Reddit Has 700 Prototypes on a Dashboard. The Column It Needs Is Shipped.

Claire Vo put Blair Metcalf on How I AI this morning to demo Reddit Proto, and it is the best version of "everyone builds" I have seen this year. I'm going to argue with the dashboard, not the harness.

The short version

Blair Metcalf, principal UX engineer at Reddit, demoed Reddit Proto on Claire Vo's How I AI on October 26, 2026: a desktop app that hides Git, a boilerplate of the whole Reddit app with the design system baked in, feature toggles, lockable navigation, persona switches, sign-in so the prototype runs on real Reddit data, on-screen annotations sent to Cursor as targeted edits, and deployment on internal Kubernetes. His numbers: 100% adoption across UX and product, over 150 people, about 700 prototypes on a public internal dashboard. The impact, in his words, is instant alignment. The handoff is a feature capsule: Figma frame through a custom MCP, generated PRD, Cursor writes code in a sibling repo, and a draft PR that "may or may not go through." The console replacing the 700-item list has people, tags, teams, and platforms, and no column for shipped or landed. Figma's design chief reported 17x more code and 1.3x more releases; 13 of the 39 agents I built were orphans. Nielsen Norman Group's survey of 521 practitioners, published three days earlier, concludes that leaders should measure whether work reaches users, not whether people use the tool. The rewrite: keep the rails, add two columns, shipped with the reviewer's name and landed with the behavior that changed.

What Reddit built

I read the transcript rather than watching the demo, so everything here is as Metcalf describes it.

The opening line is the thesis: the distance between design and product is wide, and "if we made everyone a builder, they could vibe code new features and products easily." So the team built a harness. A desktop app that gets a designer into a new project without seeing Git, a branch, or a file tree. The boilerplate is the whole Reddit app, the design system and every component and UX pattern included, and as teams add features the boilerplate absorbs them, so everyone starts from the current mountain.

Then the doodads, his word. Feature toggles. Navigation you can switch off "because some stakeholders click too much and you don't want them to click." Device, theme, and every user state down to who is a moderator. And the one I'd steal first: sign in, and the prototype runs on real Reddit data. His example was search. He annotated the screen with "make relevant type-aheads, trending, recent communities" and an overall direction, sent the annotations to Cursor as a file, told it to apply them, and it built search that pulls real results, sends them to an LLM, and ranks them. Not a mock. "If designers are mocking everything in Figma and then they show that to leadership, leadership's going to look at that and be like, this isn't relevant to me."

You Cannot Mock a Distribution is the design-handbook argument for that choice, and Reddit made it in software.

Deployment is on internal Kubernetes, which he calls basically an internal Vercel. Share a URL in Slack or a QR code, and it installs as a PWA on a phone. Leadership clicks through, feedback arrives live and in Slack, and live commenting on the prototype itself is on the way.

The numbers. 100% adoption across UX and product, over 150 people, an almost perfect mix of designers and product. About 700 prototypes on a dashboard, all of them public inside the company. And the best prototypers were not the obvious ones: a visual designer and an animator, people he describes as great directors with huge context who try things without fear of failing.

Asked where the impact landed, velocity or craft, he said neither. "It's instant alignment."

The question nobody asked

Of the 700, how many shipped?

Thirty minutes, and it didn't come up. It isn't a gotcha. The episode is about the harness and the harness is excellent. But listen to how the handoff works. The feature capsule takes a Figma frame through Reddit's custom MCP, which maps every token, component, surface, and journey of Reddit Product Language across Figma, iOS, the production repo, and the prototyping repo (he calls the visualization a Rosetta stone, and built it because nobody could see what was in the harness). Out comes a generated PRD, user goal and acceptance criteria included. Cursor writes production-shaped code in a sibling repo to keep prod clean, and a draft PR goes up. His words: "that'll get reviewed and may or may not go through."

May or may not. Nobody's counting which.

The console he's building to replace the flat list of 700 has people, tags, teams, platforms, iOS versus Android. It gives leadership visibility into who's working on what. It has no column for merged. None for landed.

