The brief
Two-person team. A real product that wasn't ready to demo publicly yet, but a tier-one VC about to write a check, and a deck deadline two weeks out. They needed a website that would survive a serious investor pasting the URL into a Slack channel. It had to feel built, not generated. Original art. Real copy. Fast. Crawlable. Hand-tuned.
And they had no money for an agency.
The team shape
One Intellilabs senior front-end engineer. Claude Code as a constant, sitting-next-to-you collaborator. Founder available for a 30-minute review every morning. That was the whole team.
This is the part that gets overdescribed, so we'll keep it short. Claude was not "generating the website". Claude was a second pair of hands on a real codebase, with a real human deciding what was a good idea and what wasn't. The senior engineer worked in their IDE the same way they always do. The difference was who they were pairing with.
"It was the first time AI didn't feel like a party trick on a real project. It felt like an extremely well-read junior who never gets tired." — Intellilabs engineer
What we built
- Next.js 15 + App Router, deployed to Vercel. Static where possible, ISR where useful.
- Sanity CMS for content the founder needed to edit weekly — blog, changelog, pricing.
- Tailwind on a custom design token set. No off-the-shelf UI kit. The site looks like the product, not like a template.
- An animated SVG illustration system for the hero and section dividers, hand-drawn by Claude based on rough sketches and iterated in Figma-then-code.
- Full technical SEO — sitemap, structured data, OG generation, Lighthouse 100 on the homepage.
The eleven days, honestly
Days 1–2: shape and tokens
Half a day on a moodboard. Half a day defining colour, type and spacing tokens. The rest writing components. Claude wrote about 70% of the first-draft component code; the engineer rewrote about half of it before commit. That ratio held for the rest of the project.
Days 3–6: pages and content
Home, product, pricing, blog template, contact. Real copy, drafted by Claude against a writing-style brief, edited live by the founder over a shared doc. The pattern that worked: Claude drafted, the engineer placed it into the page, the founder broke it on Loom, Claude rewrote, the engineer reshipped. Three loops per page, average.
Days 7–8: illustrations and animation
This was the part everyone was nervous about. It turned out fine — but only because we treated the AI as a junior illustrator working under direction, not a magic image button. The senior sketched a concept, asked Claude to render it as SVG, refined the SVG by hand, then asked Claude to add the animation. Most of the visual polish in the final site came from human eyes catching that something was 4px off.
Days 9–10: performance and accessibility
The boring part. Image budgets, font subsetting, lazy components, ARIA passes, keyboard nav, reduced-motion fallbacks. Claude wrote most of the test cases. The engineer ran them, fixed what failed, and re-shipped.
Day 11: ship
Deploy to a real domain. Final Lighthouse run. OG image regeneration. Cache warmup. Pasted in the investor's Slack at 18:42.
The outcome
The investor signed. More importantly, the founder could keep the site moving. By month two they'd shipped twelve new blog posts and three product pages themselves, with Claude Code helping where they got stuck. The site stopped being a deliverable and started being a thing they owned.
What we learned
Claude is a collaborator, not a contractor.
The moment we tried to hand it a self-contained task and walk away — "build the pricing page" — the output got generic, the design got mid, the code got fragile. The moment we sat next to it and treated it like a teammate with strong opinions and shallow context, it was extraordinary. The job of a senior didn't shrink. It changed.
Taste is the bottleneck.
A founder without taste using Claude builds a website that looks like every other AI-built website. A senior designer with taste using Claude builds something that looks hand-crafted. The model amplifies whoever's driving. We've stopped pretending otherwise.
Spend the time you saved on the boring stuff.
The day we "saved" by not hand-writing boilerplate was the day we got to spend on illustration, accessibility, and copy. That's the real productivity story — not raw speed, but a redistribution of effort towards the things that actually make a site good.
The one thing to take away
AI-assisted web dev isn't fewer engineers shipping the same thing. It's the same engineers shipping more thoughtful things, on shorter timelines, with the boring parts handled. The cost floor moved. The taste ceiling didn't.