Site Builder · roadmap · 12 September 2026
Everything that happens in between, one station at a time: what each step does, whether a script or the AI does it, what it keeps, what it removes, and what you will see on the screen while it runs.
Built and tested on 13 September 2026. The panel is live; the first real batch ran end to end on it.
/root/.config/site-builder/admin.json.claude → /login in WSL, open /root/.config/site-builder/beast-install.txt on the build server and paste its whole contents into the Beast's WSL terminal. It installs the builder and starts it (as a user service, or a restart loop where WSL has no systemd). The block holds the tokens, so it is kept only in that file.Three domains, 18.6 minutes from Build to the delivery link: riversidehutches.com and candleboy.com shipped; elrondeagles.com stopped at the photo station (6 usable photographs, a site needs 7, one photo source was refusing requests). The delivery page and all three downloads were checked from outside and match the files byte for byte.
| Station | 1 | 2 | 3 | 4 · AI | 5 | 6 · AI | 7 · AI | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| Minutes per site (median) | 0.01 | 0.77 | 0.00 | 0.53 | 0.00 | 2.19 | 11.15 | 0.01 | 0.84 | 0.39 |
| AI tokens, whole batch | 0 | 0 | 0 | 60,533 | 0 | 62,139 | 377,234 | 0 | 0 | 0 |
499,906 AI tokens for the batch in all (input, cache and output). Writing the words is most of the time and most of the tokens. What is still not fixed is listed at the end of the build status file.
Three things, then nothing until the link arrives.
A private web page on its own Cloudflare account. Works on your phone.
One per line. Ten, twenty, thirty: they all build side by side.
Then watch each site move through the stations, or close the page and come back.
The same delivery page as 11 September: screenshot, live preview, ZIP and md5 for every site, plus Download all.
Four places. Your ad traffic never touches any of them, and your video renders keep running on the Beast's Windows side, untouched.
You paste domains and watch progress.
Stores each job and every station's log, permanently. Its own empty account (#1), so it can never slow Gitku, the bridge or the dashboard.
Picks the job up within a minute and runs the stations.
ZIPs, previews, every download re-checked. The link goes back to the panel.
memory=32gb in .wslconfig (today it is 64gb, the whole machine).Every site goes through these in order. Script = ordinary code, costs no tokens. AI = a Claude call, costs tokens. Only three stations use AI, and only one of them is big.
Looks up who the domain points to today: is it registered, does it run a live website, does it carry a working company mailbox.
You see: "live company mail on Microsoft 365: confirm you own this", printed on the site's card.
Pulls the domain's history from archive.org: the old home page and the pages it linked to, from the most recent good snapshot.
You see: the snapshot date used and the pages found. No history? It says so and works from the domain name alone.
Strips everything that belonged to the old owner or would be an unbacked claim today. The full list is in Cleaning old sites.
You see: every removed item listed, e.g. "removed: phone 01234 …, 'since 1998', 3 testimonials". Kept forever.
Is the old business in our safe list? If yes, keep it. If it was medical, finance, gambling, a licensed trade or a Google local-services trade, switch to the nearest safe business the name still fits.
You see: "KEPT: tack shop" or "SWITCHED: florist → saddlery, because florists are a local-services trade".
Hands each site its own colours, typefaces, page layout, class names, animation and city, drawn from pools so no two sites match, in this batch or any earlier one.
You see: the palette swatch, the two typefaces and the layout name on the site's row.
A script downloads 40–60 open-licence candidates. The AI looks at one sheet of thumbnails and picks 8–12 that fit.
You see: the picked thumbnails, and each rejected one with its reason.
The AI writes every page's text: headlines, services, about, FAQ, legal pages. It works from the cleaned notes, in fresh words. It writes words only, no page code.
You see: word count per page, and the site's one-line voice.
Pours the words into the assigned layout and makes all 9 pages, plus favicon, social card, sitemap, cookie banner and enquiry form.
You see: desktop and phone screenshots of the home page, as soon as they exist.
The four gates from batch 0911, plus the cross-site look-alike checks. If something fails, one AI fix is allowed. If it still fails, the site ships with the problem written on its card. No second round.
You see: 4 ticks, or the exact failing line.
Zips each site, makes live previews and thumbnails, publishes the delivery page, then downloads every file back to confirm it arrived intact.
You see: the link.
The previous business existed and your client bought the domain, so we rebuild that business, cleaned. Only when its category is safe.
The safety line: the old site's text is never published. It is only notes for the writer. And the finished pages are checked for phones, addresses, prices, reviews and brand names after writing, so anything that slipped through the cleaner is still caught.
One row per site, one dot per station. Click any site to open its full log: what each station did, what it removed, what it kept. The log stays permanently, for every batch, and never gets deleted. Try it below. The entries are real findings from batch 0911, laid out the way the panel will show them.
Example layout. Every finding and reason here is real, from the 0911 build; a dash means batch 0911 had no such station. The batch header, times and progress are illustrative.
The expensive part of batch 0911 was AI agents writing page code, re-checking it and reviewing each other. The new line gives page building to scripts and keeps AI for words and judgement.
Honest caveat: the right-hand numbers are targets, not measurements. The first real batch measures them and the panel prints the real token count per site. There's also one real risk: when scripts build every page from a shared layout library, sites can start to look alike. So the library needs many distinct layouts, and the look-alike check stays on every batch.
In order. Each phase is usable the moment it's done. Estimates are my honest guess, and each phase reports its real time when finished.
Beast WSL on 32 threads (verified). The checkers, packager and publisher from batch 0911. The ad-site-batch rules written down.
Stations 1–3: archive.org pull, clean, and the removed-items list.
Estimate: ~3 h
Station 8: many genuinely different page layouts that turn words into 9 pages. The biggest phase, and the one that removes most of the AI.
Estimate: ~1 day
Stations 4, 6, 7: the three AI calls, each with a fixed shape so the output drops straight into the layouts.
Estimate: ~3 h
Picks up jobs, runs all 10 stations, writes each station's log, sends progress to the panel.
Estimate: ~4 h
Paste → Build → live rows → click for the log → link. History kept permanently. Cloudflare account #1.
Estimate: ~5 h
Your next domains. Measures real time and tokens against batch 0911, and prints both on the page.
Estimate: ~1 h
Answered by Bob on 13 September. Build starts 07:00 IST.
In PowerShell: wsl, then claude, then /login. Open the link it prints in your browser, sign in, paste the code back, then /exit.
The AI calls run through your logged-in Claude, not an API key.
One line in PowerShell, then wsl --shutdown. Renders are unaffected.
Its own logo. Bob is admin; the VA gets a separate login that can start and watch batches.