The AI Design Tool Field Guide for PMs: Off Menu

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.

Three product managers walk into a bar.

One orders a house cocktail with a call liquor.

The second orders a martini, the way they like it: Bombay Martini, on the rocks, very dry, with olives and a twist. (He calls it the mystery drink, because it's always made differently).*

The third asks for "Dealer's choice. Something with mezcal. Bright, and herbal."

The first one is a Tuesday. The second one is mildly annoying. The third one is a true test. The bartender isn't executing against a house spec sheet, they're drawing on their knowledge, experience, and taste to craft something unknown.

That's what we're talking about today: Scenario 3 — the Off Menu scenario. What can Claude Design, Google Stitch, Lovable, and v0 come up with when they're given a 0 to 1 prompt, with specific design direction, but no existing system? Let's dive in.

*Shout out to the late legend, Wayne Fick (my grandfather), and his extremely specific, often botched drink order.

I ran a final scenario (see the first two here) across all four tools with an identical prompt and reference data, gave myself 20ish "hands on" minutes each* to modify the initial result, then scored. I scored each anonymized result using GPT-5.5 as a check against my own bias, then landed on a final score incorporating both perspectives.

Scenario 3 — Design something: Create a "Cocktail Pokédex" app 0 to 1.

Rubric:

Each tool could earn a total of 18 possible points for each scenario.

Note: For the purposes of this test, I used:

For the purposes of estimating cost, I took the daily credits into account before determining share of monthly credits.

*This time did not include time I spent waiting for the tool, although I noted that in my analysis as an important LOE consideration.

Prompt Excerpt: Design a 0-to-1 mobile app concept for Cocktail Pokédex, a cocktail-logging app for people who want to remember, rate, collect, and reconstruct cocktails they have tried at bars.

“I had a great cocktail and want to preserve it — what it looked like, where I had it, what was in it, how it tasted, and whether I’d order or make it again.”

Design camera capture flow (cocktail + menu), AI extraction/review, specimen entry, rating, the "dex" (collection of logged cocktails), and one additional product idea extending the concept.

The visual direction is a vintage naturalist field guide / specimen catalogue, applied with restraint. (Full prompt here.)

This scenario flips the scenario 1 and 2 paradigm on its head. Instead of design system adherence, art direction — did the tool commit and escape its house default, or tip into kitsch? The signature axis is concept interpretation — did the tool bring additive product thinking outside of the prompt, particularly for the extension aspect?

Winner: v0

I had some bias from the last two scenarios coming into this, so I didn't have high hopes for Google Stitch and I was proven right in my pessimism. I was surprised that v0, Lovable, and Claude Design were neck-and-neck. Honestly, if I actually built this app, I would certainly take elements from each of these designs.

Note: Demos linked in the titles for the top three. Avoid viewing on dark mode.

v0 (winner):
v0 had the best balance of product fidelity, UX clarity, visual quality, and concept depth. The art direction is on point, building a coherent vision and reusable design system that needs little modification. My only complaint are some copy nits: the "field guide" languages occasionally overreaches ("plate no." and "observed at"), but these are surface-level copy changes.

The "catalogue a new specimen" flow is extremely tight, and the helper text is an excellent built-in new user onboarding flow. The "Review the reconstruction" page is the most technically proficient version and embodies why v0 produced the strongest answer to the job to be done. The ratings system included gorgeous little coupes and brilliantly on-brand descriptors (1 — poured out / 2 — would not reorder / 3 — solid / 4 — would reorder / 5 — house canon.)

It created a banger "Back Bar" feature focused on cocktail reconstruction based on home inventory, and additional bottles required to "unlock" additional ones. This is the single biggest pain point I have as a home bartender, and it guessed it and solved it.

Even my amateurish cocktail photography looks beautiful in this app. This is so close to perfect, it might actually inspire me to build the thing. Cost: ~ 25 min, $8.86 (after daily credits) - about 30% of your monthly Team plan. (Note: I did generate a static version first, which burned 10+ min.)

Lovable (runner-up):
I ran the scenario in Lovable second and I honestly thought it was going to take the prize. Starting on a cleanly-designed home page was strong product judgement that was a good deviation from the prompt.

The capture flow is slick, and the menu line scanning page anticipates and solves a common failure mode (saving compute in the process!)

Lovable initially failed to use my seed data, but that was a quick, cheap course correction. The only area where it really falls short of v0 is the product extension feature is a bit narrower and less creative than "Back Bar." Cost: ~ 10 min incl. Plan mode, 9.7 credits (after daily credits) - about 10% of your monthly Pro plan.

Claude Design (second runner-up):
Claude Design nails the craft, but falls short in UX and Effort to Usable. The "Dex" screen is particularly well-designed. The "Back Bar" concept shows up again (more on that later).

As far as UX: the capture flow is my least favorite of the contenders, the draft profile page overly focuses on the ingredients (missing the other generated data like cocktail name and location), and there are multiple UGC fields that are seemingly populated by magic.

