1. Home
  2. Work
  3. StudyfinderAI website

A website built by AI, checked by rules

StudyfinderAI sells pilots to schools and universities, careful public buyers. The site had to explain the product without promising more than it does. I led the relaunch with my co-founder. AI coding agents wrote the code. A claims register and automated tests decided what could ship.

Business
StudyfinderAI, co-founded
My role
Led the relaunch
Live since
10 August 2026
See it
studyfinderai.com
30

days from first commit to live

What it is: the time from the first line of code to the site going live.

Why it matters

AI coding agents wrote the code, and rules and tests decided what could ship.

129

commits up to launch

What it is: saved changes to the site's code before launch day.

Why it matters

each task got its own branch and agent, so every change stayed small and easy to check.

Step 02, One task, one branch
8

pages

What it is: the pages of the relaunched site.

Why it matters

every page was read once as a stranger would, before launch.

Step 04, Cold read
100

Lighthouse accessibility, mobile

What it is: the top score in Google's accessibility check, on a phone, in a lab run.

Why it matters

schools and universities are careful buyers. A site everyone can use is part of the trust. A critical finding fails the build.

Guardrails

Lab run on a local production build, 30 August 2026: performance 97 to 99, accessibility 100, SEO 100 on mobile. 0 axe violations on every page at two screen sizes. Initial JavaScript 114,5 kB gzipped, under a 130 kB budget.

How it was built

Eight steps, from rules to launch

  • AI
  • Fixed rules
  • A person decides

Step by step

What happened at each step

01Rules first

A person decides

  • A claims register came first: every public statement the site may make, and where it comes from.
  • A baseline document followed, built from research, our founder questionnaire and the brand assets.
  • The agents worked inside both.
02One task, one branch

AI

Each ticket got its own branch and its own AI coding agent. By November 2026, 30 agent branches were merged: 19 from Codex, 11 from Claude. Small, separate tasks keep every change easy to check.

03Claims check

Fixed rules, with a person auditing

  • All copy is checked against the claims register. A claim that isn't in it doesn't go live.
  • Planned features may only be worded as a pilot goal, never as something that exists.
  • A change to price, claims or data collection needs a written decision record.
04Cold read

Fixed checklist

Before launch, all eight pages were read as a stranger would read them: once, with no context. Each finding went on a tracker for the fixes.

05Automated checks

Fixed rules

Every pull request runs the same checks. One failure blocks the merge:

  • type check, build and pre-render;
  • link check;
  • performance budget;
  • contact form tests, 31 assertions;
  • browser tests and an accessibility scan of every page at two screen sizes.
06Merge bands

Fixed rules

  • Green: the agent ships the change on its own.
  • Amber: it ships and writes down why.
  • Red: the change waits for the owner.
07Launch

A person decides

A preview first. Then the domain switch, with the mail records protected so email kept working. Then a live check of every page.

08Running it

Fixed rules

  • Pages are served pre-rendered, so they load fast.
  • The contact form passes messages on and stores nothing. Its credential can send mail but can't read it.
  • Analytics only run after consent. Accept and reject carry equal weight.

Guardrails

What keeps it honest

  • No unregistered claims. If a sentence isn't in the claims register, it doesn't go live.
  • Accessibility is a gate. A critical finding on any page fails the build.
  • Speed is a gate. Going over the performance budget fails the build.
  • Least access. The contact form can send mail, never read it.
  • Fair consent. Reject is as easy as accept.

Built with

Stack, in plain words

  • React and TypeScript, built with Vite and styled with Tailwind.
  • A 3D scene that loads only when needed.
  • Hosting on Vercel, with one small server function for the form.
  • Playwright browser tests and axe accessibility scans.
  • AI coding agents: Codex and Claude.

What went wrong

What broke, and what I changed

  • Two days before launch, the homepage promised more than was true. It described one shared product and one live student journey. Neither existed. Now the first sentence says what the product is and where it runs, and nothing more.
  • Contact buttons only opened the visitor's mail app. No confirmation, and no fallback if it failed. I replaced it with a real form backend and a clear success message.
  • Wrong claims still got through after launch. Two days after go-live, the high-school page still had wrong lines about fees and the data agreement. I fixed them the same week.

For your business

What you can take from it

  1. Write the claims before the copy.

    Decide what you may promise, with a source for each line. Then AI can write fast without overpromising.

  2. Let agents ship behind gates.

    Automated checks and written risk levels let AI do the work while you keep control of what matters.

  3. Protect mail when you move a domain.

    A website launch should never take your email down. Plan the DNS change around it.

Sources: the project's QA log, claims register, work log and git history, read 2026-11-16. The numbers describe the build, not StudyfinderAI product results.

Contact

Tell me what is stuck

Book 30 minutes on video, or write me a few lines about which AI tools your team uses and where it gets stuck.

Book a 30-minute call

consulting@mario-behrends.com
LinkedInEnglish and German.