Tech / AI

AI-assisted web development: how this site is built

Two languages, light and dark, from phone to large monitor: I build this site with Claude Code and a team of AI agents. How it is set up, and where the limit is.

14 Nov 2026
Contents · 5 sections

My own websites used to stop at 80 percent. A WordPress template looked quite nice, but as soon as I wanted something different, it didn’t work.

This site has two languages, a light and a dark mode, and adapts from phone to large monitor. I have been building it with AI since March, first with OpenAI Codex, since May with Claude Code. By now a small team of AI agents with fixed roles works on it: one plans with me, others research, implement, review texts and test every change before it goes live.

The 80 percent limit has not gone away. It has moved, to me. I have to supply what goes on the site, and I have to check what came out of it. This is how the team is set up, and where exactly the limit runs today.

120+ pages in two languages
8,700 image variants, from phone to large monitor
6 reviewers for every article
5 automated checks before anything goes live

New terms are explained in the In plain words boxes, Dive deeper is for anyone who wants the details. If you only want to know how the work is split between me and the AI, read up to Under the hood.

Content first, then code

This site started on 20 March. The next day produced six planning documents before there was any site scaffold at all: who should find what here, how photography and tech sit side by side, which pages are needed, what the design could look like and which tech stack to build on. Only then came the first line of code.

Anyone who has built websites for others knows this: clients want a website but don’t know what should be on it. I was at that point too. An AI writes code in minutes, but it cannot tell me what I want to show. Picking images, writing down experiences, deciding what the site stands for: that stays my job. Skip it and you get AI slop, text that only reheats what is already on the web.

The storyline is mine, the wording no longer has to be

I’m no spelling pro and no genius with words. My father is a German teacher and used to proofread everything I wrote. I love packing knowledge into a good storyline. Writing it out properly I hated, and that is why for a long time I shared what I knew less often than I could have: preparing it was always extra work. Keeping a German and an English version in parallel is something I would never have kept up on my own.

Today I supply the raw material: what story a text should tell, what readers should know afterwards, which facts and experiences I can add. Often I simply dictate it. Claude first turns it into a build plan, asks where something is missing, and only writes once I have approved the plan. The English version is then written fresh from the same core, not translated sentence by sentence.

When it comes to checking, the split is uneven: I read every text, but code only in samples. That is the difference from what people call vibe coding, accepting code without looking at it. My code is checked by agents and automated tests, and whatever they don’t know about, I see on the page, or not at all.

From here on it’s about how the team behind this is set up.


Under the hood

I started with OpenAI Codex and little setup around it, basically a chat window. Codex was strong at writing and weaker at code. Since my other projects are more code-heavy, I switched to Claude Code in May. In August a writing skill, a review board and checking agents were added, in November the automated pipeline and a fixed division of labour. Today it looks like this:

I decide, the orchestrator plans, workers deliver, and the checks run on their own.
Who does what Top to bottom: I provide content, judgement and sign-off. Planning runs in both directions between me and the orchestrator (Opus). The orchestrator writes briefs for three workers (all Sonnet): a researcher for facts with sources, an implementer for code to spec, a verifier for five checks. At most three run at once. Below that a second strand for texts: writing skill, review board with six seats, humanizer, then I read it myself. Me Content, judgement, sign-off Orchestrator Opus plans with me, writes briefs Researcher Facts with sources Sonnet Implementer Code to spec Sonnet Verifier 5 checks Sonnet at most 3 at once The strand for texts Writing skill Review board · 6 seats Humanizer I read
In plain words

An AI agent gets a task and works through it on its own: it reads files, writes code, runs programs and keeps going until the task is done. In Claude Code, a skill is a saved routine in a text file. Instead of explaining every time how I want texts to be, it is written down once, and Claude follows it.

Opus plans with me, Sonnet does the work

Anthropic offers its models in several sizes. Opus is the strongest and most expensive, Sonnet the faster all-rounder. In my setup only Opus talks to me. It works out what I want, makes the decisions with me and writes briefs. Anything that means a lot of reading or typing to a clear spec goes to workers on Sonnet, at most three at once: a researcher who collects facts, an implementer who changes code to a brief, and a verifier who runs the tests.

