--- schema: 1 type: topic title: AI-Assisted Design Research description: A synthesis of the channel's material about AI-Assisted Design Research. canonical: >- https://liquidgold.cc/channel/mobbindesign/explore/topic/ai-assisted-design-research-signresearch markdown: >- https://liquidgold.cc/channel/mobbindesign/explore/topic/ai-assisted-design-research-signresearch.md llms: >- https://liquidgold.cc/channel/mobbindesign/explore/topic/ai-assisted-design-research-signresearch/llms.txt parent: 'https://liquidgold.cc/channel/mobbindesign.md' channel: mobbindesign updated: '2026-10-23T22:20:24.214Z' language: null children: - 'https://liquidgold.cc/watch/YbLF42BaoZs.md' topicSlug: ai-assisted-design-research-signresearch searchChunks: - ordinal: 0 text: null startSeconds: 0 endSeconds: 11 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=0' heldReason: rights - ordinal: 1 text: null startSeconds: 11 endSeconds: 46 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=11' heldReason: rights - ordinal: 2 text: null startSeconds: 46 endSeconds: 57 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=46' heldReason: rights - ordinal: 3 text: null startSeconds: 57 endSeconds: 111 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=57' heldReason: rights - ordinal: 4 text: null startSeconds: 111 endSeconds: 122 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=111' heldReason: rights - ordinal: 5 text: null startSeconds: 122 endSeconds: 169 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=122' heldReason: rights - ordinal: 6 text: null startSeconds: 169 endSeconds: 180 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=169' heldReason: rights - ordinal: 7 text: null startSeconds: 180 endSeconds: 202 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=180' heldReason: rights - ordinal: 8 text: null startSeconds: 202 endSeconds: 212 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=202' heldReason: rights - ordinal: 9 text: null startSeconds: 212 endSeconds: 244 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=212' heldReason: rights - ordinal: 10 text: null startSeconds: 244 endSeconds: 256 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=244' heldReason: rights - ordinal: 11 text: null startSeconds: 256 endSeconds: 281 watch: 'https://liquidgold.cc/watch/YbLF42BaoZs?t=256' heldReason: rights --- # AI-Assisted Design Research A synthesis of the channel's material about AI-Assisted Design Research. Canonical: https://liquidgold.cc/channel/mobbindesign/explore/topic/ai-assisted-design-research-signresearch Parent: [Mobbin](https://liquidgold.cc/channel/mobbindesign.md) A single prompt asking Claude for a fintech onboarding flow returned two drafts, and the presenter's verdict on the first was blunt: "This one, it looks fine, but it's generic." The second draft accounted for identity checks and compliance requirements, which the presenter says made it feel far more considered. The difference, according to the presenter, came from one connection. ## Grounding AI output in shipped products ### What the connection actually adds Connecting an AI tool to Mobbin MCP, the presenter says, is like handing the AI a library card to over 600,000 real app and website screens, flows, and animations from products already shipped. The claim is about access, not model quality. The tool does not get smarter. It gets a reference set of finished interfaces to reason against. The 600,000 figure is the presenter's own, and the video does not show how it was counted, so treat it as a vendor-scale claim to verify on the product page rather than a fact this chapter can vouch for. What the demo does establish is one before-and-after on one brief. The grounded flow was not longer or more elaborate by design. It was more specific to the category. A designer can test that distinction on their own work: run the same brief with and without the reference library and compare what each output has to answer for. ### Research-first prompting for a specific industry The prompt in the demo is the part worth copying. It asks the tool to research how top apps design onboarding and account opening flows across fintech or banking, find the most relevant screens and flows, and compare patterns across products. The presenter then asks for a visual report of top examples, common patterns, and the practices top companies converge on. In the captured run, the tool fetched account opening flows in the finance category and studied welcome screens from fintech apps, which the presenter reads as evidence that it understood a banking flow differs from a generic one. The report is only as useful as the comparison it forces. The stated goal is to identify what the best apps do differently, not to average them. A designer should read the report for where products disagree as closely as where they agree. ## From research to a built flow ### Attaching the brief and the research together With the report in hand, the workflow moves into a design file with a made-up design system. The presenter opens a fresh Claude session with the design tool's MCP connected, attaches the research report and a short document describing the product, and asks for an onboarding flow for Curve. The document is optional, but the presenter includes it because it helps the agent understand the full context. The instruction that matters most comes last: flag anywhere the agent intentionally deviates from a pattern it found. That flagging step is what makes the output reviewable. In the demo the agent pushed back on the brief itself. For a safe offers nudge, it placed the nudge on the homepage right after onboarding, citing an app where the setup bonus lives on the homepage. The agent explains its reasoning rather than silently following or silently overriding the brief, and that explanation is what a designer has to accept or reject. ### Why one-shotting fails The strongest direct comparison in the material is a Figma agent asked to design map and filter views for iOS. Without research, the presenter says, the output had broken UI. When the agent was connected to Mobbin MCP first, it studied references citing their sources and produced a much better result. This is one example on one task, and the two outputs are shown side by side without any measurement. It supports the research-first workflow; it does not establish how often one-shotting fails. ## Where the tool stops The video lists other uses: comparing DoorDash and Uber Eats on checkout, tipping, and order summary; pulling ten empty states from productivity apps; drafting copy tailored to users with the sources cited; and remixing ideas across industries. Setup instructions live at mobbin.com/mcp. The presenter notes that Mobbin MCP is included with any paid Mobbin plan and works with Claude, Antigravity, Codex, and Cursor. The limit the presenter states is the one a careful reader should weigh most. The tool cannot decide what to prioritize, what to keep, and what to cut. That part, the presenter says, is still the designer's call. The material does not test whether the agent's deviations were correct, and it cannot establish that a reference-grounded flow converts better. What it does establish is that the research step changes what the agent has to justify. - Give the tool a reference library of shipped products before asking for a design, and compare its output with and without that access. - Ask for patterns and trade-offs together, and read the report for where competing products disagree. - Attach the research report and a short product brief, then ask the agent to flag every deviation from a found pattern. - Skip one-shot generation for high-stakes screens; the Figma example showed broken map and filter UI without research first. - Keep prioritization, cuts, and final judgment with the designer, since the presenter says the tool cannot make that call. ## Source videos - [I Gave Claude 600,000 UI Screens… Then This Happened](https://liquidgold.cc/watch/YbLF42BaoZs.md) > Search chunk text is held until public-source provenance is verified.