Hi, I'm Masato — a Webflow specialist who now tracks the AI frontier daily.
We renewed Webharu's website — the entire site, rebuilt by AI. Three days for the build, seven including planning. This article publishes the whole production process, logs and all.
What we did
- Built 26 static pages from scratch (including the cherry-blossom animation on the top page)
- Migrated 62 articles and 24 portfolio entries from the old (Webflow) site
- Connected 4 forms — contact, document requests, and more
- Set up GA4, Search Console, structured data, the sitemap, and llms.txt
The only human was me. The hands on the keyboard were AI's (Claude Code).
I used my experience to narrow the target audience and rebuilt the site around it.
It took about a week — most of which was planning.
The seven days, day by day
Day 1: Planning and goals
Strategy first: where does traffic come from, who is the target, what kinds of customers have we actually had — inventory the facts, then set the goals the renewal must achieve.
Day 2: Migration and design direction
Scraped and migrated the old site's articles and portfolio. Checked whether the published information matched the target audience, pulled all assets local, and defined the design's foundations — palette, button placement, the direction we wanted to express.
Day 3: Fifty design patterns
New designs, HTML-based — roughly 50 patterns. Weight on the top page: tell the AI exactly what I disliked about what I disliked, generate 3 more patterns, repeat.
Day 4: Sitemap and lower pages
Only after the top page was done did I write a clear sitemap — deliberately late, because I didn't want to strap odd constraints onto the AI early. Conventionally the sitemap comes first, so this was an irregular step. From that sitemap, the lower pages were generated in one sweep.
With the top page's design settled, lower pages just follow it — the design work shrinks dramatically.
Day 5: SEO optimization and the Astro build
SEO and AI optimization across all pages, migration of CMS articles and portfolio pages, one HTML design for the detail template — then the move into Astro. Why Astro over Next.js: Cloudflare compatibility and loading speed, which I've confirmed give the best performance for an SEO-driven site.
Day 6: Integrations and text review
Google tags, forms, Cloudflare — everything wired to the brink of going live. Then a full-text review of the entire site against the plan, and a re-definition of direction using numbers from GA4. AI-generated pages contained plenty of copy and figures I'd never intended, and they all got scrutinized in one pass.
Day 7: Full visual check → launch
Every page checked by eye. Real visitors are company representatives considering hiring us — publishing wrong company information is not an option. Checked everything, then published.
That's the clean version. Honestly? Before designs generated well, I spent about a month producing garbage with AI. Tuning prompts and phrasing was genuinely hard; there were hard calls to throw away half-built sites and restart from zero, and backend-requiring features that wouldn't generate well. Reaching a setup that even runs free of server costs took all of that time.
Still, AI improves daily. It is not free, but paying roughly ¥30,000 a month for AI gets you to the point where about a month of effort produces a clean site. I built this with Claude Code (Opus and Fable), and for website production specifically, I think AI has already reached the top tier: state clearly what you want and it can build nearly anything. What information to publish, and what design earns trust — that remains human territory.
What I (the human) was doing
To repeat: everything that involves thinking and deciding is the human's job. AI can only do what it's told.
- Deciding concept and structure (what the site communicates)
- Aesthetic judgment on design
- Fact-checking the copy, and business calls (pricing, positioning)
- Reviewing each stage and giving the GO
Hands: AI. Judgment: human. That division determines quality.
I initially planned to position for enterprise clients. But pageviews and inflow data said the real audience is startups and SMBs. We've operated since 2020 and taken on JTB and Nippon TV projects — but those proved low-repeatability, barely 10% of the whole. The other 90% were startups, founders just beginning, and companies wanting renewals.
What I learned
Three conclusions up front:
- Fine-grained tweaks are slower than brute-forcing them in no-code used to be
- The wall of unfamiliar languages and tech has nearly vanished (I've shipped App Store apps now)
- Publish 100% AI output unedited and every site converges on the same face — degrading your brand
Did it save time overall? Honestly, still unclear. You can't just click the thing you're looking at and fix it: widening the gap between two sections used to be click-and-drag; with AI, "widen just that" doesn't quite work and the design stops looking right — a small tweak escalates into a whole-layout adjustment.
On the other hand, I now feel fluent in code I've never learned. I've released multiple apps on the App Store — which requires Swift — because once you can handle planning, strategy, and design, the implementation language no longer matters.
FAQ
Q. Can AI build a website alone?
The implementation, yes. But planning, aesthetic judgment, and fact-checking are human work — abandon those and you get "that AI-looking site you've seen somewhere before."
Q. How long does it take?
This site took 7 days: 3 building, 4 planning. For client projects, allow about a month including discovery.
Q. What does it cost?
AI site production from ¥500,000 (¥350,000 for the first five monitor clients); the post-launch operations plan from ¥15,000/month.
Q. Which AI?
Claude Code (Opus and Fable). Usage details are in the AI vibe coding article.
This process is now a product
AI site production: humans handle planning, judgment, and quality control; AI handles coding and page generation. We offer AI site production (Japanese) built exactly the way this site was.
- AI site production: from ¥500,000
- Post-launch operations: from ¥15,000/month (tracking and applying the latest AI/SEO developments monthly)
Interested? Get in touch.
The full build log is in the complete AI production record (Japanese).
