The AI Design Tool Field Guide for PMs: The Last Call

Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.

Do you know what tequila won the 2026 Bartender Spirits Awards, scoring 98 points? Me neither. What actually gets poured is the daily driver bottle of Espolòn Blanco bartenders reach for that gets shaken into thousands of margaritas every day.

Lab tests of AI tools are instructive, but they aren't the full story. The score can't tell you which good tool becomes a habit, only the field can. The missing ingredient is what practitioners reach for day after day in their actual workflows.

Over the past four weeks, I've introduced you to four AI design and prototyping tools, then put them head to head on three typical product use cases:

This week, I'm putting them to the test against real practitioners in the field: Ankush Singla — a former FanDuel AI Product Director — and Luke Moderwell — an ex-Mailchimp product designer turned design/engineering builder.

Ankush lives and breathes "vibe design." He described one high-stakes meeting in advance of a VP presentation the next day that was rabbit holing where he prompted Lovable in parallel, showed the prototype, and built consensus on what to present, live.

Luke wears many hats in his builder-consultant role. In one example, he's building the design system guardrails in Claude Design so that his CPA-teammate has free reign to iterate on concepts for the agentic tax product. In another, he's moving context from Slack thread to Linear to Claude Design to a multi-agent coding team in Claude Code.

I asked each of them to make a bet on the winners by scenario, then I revealed the rankings, and learned where they disagreed.

First, their picks.

Ankush: Scenario 1: Claude Design (followed by Lovable/v0). Scenario 2: Lovable, with v0 in the conversation. Scenario 3: Lovable.

Luke: Scenario 1: Claude Design (due to model interoperability with Opus 4.8) and v0. Scenario 2: v0 and Lovable as a potential close second. Scenario 3: Google Stitch (although Claude Design was his gut reaction).

Personally, I went in thinking it would be a jump ball, but I was only partially right. No single tool ran the table, but Claude and v0 were pretty consistent. If you stuck with one across these scenarios, you'd be in the Good Place with your design team. The movement happened near the edges, where late-to-the-party Lovable came from the doldrums on the first two scenarios to a commanding second place on the 0 to 1 scenario. Google Stitch showed up to the wrong party.

Interestingly, both blind bets landed close to the lab on Scenario 1 with Claude Design at the top and both overrated Lovable in Scenario 2. Ankush clocked Lovable on its 0 to 1 prowess, likely due to his experience using the tool day-to-day (that was the de facto AI prototyping tool license at his last gig.) Luke correctly guessed that v0 would perform, but his optimistic reframe on S3 missed.

My intention going into this newsletter series was never to crown a champion. That certainly would have driven clicks. Maybe even encouraged one of you to forward this to your procurement department (a boy can dream). I always envisioned this was going to land on a layered, "it depends"-style answer (my former mentors at Deloitte are beaming after reading that sentence).

If you have the luxury of choosing between tools (and many of you don't — prayers for our Office 365 + Teams + Copilot brethren), this is how I would think about it:

Pick a default — Claude Design and v0 hold across every job. 0-to-1 is the one place the leaderboard reshuffles. The only question that matters: “Does this give me what I need to walk out of the room with?”

Ankush agreed:
"You have different [outputs] you're creating for different people... It's the same case yesterday as it is today. It's just that the thing you're creating is different and higher fidelity."

Luke brought up a core tension inherent in this new world of AI prototypes in everyone's hands: high fidelity can paper over ill-conceived ideas. "[The prototypes can be] premium mediocre… incredibly premium but it's actually worthless." (See also: the typical Netflix-produced genre film.)

Ankush also converged on this idea. These tools produce an incredible output — and it can be an extremely effective communication tool — but it shouldn't be relied on exclusively. Take, for example, the Substack cancel flow scenario. Even if the v0 prototype is damn near perfect (it would be after another hour of edits), you would never toss it to your engineering team without mapping happy path / alternative paths. Clicking through every single survey option and trying to clock every aspect of the state changes would be Sisyphean.

AI prototyping is a powerful hammer (more like a cordless framing nail gun) to have in your arsenal, but you have to fight your instincts to look at every design problem as a nail.

I knew instinctively that inputs mattered, but the tests hammered that home for me even more. There are two aspects to this: 1) the design system, 2) the prompt.

The Design System: A design system is a shared set of rules, reusable components, and guidelines that inform how teams design a particular product.

When you're bringing a screenshot to these tools, they're guessing on the design system. Your mileage may vary, but if you're not using a high powered model, the result is probably going to be mid AF (the technical term for my lived experience in these tests).

Each of these tools also supports bringing your own design system, allowing you to import from Figma, GitHub, a website, or a DESIGN.md file.

Ankush: "Lovable has a way to build a design system in it, which I think is like a little bit of a cheat code… Spend six weeks building a design system in Lovable and the output is going to be pretty high quality."

Using screenshots likely disadvantaged Lovable in Scenario 1 and 2, but there's a limit to how much grace I'm willing to give. Lovable restricts the use of design systems to Enterprise plans, gatekeeping this table stakes feature to the FanDuels of the world. I think this is a product strategy miss, but I've only spent the last decade building monetization systems.

The Prompt: Knowing how to precisely describe what you're expecting and what context to provide is the other half of the battle. (Pro tip: Use another model to write your prompt.)

This is backed by the tests. Every single tool nailed the cancel survey branching logic I specified explicitly in the Substack prompt. Every tool struggled managing state logic (i.e. reflecting a post-cancelled state correctly across all surfaces) which I underspecified in the prompt.

