A Markdown File Just Replaced Your Most Expensive Design Meeting. (Google Stitch)

Three new “creative primitives” are converging on the command line—turning design, video, and 3D scenes into editable code—so teams can iterate faster, automate production, and reduce the handoff bottlenecks that have long slowed product work.

Google’s updated Stitch is the clearest example of design shifting from a canvas to an agent-driven pipeline. Instead of producing wireframes, Stitch generates high-fidelity, finished-looking UI across multiple screens at once. Users describe an app in natural language (with voice support), and Stitch outputs complete layouts with real typography, spacing, and component hierarchy. The agent keeps project context—so edits can propagate coherently across the whole set of screens rather than treating each page as a separate task. Stitch also branches and compares design directions, enabling rapid versioning and side-by-side exploration that used to be gated by the time cost of producing polished mockups. The most consequential change is export: Stitch produces a design.markdown file that captures the evolving design system (colors, typography, spacing rules, component patterns) in a format agents can read. That file can be consumed by coding agents via MCP, eliminating the classic “export to Figma, then hand off” workflow.

Remotion extends the same idea from design to video. A React framework that treats video as code, Remotion gained an agent skill for Claude Code, letting users describe a video in plain English and having Claude generate the React components that define every frame—text animations, motion graphics, captions, transitions, and even data visualizations. Remotion then renders the result into an MP4 locally. The key distinction is programmability: unlike prompt-to-pixel tools such as Sora or Runway, Remotion outputs editable code, making rerenders cheap and updates straightforward when inputs change. The practical sweet spot is “clean” motion graphics—product demos, feature announcements, data stories, and social clips—where iteration speed matters more than ultra-complex animation.

Blender MCP brings the same command-line logic to 3D. Blender’s Python API is notoriously complex, but MCP makes it accessible through natural-language tool manipulation. Users can request a scene—like a beach with lighting and palm trees—and watch objects, materials, and lighting assemble in real time. The workflow can pull in external assets via Polyhaven and SketchFab, and can even incorporate text-to-3D generation through Hyper 3D. The result is a faster path from concept to proof of concept for architecture visualization, walkthroughs, and game-adjacent prototyping, without requiring months of Blender training.

Across all three tools, MCP is positioned as the universal connector—an “USB plug” for AI—so products can expose capabilities directly to agents at the terminal. The broader implication is structural: the old 2010s triangle of product, design, and engineering often failed in practice because teams discovered buildability late and spent too long on sequential pixel work. Command-line design collapses that loop by making outputs inherently buildable and by enabling rapid iteration through language. The speaker’s bottom line is not that designers disappear; it’s that the floor for producing “good enough” visuals drops dramatically, shifting the scarce skill toward judgment, intent, and polishing. The winners will be those who can articulate what should be expressed visually—and then refine it—while agents handle the mechanics and the scheduling of repeatable creative pipelines.