Greg Isenberg went live on a weekend, opened claude.ai/design, and said the thing every founder was thinking: the YouTube tutorials about it "didn't feel authentic to me."
So he stress-tested it on camera instead. Crashes included.
I watched the full hour so you don't have to. This Claude Design review covers what the tool nailed, where it broke live on stream, and the one output that made Greg say "this is probably the best deck I have ever seen created by any LLM, period."
Here's the source video. Worth watching at 1.5x if you're deciding whether to give it a weekend:
What Claude Design Actually Is
Quick facts before the story.
Anthropic shipped Claude Design on April 17, 2026 through Anthropic Labs. It's a research preview, powered by Claude Opus 4.7, and it lives at claude.ai/design for Pro, Max, Team, and Enterprise subscribers. Not on the free tier.
It makes wireframes, high-fidelity prototypes, slide decks, one-pagers, and (in theory) videos. You can feed it a screenshot, a codebase, or a Figma file as context, and it has its own usage quota separate from your normal Claude chat limits.
Greg has no affiliation with it. Neither do I. That's what makes his hour interesting: it's a founder poking a tool with his own tokens, in public, with the failures left in.
The Test: An App Idea Picked in 60 Seconds
Greg didn't prepare a demo script. He pulled an idea from ideabrowser.com on stream: a brain-training app for seniors. Cognitive decline robs seniors of independence, the app adjusts difficulty automatically, personalized brain workouts. He named it Senior Brains.
And he meant it as a real business, not a prop. His read: "I could see how this could be a 5, 10, 15 million-dollar ARR business." Distribution through Instagram Reels and Facebook, because the seniors are on Facebook and their adult kids are on Reels.
That framing matters. He wasn't testing whether Claude Design makes pretty pictures. He was testing whether it moves a real idea forward.
Wireframes First: The Token-Saving Move Most People Skip
Every tutorial one-shots a polished landing page. Greg did the opposite and started with the lowest-fidelity wireframe possible.
Two reasons.
One: tokens. "If I just go and one-shot it, I'm going to end up spending a ton of tokens." Multiple viewers in his chat reported burning through their entire allowance in about two hours.
Two: constraints. A wireframe forces you to decide which features actually exist before you fall in love with gradients. His rule on fidelity: go low-fi or high-fi, never mid-fi. "You can't be half pregnant."
Then came the part that genuinely surprised him: the questionnaire. Before generating anything, Claude Design asked about the primary device, which screens to mock, gamification elements, accessibility defaults (it suggested large text and high contrast on its own, because it understood the users were seniors), and how prominent the family caregiver should be in the app.
Greg's reaction: "I'm blown away by how good these questions are." It extrapolated from a two-line idea "like a product manager."
His sharper point came later. Answering those questions is where you, the founder, get clear on what you're building. "If I let it do all the thinking, I find myself creating projects and never finishing them." The questionnaire isn't friction. It's a free product management session.
Three Directions, Like an Agency Deliverable
The wireframes came back as three distinct directions, each with a story:
- Direction A, the warm stack: card-based home, one clear action, mascot as a small sidekick. Duolingo-adjacent but calmer.
- Direction B, mascot forward: the mascot (a bean named Bean) is the navigator and talks to you like a chat.
- Direction C, calendar ritual: today's session as a single scroll. Less gamey, more "morning coffee and crossword vibes."
Greg co-founded an agency that builds AI apps for some of the largest companies on the planet, and his comparison was blunt: this is the direction A / direction B / direction C presentation an agency would charge real money for. He'd spent zero dollars at that point.
His chat voted A. They moved on.
Then It Broke. Twice.
The moment Greg asked for high-fidelity versions, Claude Design errored out. Refresh, retry, error again. Live, in front of everyone.
"This is why I'm doing this live stream. When you go on YouTube and see tutorials, this gets edited out."
The debugging surfaced three real limitations:
- You can't run two tasks at once. Greg kicked off a pitch deck in a second project while the app job was cooking, and the first job silently froze. No warning from the product.
- There's no stop button when a task hangs.
- Leaving a project folder and coming back loses your context. The questionnaire he wanted to revisit was just gone.
His summary of the experience: "When I'm on Claude Design, I'm walking on eggshells."
Worth knowing before you plan a client deliverable around it. It's a research preview and it behaves like one. The lesson he kept repeating: it's going to break sometimes, and you just have to keep going.
The Sequoia Deck That Stole the Show
While the app job was stuck, Greg asked for something absurd: a VC-style deck to raise $2 million from Sequoia for an app that didn't exist an hour earlier.