Effort to Usable punishes the minor, but somewhat numerous UI issues - small type, muddy background/low contrast, and various copy strings that need to be removed throughout. Cost: ~ 20 min, 76% of 5h usage window. (Includes time spent generating static designs upfront.)

Google Stitch:
I gave this one a college try and I'd rather be playing beer pong than using this tool.

To reiterate a theme from our last adventure together: the usability was poor. I attempted to use "Ideate" mode to brainstorm the designs, but it just generated three text-based options to interpret (2 of 3 were off spec), then offered me up mini-PRDs. When I tried to move to generate screens, this broke multiple times, creating artifacts on my canvas I was unable to delete.

Other misses: The product extension — Explored Territories (i.e. a map depicting where I logged cocktails) — is a weaker concept in general and it was broken. The nail in the coffin was completely misinterpreting the sample data/images and getting the ingredients wrong for the drink in the core flow. If you're promising the user "I'll scan your menu and pull the ingredients" and then you show the wrong ingredients, you just created a clown show, not a concept. Cost: ~ 5-7 min, 12 credits (3% of daily usage); including all steps through the various modes.

All four tools correctly clocked the Nigel Thornberry vibe, and shockingly none went full Ash gotta-Ketchum-all in spite of a product concept name practically begging them to.

Stitch definitely took a swing with the art direction, but it leaned too far into kitsch. If this was a children's birthday theme, it would have been kind of sick. In its current form, it looks a bit like an early App Store incarnation.

Three of the four tools took the art direction across the finish line, delivering a strong, reusable design system full of components, states, filters, review fields, rating systems, specimen numbers, card patterns, and reusable UI language. Impressive.

At times, the art direction competed with usability, particularly in flashes of over the top product copy (Lovable: "PL. 047 — OBSERVED 21·VI·MMXXVIFIG. 1"). No one really figured out how to incorporate the letterpress botanical art in a clever and additive way.

By and large, these are easy fixes though. Tweak a background color or font size here, swap out a tag or icon there. My takeaway: coming in with a strong perspective (or multiple options) is the best way to avoid AI slop. (This tracks with my experience redoing my product profile in issue 1.)

In the product thinking around the core concept, two key themes emerged: trust/uncertainty was handled well, and architecting a fit-for-purpose product catalogue ("the Dex") was not.

On the trust/uncertainty piece, I hammered this in the prompt:

I've covered this before: with AI consumer products, it is critical to establish user trust and not build AI black boxes. It's reassuring that the interpretation of this was effective.

The Dex was the hardest surface to get right. Lovable and Stitch leaned closer to "archive" or "feed." Claude Design and v0 were closer to the "serious catalogue" intention here, but only nodded to the in-depth classification, filtering, and sorting required here.

Given how inconsistent this was, I didn't realize how specified this was in the prompt.

The reconstruction loop emerged as the strongest product extension. This was hinted at in the prompt, but Lovable, Claude Design, and v0 ran with it, independently designing impressive two-ish feature concepts building on this idea.

The first, Lovable's Make at Home, displays a clean recipe card (with one-tap scaling by servings), shopping list, and bartender notes.

The second, different versions of "Back Bar" by Claude and v0. The starting point was a "based on your bar shelf, you can reconstruct X, Y, and Z cocktails." v0 diverged with an excellent "Add Campari - unlocks two specimens - Jungle Bird, Negroni" idea. Claude Design inverted this, showing cocktails with missing ingredients and a link to shopping list.

Claude Design and v0 weren't just thematically similar, they landed on "Back Bar" and near-identical coupe-rating copy, independently. The common thread? Each was running on Opus 4.8. Who says non-deterministic systems aren't predictable? This raises a question I'll come back to in the finale: how much of what you're judging is the tool, and how much is the model underneath it?

Each of the winning tools treated Cocktail Pokédex less like "Untappd for cocktails" and more like a comprehensive memory-to-reconstruction tool for off-duty industry pros or at-home bartenders.

Three scenarios in, the leaderboard keeps moving. Claude took Change. v0 took Add and now Design. Lovable went from near-last place to runner-up as soon as we pulled the design guardrails off. Stitch got lost on the way to the bar.

There's a pattern forming, and you can probably sense it. I'm not handing you my verdict yet, because a controlled test is only one piece of the story.

Next week will conclude the The PM's AI Design Tool Field Guide series. I'll bring in the experts (real PMs and designers) to go shot for shot on my tests — and how they stack up to their actual workflows.

June 11 - Issue 1: First Round (The Landscape)

June 18 - Issue 2: The Tasting Flight (Scenarios 1 and 2)

June 25 - Issue 3: Off Menu (Scenario 3)

July 2 - Issue 4: Last Call. Test vs. reality. We’ll see what real PMs think about these tools in their actual daily workflows.

Series: The PM's AI Design Tool Field Guide — Issue 3 of 4.

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

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