An example from this article: how images are delivered, how the colour mode switches and what each check catches was collected by a researcher, with the path of the source file for every statement. Opus wrote from that and looked up the most important numbers again itself.

So that not every agent reads everything at once, there are rules that only load when they are needed. Claude reads the project file CLAUDE.md at every start: how the site is built, which commands exist, what must never happen, such as putting anything live without my go. The style rules for texts, on the other hand, only load when Claude touches a text, the pricing rules only at the pricing engine. Why that matters I explained in the travel article with the picture of the desk.

Six reviewers for every text

Every article goes through the same chain. A writing skill guides me through build plan and draft. Then a review board reads the text, six agents with one rubric each: readability, what can be cut, positioning, technical credibility, search intent and dramaturgy, meaning whether the structure holds. Each seat sees only its rubric and the text, never the other seats’ verdicts. The board never changes anything itself. It reports, and I decide on each finding.

It had quite a bit to say about the first draft of this article. Five of six seats found one sentence in the go-live section unnecessary, and it is gone. Dramaturgy found the technical part too much like an inventory list, so it is now built around the team. The cutting seat wanted the sections on images and graphics removed entirely. I shortened them and kept them, because they show what would not exist without AI. And the positioning seat found my line about spelling too modest. That one stays.

At the very end every text goes through the humanizer, a freely available skill that finds typical AI patterns: inflated wording, lists of three on principle, sentences that only sound good. To match my voice it gets a few paragraphs of mine as a sample. Then I read.

Dive deeper

Every seat is the same agent with a single tool, reading files. What it checks is in a rubric file that the brief hands to it. The dramaturgy seat asks, for example: do the first sentences carry a concrete result, number or scene instead of a definition? Does a reader who stops after the first third have the core message? Do the headings alone carry the argument? Does the conclusion come before the FAQ? On request a seventh seat reads the text like a stranger, once as a layperson and once as a practitioner. Already published articles come back from the board with a verdict: keep, rework, cut hard or start over.

Nothing known slips through

The dangerous changes on a site that an AI helps build are the ones that are right in one place and break something in another. Or that land on the German page and not on the English one. So every change has to pass five checks:

Nothing goes live until all five are green. Locally and on GitHub, every time.
Five checks A change passes five checks in a row: type check (does the code fit together), language parity (DE and EN complete), price regression (prices unchanged), build (do all pages build), link check (no dead links). When all are green, my sign-off follows, then the site goes live. The same five checks run twice: locally in the verifier agent and on GitHub in CI. Change 1 Type check does the code fit together? 2 Language parity DE and EN complete? 3 Price regression prices unchanged? 4 Build do all pages build? 5 Link check no dead links? all green my sign-off live runs twice locally: verifier GitHub: CI
  1. Type check: finds code that doesn’t fit together, for example a component called with the wrong props.
  2. Language parity: every German page needs its English counterpart, every article its twin and every page an entry in the page inventory. Exceptions sit on a short, explicit list.
  3. Price regression: the pricing calculators are checked field by field against an independent reference calculation.
  4. Build: all pages are built. Broken article metadata and broken imports show up here.
  5. Link check: every internal link, every language switch and every hint for search engines must point to a page that actually exists.

These five steps run twice. First locally: before Claude reports anything as done, the verifier runs all five, and it only checks, it never fixes. Then again on GitHub, as soon as a change is up for merging. It has become rare for the second round to find anything. In June it was still red: the type check had found 13 errors, all around the way images are described in the code. Back then the local verifier didn’t exist yet.

Only what I have signed off goes live. After that everything runs automatically: checks, build, and only when everything is green does a separate step upload the finished files to the server. That step looks after itself:

  • It logs in with a key; there is no password stored anywhere.
  • It checks that the target really holds this site and that the build is complete.
  • It never deletes more than a fixed number of files in one run and leaves files that only live on the server alone.
  • Started by hand, it first only does a dry run and shows what it would upload and delete.

So nothing that one of the checks knows about can slip through. What none of them knows about comes further down.

In plain words

A branch is a working copy next to the main version. Work is built and checked there without the live site noticing. A pull request is the request to take that working copy into the main version; taking it in is called a merge. With every such request GitHub starts the checks automatically. This is called continuous integration, CI for short.

