Eve's designers stopped writing handoff docs and started shipping pull requests

The old handoff was a Notion doc full of screenshots. The new one is a pull request. That single change moves where design ends and where code begins, and it is reshaping who does what on your team. Let me catch you up.

The deep cut

• The handoff is the workflow now. Ayon at Eve ships production code to close the last 20% instead of writing annotated docs.
• Speed without quality doubles the bugs your users hit. Tyler Folkman's math: ship twice as fast at the same defect rate, customers see twice the failures.
• An eval scores the system, not the shopper. Albertsons' assistant lifted baskets 26% by catching forgotten milk, something no eval would test.

The doc nobody wanted to write

For years the design-to-code gap got patched with paperwork. You built the vision, then wrote a Notion doc listing every deviation with screenshots, and you still landed short. Ayon, an AI-native designer at Eve, says engineers shipped 75 to 80% of what he built, and closing the rest ate hours. Now he skips Figma for 95% of his work and edits the production codebase directly, tweaking layout, spacing, and type with AI tools.

The point is not that designers become engineers. It is that the last 20%, the polish that used to die in a backlog, now ships. Ayon's team, mostly people with no technical background, are deploying on Vercel and using Git. His onboarding move is worth stealing: build a thin sandbox that mirrors your real app, so designers can run it, break it, and prototype with fake data without waiting on migrations.

The floor plan draws itself

The same shift is hitting spatial design. JLL Design rolled out laiout across all 65 studios after a survey showed a 50% cut in test-fit time. Work that took days now takes as little as a few minutes, up to 40 on average, and the CEO claims production time drops up to 93% depending on complexity.

Here is the tradeoff worth naming in your review. The tool generates the first draft; designers shape the final space. JLL's operations lead says the real change is where hours go, toward refinement and client engagement instead of cranking out initial options. That is the pattern across every one of these stories. AI takes the production half, and the human keeps judgment. The risk is treating the draft as the answer.

Speed is a bug multiplier

Faster is not free. Tyler Folkman, who runs product and AI at JobNimbus, gave the math that should sit on your wall: if you ship twice as fast and hold the same defect rate, your customers hit twice as many bugs. He points to that as the reason uptime at big companies has been rough lately.

Folkman's fix is to build repeatable loops, agents that fetch their own inputs, do the work, pass a gate, and write a named artifact. His favorite one pushes back on the team instead of doing their work: did you talk to customers, where does the data come from? The warning underneath is sharp. When a bot hands you a pile of output you can't evaluate, sending it up the chain "is not doing product work. That's outsourcing your thinking."

The eval that forgot the customer

Everyone is calling evals the new PRD, and they are right that pass/fail tests beat guessing requirements for a system that outputs something different per user. But Jeff Gothelf catches the trap: an eval scores system behavior, not customer behavior. Albertsons told the Wall Street Journal that shoppers using its full AI assistant spend up to 26% more, because they stop forgetting items. No eval would ever have surfaced that. The store doesn't know you forgot the milk. Only you do.

His fix is one line added to every eval: name the specific customer behavior you expect to change. Then name the score you'll tolerate, the person who fixes an outcome miss, and the date you refresh the eval from real production data. Basket size is the business result. Not forgetting the milk is the experience that earns the repeat visit.

Three questions for your team

• What does our real definition of done say about customer outcomes, not just eval scores? If it only checks system behavior, we are shipping features nobody asked for.
• Have we measured our defect rate against our new shipping speed? If we doubled speed and held quality, we doubled the bugs our users feel.
• Do our designers have a sandbox and Git access, or are they still writing handoff docs? Pick one product area, one tech lead, and one designer, and close that gap this quarter.