We Let Claude Build Our Entire Website — Here's What Happened
The website you're reading right now was built entirely by Claude Code. 61 source files. Roughly 3,000 lines of TypeScript, MDX, and config. 25 commits over three days, March 5 through 7, 2026. I wrote none of it.
That's the headline. Here's the honest version.
What We Actually Did
"We let AI build our website" sounds like we kicked back and watched. We didn't. We spent weeks on the inputs before Claude touched anything, and that turned out to be the entire game.
Before a single file was generated, we'd produced a design brief with 11 brand colors and three self-hosted font families. A PRD at version 2.0 covering every user story and data model. An implementation roadmap with sprint-level detail. HTML blueprints for the homepage, blog feed, and blog post layouts. And a CLAUDE.md file listing locked architectural decisions. Things like "Tailwind + design tokens over CSS-in-JS" and "static generation for content pages." Claude was not allowed to deviate from any of them.
That last part mattered more than anything else. Every successful Claude Code story I've come across, including Anthropic's own disclosure that 90% of Claude Code's code is written by Claude Code, comes down to the same thing. The documentation you feed it is the productivity multiplier. Not the AI itself.
Google principal engineer Jaana Dogan shared that Claude Code generated a distributed agent orchestrator in 60 minutes. Her team had spent a year building one. But I guarantee that prompt wasn't "build me a distributed agent system." It was backed by deep context about what the system needed to do.
What Went Right
Scaffolding was flawless. Project setup, Tailwind config, Next.js App Router structure, dependency management. All nailed on the first pass. The kind of boilerplate that eats a developer's entire first day. Claude did it in minutes.
Design token extraction was the thing that surprised me most. We handed it HTML blueprints with our color palette and typography, and it pulled the exact values into Tailwind's @theme configuration. No drift. No interpretation errors. I genuinely wasn't expecting that level of precision.
TDD with AI works better than TDD without it. When I told Claude to write tests first, then implement, the failure rate dropped noticeably. The constraint forced precision. Left to its own devices (and this is a consistent pattern) Claude optimizes for getting something working. Not for getting it right.
SEO generation was nearly perfect. Sitemap, robots.txt, JSON-LD structured data, OpenGraph images. The invisible infrastructure that's tedious to build and easy to forget. Claude didn't forget any of it.
What Went Wrong
Creative decisions need a human. Claude will give you a competent layout every time. But competent isn't interesting. The homepage it generated was technically correct and completely forgettable. Every design choice defaulted to the safest option. I had to push back on nearly every visual decision that involved taste rather than specification. If you've ever used a website and thought "this looks fine but I can't remember a single thing about it," that's what Claude produces by default.
Long context windows don't replace understanding. Around commit 15, Claude started losing the thread. It would reference patterns from earlier in the session that it had already refactored away. The context window is large, but it's memory. It's not comprehension. Those are different things.
Debugging is where it struggles most. When something broke, Claude's instinct was to add code. Another wrapper, another null check, another fallback. The fix was almost always to remove something. AI confidently writes broken code and tells you it's working. You need to actually verify. Every time.
A METR study found that experienced developers estimated AI made them 20% faster, even when they were objectively 19% slower. I felt faster the whole time. Whether I actually was? No idea. I keep thinking about that gap between perception and reality. It's uncomfortable and I don't have a neat answer for it.
The Numbers
42% of committed code globally is now AI-assisted. That's the current number, not a prediction. But AI-generated code produces 1.75x more logic errors than human-written code. 48% of it contains security vulnerabilities. The Sanity engineering blog put it bluntly: the first attempt is 95% garbage. Experienced engineers iterate three or more times on AI output before it's production-ready.
We iterated. A lot. The 25 commits in the git log don't tell you about the dozens of outputs we rejected, the design choices we overrode, the moments where Claude confidently presented something broken as finished.
The Verdict
Claude Code is the best implementation partner I've ever worked with. It executes specifications with remarkable speed. It cannot tell you what to build or why it matters. The gap between those two things is vast and I don't think it's closing anytime soon.
The website works. It's fast. Lighthouse scores are above 90. The MDX pipeline renders content cleanly. The design system is consistent. Claude did all of that.
But the voice you're reading right now? The decisions about what ByteMelt should feel like, what it should say, what to leave out? Those are still ours. I suspect they will be for a while.