A few interesting wrinkles: Claude and Lovable ask you questions when you under-specify, and pros under-specify on purpose to mine for ideas. "I under-specify intentionally because I know the things I care about... then I take the interesting things it comes up with and combine it with the things that I know I want" (Ankush).

Luke's workflow informs how a designer-by-trade actually uses Claude Design: starts with art boards (a la Figma) to explore component-level concepts, brings them into full prototype context, and hands off the output directly to Claude Code.

Across each tool I tested, model selection was available, except in Lovable's case where they make that decision for you. Claude Design and v0 both ran on Opus 4.8 by default, and I didn't downgrade. It mattered — the best model generally produced the best results.

Unlike "typical" AI model usage (text generation), at this stage in the lifecycle of AI + design, I'm not sure I would be comfortable downgrading or know where I could take an inference hit without sacrificing quality.

Unsurprisingly, in both of my conversations — with a PM with "AI" in their title, and a deeply AI-pilled product builder — the model quality topic came up repeatedly.

Luke argued Claude Design's inherent advantage with the Apple Silicon/vertical integration analogue. "Claude should win everything pretty handily because they should be the best at using their own models... these are coding tasks... we're building a really sleek web app that you think looks like a design."

Ankush didn't think Stitch was in a fair fight given the underlying model. "You're comparing apples to oranges in terms of model quality... Opus 4.8 to Gemini 3.1 Pro... and 3.5 Pro is already coming out soon." He's right, and this is a structural consideration. Claude Design and v0 both ran on Opus 4.8, which stood as the strongest model for reliable, long context work at the time of testing (calm down, Fable stans). The irony: Google had shipped a newer model — 3.5 Flash — that beats 3.1 Pro on exactly these coding and agentic benchmarks. It just wasn't in Stitch yet.

Unlike the big ticket software of yesteryear (that you probably downloaded from The Pirate Bay), the AI prototyping sticker shock comes when you actually use the tools. While the pricing pages promise clean monthly recurring bills in round numbers ("$25") and usage pools ("100 credits per month"), this hides the lived cost-to-quality ratio of actually using these tools in the wild.

The Tab:

Stitch - While the cheapest (S1: 3% of daily usage / S2: 20% / S3: 3%) in terms of credit budget and burn (and actually free), it produced the lowest quality outputs (7 / 10 / 8), therefore worst value. There's no world in which I would pay for this tool in its current form, but based on Google's generous AI bundling scheme, there's probably no world in which they would charge for it separately either.

Lovable - Cheap and good in the early phase. Recall that Lovable put up numbies for the zero-to-one scenario (16.5), burning only 10% of your monthly budget. This is the efficiency king for concept and 0-to-1.

v0 - The price of excellence (S2: 15 / S3: 17.5) is paid in dollars ($11 / $8.86). If you need high quality output, near zero-shot, v0 is your friend.

Claude Design - Anthropic's impressive entry inverts the cost equation. In dollars, each of my total-session-burn design outputs cost $0.50 or less at the time of writing (assuming 40 sessions/month). In opportunity cost, doing this midday is tragic.

This matters a lot to the indie dev (or hapless AI newsletter writer). For someone in a corporate gig like Ankush, cost is a non-factor. "We have budget for it... I'm not usually running close to the budget." Fair — it's an internal tool. Tokenmaxxing discussions aside, marginal inference cost should probably come into play for a well-funded PM when they're building an AI tool, not using one to do their job.

If you're pulling down $100/hr and spending $15/hr a few times a week to create prototypes, there's a reasonable chance you're doing it in the service of avoiding $10,000 endless-meeting-spiral collaboration snafus. Worth it.

Ankush, Luke, and my test results tended to agree on a lot. We collectively had a lot of belief in Claude Design, were surprised by v0, and generally were unsure/not optimistic about Google Stitch (some of the doubt on the latter came from inexperience with the tool rather than conviction upfront that it would be truly bad.)

Where we disagreed:

Ankush: "I would never have cared about state management. But I get why you would." My monetization PM is showing here. Scenario 2 was firmly in my lane. In fact, I've built cancellation, dive and save, and pause. These are boilerplate B2C streaming subscription flows. I instinctively know how they work together, so I am going to be the tools' harshest critic here.

Luke and Ankush both hit on the idea of shifting goalposts. Ankush: "I think you are evaluating production readiness and how quickly to get there, versus illustrating a concept." Luke: "It's closer to a real app than to a wireframe and so you grade it lower because it tried and didn't nail it."

Inherently, there is a bias introduced by implicitly scoring against "production readiness" vs. the "good enough for a peer stakeholder chat" reality that most of these outputs are used for in real life. And, in fairness, I also graded these scenarios on a near-zero-shot output. The field treats prompting as iterative and even strategically under-specifies to mine for ideas.

These are totally fair points. This lab is a point-in-time slice with a monetization PM-coded rubric, not a benchmark. Ultimately, my goal is to steer you in the right direction, not make the decision for you.

At this point, the floor is sticky, the bartenders look haggard, the jukebox is on its 21st play of "What's New, Pussycat." (For some of you, you've stuck with this series for four weeks. Hats off to you, your LinkedIn badge is in the mail.) You look up from your tablet with bloodshot eyes and mutter, with a gravelly voice: "What's the 'so what'?"

Experts?

Luke: "Claude Design is head and shoulders above the rest."

Ankush: Lovable = daily driver, because of the enterprise license and speed-to-concept. v0 is the surprise.

My takeaways for PMs jumping into this space:

Ultimately, the only right answer is the tool that answers this question in the affirmative: "does this give me what I need to walk out of the room with?"

Send me questions, feedback, and cocktail recipes:
[email protected]

_{Icons made by Icongeek26 from www.flaticon.com.}