AI · 3 min
Let an AI write your Kaury site
One prompt, one spec file, one compiler that explains its errors. A loop that turns Claude, ChatGPT or Cursor into a reliable web developer.
Kaury was designed so that a language model can write it well: one way to do each thing, the same shape on every line, a whole site in a few dozen lines.
The prompt
Read https://dev.kaury.studio/kaury-ai.md — the full specification of the Kaury language.
Then write a complete Kaury site (one site.kaury file) for: a bakery in Lisbon,
with a 3D croissant, the menu, opening hours and a contact form.
kaury-ai.md is about 300 lines. It holds the syntax, the web elements, styles, animations, 3D and a list of golden rules. Assistants that browse can read the address; for the others, paste the file.
The loop
- Save the answer as
site.kaury. - Run the checker:
kaury check --json
- If there are errors, give the JSON back to the assistant. Each error has the line, the column, the word and the fix:
{ "severity": "error", "file": "site.kaury", "line": 4, "column": 20, "message": "\"conut\" does not exist.", "fix": "did you mean \"count\"?" }
- Repeat until
kaury checkis green, thenkaury dev.
The checker also renders every page, so what only fails when a page is displayed is in the list too. And when the assistant writes HTML, CSS or JavaScript by habit, the fix is the Kaury form: font-size 20 answers in Kaury: size 20, && answers "&&" is written "and".
An agent such as Claude Code or Cursor can run this loop on its own.
Why it works better than JavaScript
- Less to write, less to get wrong: a full site is one answer, not twenty files.
- No version soup: there is no React 17 vs 19, no router to choose, no CSS framework to guess.
- A strict compiler: mistakes are caught before the browser, with a fix the model can apply.