Gemini 3.7 Flash Review: One Prompt, One Full HVAC Website

Google shipped Gemini 3.7 Flash this morning, and it claims significantly improved coding. So I ran the test I run on every new model: one big prompt, one complete local-business website, no follow-ups, no hand-holding.

The build was a full site for a fictional HVAC company — Comfort Pro Heating & Air Conditioning — in Antigravity, using Next.js, TypeScript, Tailwind and shadcn/ui. It ran for about five and a half minutes. I scored it 7–8 out of 10 on the first prompt.

The prompt

This is the whole input. One message, no follow-ups:

Hey, let's build a complete modern website for an HVAC company called Comfort Pro Heating and Air Conditioning. Tech stack, let's use Next.js, TypeScript, Tailwind, and shadcn for any forms. The company provides AC repair and installation, furnace repair and installation, heating, cooling, and 24-hour service. Now also create a high-converting professional design to generate phone calls and service requests. Include a hero section, services section, why us, customer reviews, service area, financing, FAQ, and a strong call to action with strong headers and footers. Let's make it look real premium for a local HVAC, and I don't want it to look like AI slop. Use imagery that you can create, make a full responsive website, and continue to go until it's done.

What happened, in order

It wrote its own plan

Before touching a file it produced an implementation plan without being asked. I use Claude Code for most of my work and I plan everything there, so seeing Flash do it unprompted was a good sign.

It built the component tree

Global CSS, layout, then Header.tsx, Hero.tsx, EmergencyBanner.tsx, ServiceDetailModal.tsx, ServicesGrid.tsx, CostEstimator.tsx, WhyChooseUs.tsx, SpecialOffers.tsx, ServiceArea.tsx and the rest — a few hundred lines each.

It broke, then fixed itself

The first run threw a syntax error. It traced the problem to the WhyChooseUs.tsx component and repaired it without my input.

It reviewed its own work

It spun up the dev server and took screenshots through my Puppeteer MCP, then walked its own sections checking them over. That part genuinely impressed me. Total time: about five and a half minutes from prompt to a running site.

What it got right

• The quote form is built into the hero. I didn't specify placement. Putting the service request above the fold is exactly what converts on a trades site.
• A working cost estimator — functional, and a nice touch I didn't ask for by name.
• A financing section. I did ask for this, and it delivered — flexible monthly options, presented well.
• Every requested section shipped. Services, why us, reviews, service area, FAQ, footer. Nothing dropped.
• A clean mobile menu, with a book-a-call action right in it.
• The forms and modals are real shadcn. The booking popup is clean.
• It caught its own syntax error and fixed it.

What it got wrong

• The header overlaps and runs to two rows. Cluttered, and the first thing you see.
• It won't break responsively. The header doesn't collapse fast enough as the viewport narrows. In my experience most models except Claude struggle here, and Flash is no exception.
• Placeholder fake reviews. Real testimonials or clearly-marked placeholders would be better than invented ones.
• The same form on every call-to-action. Every button opens an identical modal. Different CTAs should capture different intent.
• Stock imagery, not generated. I asked it to use imagery it could create and it appears to have sourced existing images instead. Partly my fault — I should have said "generate your own AI images" explicitly.
• A few sections read crowded, and the icons don't quite land, though they do match their headings.

The verdict: 7–8 out of 10

On a single prompt, that's a solid result. Everything I asked for is on the page, it self-corrected, and it checked its own work — for a five-and-a-half-minute build with zero follow-ups, that's real progress.

It is still not, in my opinion, where Anthropic or OpenAI are. Those are a little bit ahead, particularly on responsive layout. But the gap is narrower than it was.

The header and responsiveness issues are a second prompt's work. If you're building production sites with this, plan on a review pass — which you should be doing anyway.

Should you use it for client work?

If you build for trades businesses — electricians, plumbers, HVAC — Flash will get you to a credible first draft fast. It will not get you to a shippable site without a human pass over the layout, the copy, and anything that claims to be a real customer review. That's the honest state of it as of today.

If you'd rather skip the review pass entirely, that's what we do — see custom CRMs for HVAC contractors or read our take on building vs. buying field service software.