Dive deeper

The price regression is a so-called golden master test. The check script holds the old pricing formula as its own, independent calculation. For a fixed set of wedding and portrait cases in every pricing mode, the script and the site’s engine calculate in parallel, and every field has to match. That way Claude can rebuild the engine without a price quietly shifting.

The link check runs the tool lychee over the built pages, entirely offline. Even addresses that point to the site’s own domain, such as the language hints for search engines, are mapped to the built files and checked too.

What the team built that I would not have built alone

The site is built with Astro. Astro generates every page in advance as a finished HTML file. No server computes anything when a page is opened, and the only JavaScript of my own running in the browser is a small script for colour mode, menu and table of contents. The pricing calculators are the only interactive parts.

Images at the right size. Every photo goes through one image component. It produces the photo in several widths as AVIF and WebP and only loads it when it scrolls into view. The phone gets the small version, the large monitor the large one. Across the whole site that is around 8,700 image variants and about 1.5 gigabytes. Without a cache, a build in the pipeline once took an hour and a half; with the cache only new images get processed.

One graphic, four versions. Every graphic in my articles exists in German and English, light and dark. That sounds like four files per graphic. It is one: a small component with a drawing written as code, an SVG, and the labels in both languages. The colours are not in the drawing; they come from the same colour variables as the rest of the site. When the mode switches, the graphic changes with it. A changed label is one line of text, not four new images.

The same graphic in two languages and two colour schemes. I only have to build it once.
One graphic, four versions Four identical small diagrams in a two by two grid: text, check, live. Top left German in the light scheme, top right English light, bottom left German in the dark scheme, bottom right English dark. German · light Text Prüfung live English · light Text Check live German · dark Text Prüfung live English · dark Text Check live

Two languages that match. The language switch leads to the translated page. If there is none yet, you land on the overview of the other language instead of an error page.

Dive deeper

Up to 1920 pixels wide the base font stays at 16 pixels; above that the whole page grows with the window instead of sitting as a narrow strip in the middle. It is a single rule. Almost all sizes are set in rem, relative to the base font, and that grows with the window width:

html {
  font-size: clamp(100%, calc(0.263vw + 10.95px), 131.25%);
}

At 3440 pixels that gives 20 pixels, capped at 21. Type, spacing and the maximum page width grow together.

Where the AI got it wrong

Three mistakes from the past months, and all three looked right at first glance.

Umlauts. The German texts of the portrait pricing calculator said “fuer”, “groesser” and “Verfuegbarkeit”. The AI had written German with ae, oe and ue instead of real umlauts. It was cleaned up in June, in two passes over almost 90 lines. Since then the rules say: visible text always uses real umlauts.

A rule that didn’t apply. This week a rule was supposed to cap the graphics at a moderate width. It was in the stylesheet, it just didn’t take effect. Only a second attempt made it work. None of the five checks notices something like that. You only see it on the page.

A safeguard that only looked good. The first safeguard for going live looked convincing in the code and would have done nothing at my host. It was replaced the same day by the guards described above.

The limit is with me now

The 80 percent limit used to be technical: the template couldn’t do more. Today it is with me, in three places. I have to supply the content; no agent knows what I want to show. Design remains a gap, because neither Claude nor I are designers. And consistency: a change that isn’t carried through to every page, or one with side effects somewhere else. The checks help with part of that. For the rest, the only thing that helps is looking at the site, on the phone and on the monitor, light and dark.

The bigger gain for me isn’t the technology, though. Writing things out was for a long time the reason I shared what I knew less often. Today it is no longer the bottleneck, and this site is the result.

FAQ

Do I need to be able to program?

Not for a simple site. For a setup like this one it helps: Git, a command line and enough understanding to judge a sample of code. Most of the work is still in the content.

Was this text written by an AI?

Co-written. Storyline, facts and experiences come from me, Claude did the wording, six agents and I reviewed it. Every sentence here has been through my hands.

Can I rebuild the setup?

The building blocks are freely available: Claude Code, Astro and the humanizer skill. My rules and rubrics are tailored to this site. A sensible start is a short CLAUDE.md, one automated check for the mistake that annoys you most, and a skill for whatever you explain to the AI most often.

More on the topic: how I plan my trips with AI agents is in AI agents instead of a travel agent. More AI projects are in the AI section.

Work in progress