Casel journal · ideas, language & the web
Blog
Two kinds of writing here. Practical guides — how to make a website with AI, what one actually costs, how to build in your own language — and working notes on the unglamorous layers underneath: tokenisation, typesetting, evaluation.
Featured
14 field notesFrom thought to a finished site
Build a website in the language you think in.
Describe the work naturally. Casel turns it into a thoughtful, publish-ready website.
Start buildingLatest posts
Practical guides and working ideasHow to create a restaurant website with AI: brief, prompts and checklist
Turn your menu, atmosphere, location and reservation path into an AI-ready restaurant website brief—without letting the generator invent the facts.
How to make a website with AI, step by step
The five decisions that decide the result, the two things AI is genuinely better at than a template, and the parts it still cannot do for you.
What a small business website actually costs
The build is rarely the biggest line. Here is the rest of the bill — domain, hosting, photographs, copy, and the hours nobody quotes for.
How to make a website in your own language
The language you think in and the language your customers read are two separate decisions. Getting them the right way round is most of the work.
llms.txt, explained without the hype
A plain-text map of your site for answer engines. What it is, what it definitely is not, and why ours is generated rather than written.
When one page is the whole website
Five pages because that is what websites have is the most common way a small site gets worse. A test for whether you need the second one.
Notes on typesetting nine scripts without breaking them
Line breaking, conjuncts, ascenders and mirrored layouts. The unglamorous work that decides whether a page reads as professional.
Writing so an answer engine can quote you
AEO is not a new dark art. It is the old advice — be specific, be structured, be quotable — with a machine as the reader.
Your language may cost more tokens than mine
Tokenisation was fitted to Latin script. Everyone else pays for that in latency, context and money.
The stack is English-first, all the way down
Not just the training data. The tokenizer, the evals, the docs, the error messages and the folklore about what to type.