A production-grade blueprint: how a design system becomes infrastructure AI agents can consume directly, and turns into $1 billion in business results.
Preparing Rocket for the AI era | Config 2026 · Emily
20 min total·Actually worth watching closely: ~13 min·3 must-watch clips
- 0:12 – 5:10Listen
Who Rocket is: scale, mission, and why AI isn't optional
Introduces the scale Rocket operates at — 1 in 6 American homeowners, 800,000 chats and 160 million calls a year — and the context and motivation behind the AI-native transformation.
At that scale AI isn't optional, and the team is already shipping five times faster than two years ago.
Mostly spoken argument and background with no key demo footage — fine to listen to like a podcast.▶ Jump to 0:12Speaker · Emily - 5:10 – 7:13Watch
Semantic tokens and canvas annotations: Nova's foundation
Every color, radius, and motion curve is a semantic token named for why it exists, with canvas annotations capturing the development requirements and interaction states a screenshot can't carry.
Every visual decision traces back to a named decision, which is what makes Nova portable across platforms, themes, and AI agents.
From 310s there's concrete footage of the token naming and annotation system — seeing the examples conveys how granular "semantic" really gets better than hearing it described.▶ Jump to 5:10Speaker · Emily - 7:13 – 12:10Watch
Figma Make live prototype: research anyone can run
A shared AI chat prototype built in Make, a new component wired in with two messages, published to a research panel for a blue vs. green variant test, with findings embedded back on the design file canvas.
Research participants experience something close to the real thing rather than a static screenshot, and no engineer is needed anywhere in the chain.
529s (node editor), 590s (two messages to add the component), and 663s (findings embedded back) are continuous live operation — that's where this section's value sits.▶ Jump to 7:13Speaker · Will Habeck - 12:10 – 14:16Watch
Preventing drift and automating handoff: Check Designs + MCP annotations
Check Designs finds hardcoded values that escaped the tokens and fixes them in one click; a home-grown skill uses the Figma MCP to auto-annotate layers, surfacing missing icons and accessibility questions ahead of handoff.
The details a manual handoff loses most easily get caught automatically, before delivery.
738s and 799s are two tight tool demos — hover-to-locate and auto-annotation only land if you watch the screen.▶ Jump to 12:10Speaker · Will Habeck - 14:16 – 18:06Watch
Agents implementing components: the three-step skill and Playwright verification
The skill enforces three steps — initial draft, adjust against the annotations, replace with design system classes — and feeds in all the context automatically, so one message produces a compliant component; another skill screenshots Storybook with Playwright and compares visual and code against the design.
Context management is the key — you can't hand the agent the whole Figma MCP output at once; the structured, staged approach is what makes the one-shot hold up.
The one-shot generation at 940s and the automatic review pass at 1007s are the talk's peak demos — the comparison only works on screen.▶ Jump to 14:16Speaker · Will Habeck - 18:13 – 20:27Listen
Closing: AI amplifies the foundation, designers decide what to build
Recaps the business results the whole chain produced ($1 billion in incremental originations in Q1 2026) and lands on the core argument: speed is no longer scarce; naming, structure, and context are what AI can amplify.
When everyone can build fast, the designer's value is deciding what's the right thing to build and what's true to the brand.
A pure argumentative wrap-up with no demo footage — good for listening straight through.▶ Jump to 18:13Speaker · Emily