- Claude Code — what it means
- Claude Code is a development tool that lets you generate, edit, and run code while conversing with Anthropic’s AI, Claude, directly in your terminal.
Hi, I’m Masato. I work as a Webflow specialist while chasing the leading edge of AI.
In July 2026 I fully rebuilt Webharu’s own website with Claude Code. Twenty-six new pages, 62 migrated articles, and four form types connected — finished in 3 days of building (7 days including planning). In this article I reorganize that build into a set of steps, written for technically inclined readers who want to try it themselves.
✓ In this article
- The five steps for building a corporate site with Claude Code (plan → design → build → verify → launch)
- The real timeline and the tools used to build 26 pages in 3 days
- Real examples of doing post-launch work with AI too (bulk AIO implementation, building a calculator tool)
Here is the conclusion first. You can build a 26-page corporate site with Claude Code alone. In my case, though, it took about a month — and a mountain of throwaway output — before I could reliably generate good design. What decides quality is less the procedure itself than where the human makes the judgment calls.
Prerequisites: the tools and the cost
- Claude Code (Opus and Fable): about ¥30,000/month in AI subscription costs
- Astro: a static site generator. I chose it for Cloudflare compatibility and page speed
- Cloudflare: hosting. Server cost is zero, and running cost is ¥1,500 a year for the domain and nothing else
The only human involved was me — though I already had a base of design, build, and SEO knowledge going in. The old site ran on Webflow, with subscription costs of just under ¥100,000 a year. After the switch, only the domain fee remains.
Step 1: Plan (Day 1)
The first thing you do is not code. It’s an inventory. Where does traffic come from, who is the target, what kind of customers have we actually had? I laid out the numbers I had and put into words the goal the rebuild had to achieve.
In my case I originally intended to angle the site toward large-enterprise work. But roughly 90% of our actual customers were startups and small businesses, so I reset the target to match. That is a judgment AI cannot make for you. If you enter the build with this still vague, the text review in the later steps turns into a complete redo.
Step 2: Design (Days 2–4)
Data migration and defining the design
I scraped 62 articles and 24 case studies out of the old Webflow site and pulled them down locally. In parallel, I defined the foundations of the design in text: colors, button placement, the direction I wanted to express. That definition became the base of every prompt from then on.
Fifty rounds on the home page
The new design was produced in plain HTML first. I put the weight on the home page and repeated a loop — generate three variations, say which parts I disliked and why, get three more — for roughly 50 variations in total. This is where the back-and-forth with AI is heaviest. The trick is to give reasons rather than impressions: say “something feels off” and you get the same thing back.
The sitemap comes later, on purpose
I built the sitemap only after the home page was settled. Conventionally the sitemap comes first, but I judged that locking it early puts strange constraints on the AI. Once the design is fixed on the home page, you can point the AI at it and mass-produce the lower-level pages in one go.
Step 3: Build (Day 5)
Next, port the finalized HTML design into Astro. I picked Astro over Next.js because, in my own testing, Cloudflare compatibility plus load speed made it the best-performing combination as an SEO medium.
Along with that came SEO and AI optimization across every page. I had Claude Code handle structured data, the sitemap, and llms.txt as well. For the articles, I first built one detail page design in HTML, then used it as a template to pour all 62 articles through.
Step 4: Verify (Days 6–7)
I connected Google Tag, Cloudflare, and the four form types, bringing everything up to just short of production. The most important task here is reviewing the text. The AI-generated pages contained a great deal of wording and numbers I had never intended. Check them against the plan and fix them in one pass.
Finally I reviewed every page with my own eyes. Errors in company information go straight to trust, so I don’t hand that to AI. The AI does the hands-on work; the human makes the calls. I believe that division is what decides quality.
Step 5: Launch and results
Launch means deploying to Cloudflare and handling the migration from the old URLs. The result: the day after the rebuild, 24 of the 30 old articles still held their index status (measured in Search Console). Do the migration design properly and a rebuild does not blow up your SEO.
The detailed day-by-day record, and what the human was thinking at each point, is in our article on the full rebuild process.
Post-launch operation works with Claude Code too: two things we did
Bulk AIO implementation across 63 articles
In July 2026, after launch, I bulk-implemented key-points boxes, FAQs (with FAQPage structured data), one-sentence definitions, and Speakable markup across 63 blog articles. Our internal score went from an average of 47.8 to 97.3. While I was at it, I fixed over 450 typos and ran three 301 consolidations on articles that were cannibalizing each other’s keywords. That scale is not realistically doable by hand, but it worked because the articles were structured as JSON and generated through templates.
Building a calculator tool
I also built our calculator for the transitional relief under Japan’s qualified invoice system (インボイス制度) with Claude Code — the tool page itself is in Japanese. Being able to complete a piece of development like “add one tool to the blog” without outsourcing is a major benefit of switching to an AI build. I share the ongoing measurements from this area first in our newsletter.
Being honest: what was slow and what didn’t work
- Fine adjustments are slower than in a no-code tool. “I want slightly more space between sections” tends to escalate into a global adjustment, and there were moments when brute-forcing it in Webflow would have been faster
- It took about a month before design came out reliably. Choosing the right words for prompts is hard, and several times I decided to throw away a half-built site and start again from zero
- If you launch something 100% left to AI, every site ends up with the same face. That lowers brand value, so aesthetic judgment should stay with the human
Even so, the wall of unfamiliar languages and technologies has essentially disappeared. We have productized this approach as AI Website Production (about one month including discovery, from ¥1,000,000). Try working on your own site with the steps in this article first, and if it looks difficult from there, get in touch — that’s perfectly fine.
FAQ
Can you build a corporate website with Claude Code alone?
The implementation, yes. The author built a 26-page company site in 3 days of building. Planning, aesthetic judgment on design, and fact-checking remain human work, though — abandon those and you end up with a site that looks like AI output you have seen somewhere before.
How much does it cost?
In the author’s case, about ¥30,000 a month in AI subscription costs. Hosting is on Cloudflare with no server fee, so running cost came down to ¥1,500 a year for the domain alone. In the Webflow era it was just under ¥100,000 a year.
Why choose Astro as the framework?
Because Cloudflare compatibility and load speed make it, in the author’s testing, the highest-performing option as an SEO medium. Next.js was a candidate but was not adopted.
Did SEO drop because of the rebuild?
The day after the rebuild, 24 of the 30 old articles still held their index status (measured in Search Console). The premise is that the design work, including migration from the old URLs, is done up front.
