The internal AI tool that's transforming how Stripe designs products | Owen Williams
TL;DR
Stripe built its own AI prototyping stack because generic tools produced “Tailwind indigo blur” instead of Stripe-quality dashboards — Owen Williams spent roughly 18 months creating ProtoDash, an internal React-based prototype environment with Cursor rules, Stripe’s “Sail” design system, and MCP integrations that gets teams “90% of the way” to realistic product mocks.
The real surprise is who adopted it: PMs may use it more than designers — Williams says he was initially nervous about “PMs designing,” but found the tool helps PMs unblock themselves, communicate more clearly with designers, and run earlier user research without waiting for polished Figma work.
The killer use case is prototyping complex, data-heavy products that Figma struggles with — Stripe dashboards need filters, states, zero-data views, internationalization, messy data, and multi-step flows, and Williams argues it’s “nearly impossible” to model all of that convincingly in static design files.
ProtoDash evolved from a local prototype starter into a browser-based internal product, ‘ProtoDash Studio’ — designers can spin up a dev box from a URL in about two minutes, prompt an embedded LLM in-browser, generate variants, annotate UI directly for fixes, and even run design reviews without the usual Google Doc feedback dump.
The tool is changing handoff, not just mockups — one Radar team recreated major fraud-product flows in ProtoDash so precisely that engineers could use the prototype pull request as a practical source of truth, inspecting real padding and behavior instead of translating redlines from static files.
Williams’ broader point is that AI makes bespoke internal tools newly feasible — instead of waiting for centralized tooling teams, product orgs can now build niche systems tailored to their own review culture, quality bar, and workflow cadence, then keep evolving them through pull requests from designers themselves.
The Breakdown
Why Stripe designers were tired of “indigo blur” prototypes
Claire Vo opens by framing the problem perfectly: AI prototypes often look like generic Tailwind demos instead of real products. Owen Williams, a design manager at Stripe, says that in design reviews he kept seeing convincing-but-wrong dashboard mockups with odd navs, off fonts, and what Stripe jokingly calls “indigo blur” or “blurple slop.” His core insight was simple: Stripe already has a highly predictable design system, so AI should be building from those real components instead of improvising.
The first version: a React shell, Cursor rules, and Stripe’s design system
Williams built ProtoDash himself, leaning on his engineering background and a long-standing love of “technical designers.” V1 was basically a prototype starter kit: React routing, Stripe’s internal design system “Sail,” an MCP integration, and a thick layer of rules telling Cursor what to do first, what to avoid, and how not to hallucinate missing components. The goal was lowering the barrier so far that a designer only needed to know something like npm run dev and could still generate a realistic Stripe dashboard.
From local setup to dev boxes and clickable design reviews
At first, people ran ProtoDash locally, but the better move was deploying it on Stripe’s dev-box infrastructure so users could spin up an environment from a URL in about two minutes. That changed the social experience as much as the tooling: instead of another Figma slideshow, reviewers could click around a live prototype. Williams lights up talking about this shift — “demos not memos” — because it turns reviews into real product conversations instead of JPEG critique sessions.
Why dashboards, edge cases, and messy data pushed this over the line
Claire pulls out the deeper reason this matters: Stripe builds data-heavy products, and it’s brutal to fake all the filters, states, zero-data views, and long-tail scenarios in Figma. Williams agrees and shows how ProtoDash can quickly simulate startups vs. enterprise customers, localization issues like Dutch text blowing up layouts, or “messy data” conditions. The point isn’t just prettier mockups; it’s getting much closer to how real users will actually experience the product.
ProtoDash Studio: the browser-based vibe coding layer
Williams’ “dream” was something “like v0, but for us,” so he added a browser-based layer called ProtoDash Studio. Instead of opening Cursor, users can go to a web UI, browse prototypes, remix other people’s work, and prompt an embedded LLM directly in-browser to create new variants. He describes building it mostly by “yelling at Claude Code for 18 months,” which captures both the scrappiness and the weird freedom of internal tooling that can be useful without needing production-grade polish.
Annotation, self-testing, and AI-assisted design reviews
The wildest part of the demo is how much workflow got folded into the tool: AI can take screenshots, inspect the browser, self-test its output, and iterate if something breaks. Users can annotate specific UI elements directly instead of describing “the element with class name 82F,” and there’s even a review mode where teams leave comments on the prototype itself, summarize feedback with AI, and queue fixes automatically. Williams’ favorite outcome is that the tool doesn’t just collect feedback — it helps close the loop and generate the “here’s the fix” follow-up designers usually do by hand.
PMs, handoff, and the bigger case for internal AI tools
Williams admits he got “a little nervous” when PMs started using ProtoDash, but now sees it as a feature, not a bug: PMs can prototype ideas, unblock themselves, and show something tangible before arguing over staffing. He also shares a Radar example where designers rebuilt major fraud-product flows so faithfully that engineers could use the prototype PR almost like source material for implementation. Claire’s bigger takeaway is that this is the underrated AI opportunity right now: not replacing SaaS tools wholesale, but building weirdly specific internal products that match your team’s culture so well they actually change how work gets done.