What came back flipped him from curious to converted. The deck had:
- A market wedge: 58 million Americans over 65, and 1 in 3 will develop cognitive decline
- Competitor framing: today's options are brain-game toys or a 6-month wait at a memory clinic at $400 out of pocket
- A buyer insight Greg loved: the adult child is the buyer, mom is the user. The daughter sees a Meta ad ("the gift that keeps mom sharp") and buys the $228 annual gift
- Unit economics: $62 blended CAC, $228 year-one LTV, 3.3-month payback, 3.7 LTV to CAC
- An expansion story: start consumer, grow into senior living facilities
- Full speaker notes for a 5-minute pitch
It even pulled scientific backing on cognitive training trials without being asked.
That's when he said it: "This is probably the best deck I have ever seen created by any LLM, period."
My grain of salt, founder to founder: those numbers are projections a model made up for a company that doesn't exist. A $62 CAC is fiction until an ad account says otherwise. Treat the deck as a thinking tool that forces you through market, buyer, and economics in 10 minutes. As that, it's absurd value. As something to actually send Sequoia, obviously not.
The Draw-On-It Trick That Shows Where This Is Going
The high-fi designs eventually landed, and they were good. Onboarding with big tap targets, one idea per screen, a family-cheering feature where relatives send hearts while grandma trains her memory.
Then Greg tried the annotation tool. He drew a rough note on the results screen: put a Facebook share button wherever conversion would be highest.
Claude Design added the button and wrote the copy itself: "Share this win."
Greg, who has built social products used by millions, caught the detail instantly. "Share this win" converts a lot more than "share this on Facebook." He didn't ask for that copy. The tool made the stronger call on its own.
That's the part of this walkthrough I keep thinking about. What sticks isn't the pretty screens, it's that the tool made a growth decision better than most junior marketers would.
Where It Flopped: Video
Greg ended the stream by asking for a 30-second ad for the app. Target: the 35 to 45-year-old children of the actual users.
It produced an animated slice-of-life spot. Ruth loses her glasses, daughter Sarah gifts her the app, warm ending.
His verdict after a revision: "It's better, but it sucks." Score: "five on 10 at best." A viewer ran a similar prompt through a dedicated video model and the difference wasn't close.
So the honest scorecard is lopsided. Wireframes: best in class, his words. Decks: unbelievable. Visual design: really good. Video: skip it.
What Bootstrapped Founders Should Steal From This
I've covered AI design twice before on this blog: Meng To's AI design workflow and the 5-step process for designing an app with AI. Claude Design compresses the front half of both into one tool.
Three moves worth copying this week:
→ Wireframe before pixels, always. Greg does it to save tokens. Do it because the cheapest place to kill a bad feature is a black-and-white sketch.
→ Answer the questionnaire yourself. That's where you get clear on what you're actually building. Delegate the drawing, not the thinking.
→ Run the fake deck exercise on your real business. Ask for the deck, then argue with the CAC and LTV it invents. If you can't say why its numbers are wrong, you don't know your funnel yet. And remember a deck-slide funnel only works in real life if the boring middle (qualifying leads, following up, onboarding users) actually runs without you. TwiLead has a solid breakdown of business automations if you want a map for that part.
Because the uncomfortable shift is that design used to be a moat, and a $10K agency deliverable now takes an hour plus some patience with error messages. The moat moved to the parts Claude Design can't do for you: picking the right wedge, and talking to actual seniors (or whoever your Ruth is) before you build.
That's the stuff I dig into every week on the podcast with bootstrapped SaaS founders doing $100K to $10M a year. Operators one step ahead of you, walking through their real numbers.
Listen to the Profitable Founder Podcast
FAQ
What is Claude Design?
Claude Design is Anthropic's AI design tool, launched April 17, 2026 as a research preview from Anthropic Labs. It generates wireframes, high-fidelity prototypes, slide decks, and one-pagers from a text prompt, a screenshot, a codebase, or a Figma file. It runs on Claude Opus 4.7 and lives at claude.ai/design.
Is Claude Design free?
No. It's available to Claude Pro, Max, Team, and Enterprise subscribers, with its own usage quota separate from chat. Budget matters here: viewers in Greg's stream reported burning their full allowance in about two hours of heavy generation, which is exactly why he starts with low-fi wireframes instead of one-shotting polished screens.
Is Claude Design good for video?
Not yet. Greg rated the video output "five on 10 at best" after two attempts at a 30-second ad. Wireframes, product design, and decks are its strong lanes. For video ads, a dedicated video model still wins by a wide margin.
Can Claude Design build a working app?
No. It produces designs, not shipping code. Greg's point on stream: you'd take the approved direction into Claude Code or another coding tool to actually build it. The realistic pipeline is idea, wireframe, high-fi design, then a separate build step.
