Our company needed a public website. We had no in-house developer to spare, no agency budget I wanted to spend, and a VP of Operations who has never written production code. That last person is me. I shipped the site anyway, end to end, in June and July of 2026.
I want to be precise about what "shipped it myself" means, because I did not become a coder. I directed AI coding tools through every step: page structure, copy, design decisions, the terms of service and privacy pages, and the deploy to Vercel. My job was the same job I do everywhere else in the company. Define the outcome, set the standard, check the work, refuse to accept anything that misses.
The build itself was less dramatic than people expect. I described each page in plain language, reviewed what came back in a browser, and sent back specific corrections. Not "make it better." Specific: this heading is wrong, this claim is not one we make, this section belongs above that one. The legal pages went through the same loop. I read every line of the terms and privacy pages myself before they went anywhere near the live site, because a lawyer can review a draft but nobody else was going to catch a sentence that misdescribes what our company actually does.
The part that failed
The first drafts of the site copy were competent and dead. They read exactly like what they were: text generated by a model with no opinion, describing a fintech company the way a brochure describes any fintech company. Smooth sentences, no person behind them. If I had shipped that version, the site would have told visitors one true thing, which is that nobody at the company cared enough to write it.
Worse, the tools kept improvising facts about us. Company naming drifted between drafts. Descriptions of what we do got embellished in directions we never approved. Each individual output looked fine. Read three of them in a row and the seams showed.
The fix was not better prompting in the moment. It was writing rules once and making every tool follow them. I built an explicit voice document: sentence style, banned phrases, the register the copy has to hit, with examples of approved text to match against. I built entity-naming rules that state exactly which company names may appear in public copy and which may not, so the tools stop guessing. Then I went back through every drafted page and re-edited it against those rules before launch.
That rework cost me real time, and it was the most valuable part of the project. The rules now outlive the website. Every piece of copy any AI tool produces for us gets checked against the same document, and the drift problem has not come back.
What I verified before launch
Nothing went live on my word that it looked done. I loaded every page in the browser. I clicked every link. I read the legal pages top to bottom. I checked that the company was named correctly on every page, because that is the kind of error that survives ten reviews of "the design." Only after that did the site deploy to production on Vercel.
The takeaway I would hand any founder: the binding constraint on shipping with AI tools is your standards, not your technical skill. The tools will produce something plausible on the first try, and plausible is the trap. If you can write down what good looks like for your company, in enough detail that a stranger could apply it, you can direct these tools to real output. If you cannot write that down, the tools will happily ship your vagueness for you.
Write the standards document first. It feels like overhead. It is the actual work.