This page in plain text
How this site is built
With WordPress, Claude Code and one habit: nothing is finished until someone has looked at it. This page says what the site runs on, the rules we work by, what went wrong along the way and what we changed because of it.
Counted on 21 November 2026, three days after the first change was recorded: 56 pages built as presentations, 23 check scripts and 50 recorded changes.
What it runs on
- WordPress, with its own plain default theme, kept for speed. The colours, the header and the footer are files we can read and put back.
- One plugin of our own. It shows a page inside a frame and it runs the chat. It began as a copy of a plugin we wrote for another site, and it has no build step.
- Claude, through AI Engine. The chat on every page is Claude. AI Engine is the WordPress plugin that connects the two, so there is no server of our own to run.
- Claude Code, the coding agent that did the building. We say what a page is for and look at what comes back.
How one page is made
- A presentation. The part you look at: one page with its own layout and motion, shown in a frame. It can be opened, tested and replaced without touching the rest of the site.
- The same content as plain text, which is what you are reading. A search engine sees no words in a frame, and neither does the chat. The text is what both read.
- A chat that reads that text. It answers from the page and is told to say so when the page does not cover the question. A guest gets 30 free questions on the smallest Claude model; a member’s run on the next one up.
- A check. A script that opens the finished page at phone and desktop width and asserts what should be true.
Which model the chat runs on came from a test: ten questions put to three sizes of Claude, with the cost and the time of each. The figures are on the model guide.
The rules we work by
Six of the lines the agent reads at the start of every session. They sit in a plain text file in the project and came with us from an earlier one. The file grows: a lesson from this build is added as a new line or a new check.
- Render it and look at it. Do not reason about how it will look.
- Measure, then infer, and label which is which. A fault report gives the value before it offers a cause.
- A check asserts on the answer, never on the absence of other output.
- Check every write by reading it back, never by trusting the word “updated”.
- Anything done by hand twice becomes a script or a check. Twice, not three times.
- A lesson is not learned until it is written down, in a rule or a check. Prefer a check: it cannot be skipped on a bad day.
Our own mistakes
The ones that cost most were ours, not the AI’s. Working with coding agents over time, these four cost us the most rebuilding, and each one we fixed by changing how we work, not by finding a better model.
- Too few design documents. We described what we wanted in a sentence and let the agent fill in the rest. It filled it in with guesses.
- No test plan. Nobody had written down what “working” meant, so there was nothing to hold the result against.
- No separate QA. The same session that built the thing also judged it.
- Too little use of goals and skills. We explained the same job again each time, where we could have written it down once and stated the goal once.
How we work now
- A design document. What it is for, who uses it and what it must not change, written before any building.
- A design checklist. The document turned into numbered outcomes that can each be ticked or not.
- The build. An agent works to the checklist, with the goal stated once. Skills carry the jobs we repeat.
- A separate QA checklist. What will be tested and how, written apart from the build.
- A separate QA run. A fresh session runs the checks and reports clean, not clean or blocked, with the evidence.
We keep adding skills and keep rewriting the rules. Every mistake becomes a line somewhere, and over time the system gets better at not repeating them. The files themselves are on their own page, and Write your first skill shows how small one can be.
Eight things that broke, with the numbers
- The model copied its own rules into the answer (257 words, then 130). The brief builder’s instructions gave the rules after describing the brief’s last paragraph, and the smallest model wrote the rules into the brief. With the rules first, it stopped.
- Replies came back empty or cut off (4 bad replies in about 20). The chat’s allowance per reply was too small for a longer job. With a larger allowance for that one job, nine of nine came back whole.
- The check scrolled the fault out of sight. A check asked “is it inside the card?” and passed, while on a phone the content sat below the card’s edge. The test’s own click had scrolled it into view. The check now asserts the card has not scrolled.
- A leftover we counted and did not read. After a rename, one line still used the old name, which would have fed the plugin’s own blocks to the chat. A search had printed that line and we passed over it.
- A link landed 6px behind the header. A glossary term’s card was still rising into place when its position was read. Cards that are linked to no longer animate.
- A tool shared the wrong folder. In the server walkthrough, the server reported the folder we had started in, not the one we named, and Claude said there were no notes. We kept the mistake on that page.
- A source refused to be read by a script (12 of 26 dates). For the timeline, one maker’s site would not answer an automated request, so twelve dates were read in a browser by hand. Two launches whose pages showed no date were left out.
- A list went stale before it was published (5 moved, 3 dropped). Checking the directories, five projects had moved to new addresses and three were left out as archived or abandoned.
What we do not know
These are open as of 21 November 2026.
- Most pages’ chats have never been asked a question by us. The chat was tested on a few pages, not on each.
- The header’s glass effect has not been looked at in Safari.
- What we say about the tools we use less rests on what they are known for, not on runs of our own.
- The two walkthroughs were each run on one day, with one version of Claude Code.
To build something small the same way, start with the build job, or read Claude Code or Codex?
Ask Another AI Guy
I'm just Another AI Guy.
You're chatting with our fast model. Create a free account for the more thorough one, and more messages.