Building a bilingual site with Claude Code, from first ask to live in 22 minutes

The site you are reading was built with Claude Code, an AI agent that works in the terminal: it reads files, writes code and runs commands on its own. This note covers how long it really took, what made it fast, and where AI could not stand in for a person. Every number comes from the git log and the actual work history.

The job

• A site for work and IT articles, static, built with Astro 5
• Two languages: Thai first, English under /en/
• Hosted on an Ubuntu server that already runs other sites behind the same nginx
• The first instruction was one casual sentence, roughly “make another site for IT, programming and security knowledge, as articles”

The real timeline

Time (8 Oct)	What happened
13:40	First sentence of the request
13:49	Thai site live at 6amdev.com with SSL (first commit, 26 files)
13:54	Asked for a second language
14:02	Thai and English live, with share images and security headers

22 minutes from the first sentence to a bilingual site anyone could open.

Why so fast

Twenty-two minutes sounds exciting, and it is misleading without the rest.

1. The article system already existed. Earlier that morning, about 45 minutes went into an article system for another site (fxboat.com): categories, levels, learning paths, a table of contents and FAQs. This site reused it as-is.
2. The server was ready. nginx and certbot were already installed; the job was adding one more site.
3. The rules live in a file. Each project has a CLAUDE.md that says how to deploy, what not to touch and what to check before calling something done. The AI reads it every time it starts.

The speed comes from what was prepared. AI assembles it quickly.

Counting the article system, the honest figure is a little over an hour, which is still far faster than doing it by hand.

Where a person was still needed

What the site says. The next day the home page was reworked to put working with AI up front. The AI wrote headlines that went too far, like an announcement, and had to be pulled back to something more neutral. What a site should say, and how loudly, is not the AI’s call.

Personal details. The AI did not guess the contact details. It left [placeholders] and kept the site out of Google until real ones arrive. That is the right behaviour, and it also means the site is not fully done until the owner fills them in.

Checking on a real screen. Some problems only show on screen, not in the code:

• Thai loanwords in big headings broken mid-word, like “คลาวด์” (cloud)
• An English headline wider than a 375px phone screen
• The Thai repeat mark (ๆ) falling to the start of a new line

The AI caught these itself by opening the preview at phone width, but it looked because the rules file says “before calling it done, check at 375px in both languages”. Without that line, there is no telling whether it would have looked.

What tripped it up

• Started in the wrong folder. The session began in another project, so the preview launcher could not find this site’s settings until one was added.
• A long script broke halfway. A multi-file edit sent through the shell in one go failed on a quote character. The AI switched to writing a script file first and running that, which is more reliable.

Both cost a few minutes. They show the AI does not get everything right first time; it just recovers quickly.

The numbers

Request to bilingual site live	22 minutes
Files in the repo	50
Code and content (lock file excluded)	about 4,000 lines
Pages built	19 (before the lab notes section)

Verdict: keep using it

Yes, especially for work with a clear pattern: static sites, repeated deploys, and anything that builds on something that already exists.

Getting results like this takes two things: prepared pieces for the AI to assemble, and rules written in a file. The decisions, such as what the site should say, which details can be public, and whether the work is ready for production, stay with a person.