I have a number for what happens without that column. 17x the Code, 1.3x the Releases. Design's Job Is the Spec Now. carries Loredana Crisan's figures as TechCrunch reported them: 17x more code at Figma, 1.3x more releases. Generation went up seventeenfold and release went up by a third, at the company that sells the canvas. And my own fleet. Of 39 PM AI agents I built across four product orgs in 80 days, 13 were orphans, with zero references from anything outside the fleet. Built, reviewed, admired in a Slack thread, never read again. A prototype dashboard with 700 rows and no outcome column is the same shape with a nicer UI.

NN/g published the counterweight three days earlier

Maria Rosala and Rachel Banawa at Nielsen Norman Group surveyed 521 UX practitioners between August 24 and September 11. AI saved time on every task respondents used it for; the mean time-savings rating was 4.2 out of 5. And sentiment was 3.57 out of 5, barely above neutral. Of the 226 who left a comment, 69% expressed negative sentiment, 41% were purely negative, and the most common reason was unreliable output and the burden of checking it. One respondent: "I now sometimes find myself trying to quickly fix completely unusable prototypes made by a product manager. And I'm expected to come up with a solution just as quickly as AI does."

Their conclusion is the right one, and it's the column. The unit of analysis should be the system, "the total time and effort, across everyone involved, needed to ship work of acceptable quality." A prototype generated in minutes that costs a designer days to make usable is a net loss. Leaders "should not measure whether each individual uses AI, but whether work reaches users faster and better."

Reddit measures adoption and count. NN/g says measure arrival.

To be fair to Reddit, the PM in that NN/g comment did not have a design-system MCP. Wayne Lin's Figma post the same day makes the point from the vendor side: the agent is only as good as the context it gets, and a team's judgment about when to use which component lives in guidelines attached to the library, not in the prompt. Metcalf built that in a custom MCP, iterated it through CI feedback and engineer feedback, and that is probably why 150 people use it instead of 15. The rails are why the prototypes are good. They are not why anything ships.

The rewrite

"Make everyone a builder" becomes "make everyone a builder, and put the shipped column on the dashboard."

Two columns on the 700.

Shipped. The PR that merged, the date, and the name of the reviewer who approved it. Not "sent to engineering." Merged. The feature capsule already produces a draft PR, so the data exists; it's just not joined back to the prototype that spawned it.

Landed. The behavior that changed. What Landing Actually Is has the test: point at something a user used to do differently, now does with your feature, and would be annoyed to lose. One line, one date, one name.

Then the number the dashboard should lead with: how many of the 700 have a blank in both columns after a quarter. That's the alignment-only count. Instant alignment is real and it's worth a lot. Metcalf's own line is that you don't have to take nine months to get a feature into prod to find out if it works. But alignment is the price of a landing, not the landing. Twelve Builders, Twelve Products, Zero Landings is the same argument at the org-chart level: twelve builders delivering twelve products is only a win if twelve products can land, and the CPO Insights Report, as cited there, put speed to market as the number one internal challenge, with the bottleneck moved from building to launching.

Under AI Product Management the claim is that AI collapsed the cost of the parts PMs were trained to be good at and left the parts nobody trained for. Reddit collapsed the prototype. The part nobody trained for is the column.

This week: if you have a prototype dashboard, a Figma Make folder, a Lovable workspace, anything with a count on it, add the two columns by hand for the last quarter. Count the double blanks. Then show the count next to the adoption number the next time the harness is in a leadership review.

Related answer: What should a prototyping dashboard measure besides adoption?

Sources: Claire Vo with Blair Metcalf, "Why Reddit built its own AI prototyping tool | Blair Metcalf," How I AI, October 26, 2026; the transcript read in full, the video not watched, and every claim about Reddit Proto is as Metcalf describes it there. Maria Rosala and Rachel Banawa, "AI in UX: Work Is Faster, But Practitioners Have Mixed Feelings," Nielsen Norman Group, October 23, 2026, read in full. Wayne Lin, "How to put your design system to work with the Figma agent," Figma, October 23, 2026, read in full. Falk Gottlob, "17x the Code, 1.3x the Releases. Design's Job Is the Spec Now." (with Loredana Crisan's figures as Ivan Mehta reported them at TechCrunch), "39 PM AI Agents Deployed: What Stuck, What Died, and Why," "What Landing Actually Is," "Twelve Builders, Twelve Products, Zero Landings," and the design handbook chapter "You Cannot Mock a Distribution."