context file, multiple creative disciplines, no handoff meeting in between.
Those tools exist now, and they compose through MCP.
What doesn’t compress
A note of caution: this should not be oversold. The outputs range from “polished enough for a prototype” to “you’d be embarrassed to show a client.” Blender MCP will assemble a beach scene; it won’t sculpt a character face. Remotion renders clean motion graphics; complex multi-element animations still break down.
But here’s what the framing often misses: the expensive part of creative work was never polishing the finished artifact. It was reaching something worth polishing. Getting five UI directions in twenty minutes instead of two weeks each removes the premature convergence gate — you can stay in option-generation mode longer, try the approach that felt too risky to prototype.
What doesn’t compress is judgment. Taste. The experience that tells you the output is actually right for this audience, this moment, this problem. Designers aren’t being replaced; they’re at the command line now, iterating at the speed of language instead of the speed of drag-and-drop. Pixel-pushing was the bottleneck. Experience judgment was always the point.
Where this goes
DESIGN.md, or something very much like it, is likely to become standard practice in the same way a README or a package file is standard practice — a document created on day one and maintained throughout. The coordination cost it’s targeting is one of the oldest in software: keeping multiple people’s understanding of the same thing from diverging over time.
We haven’t solved that problem. But we’ve finally got a format that at least doesn’t lose data by design.
Text summarized and optimized using Anthropic’s models and reviewed by a human.