Hello, this is Masato — a Webflow specialist who also follows the leading edge of AI at all times.
We have renewed the Webharu site. The whole site was renewed with AI. Three days for the build alone, seven including the planning. Here I am publishing the entire 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 site (Webflow)
- Connected 4 forms, including the contact and brochure-request forms
- Set up GA4, Search Console, structured data, the sitemap, and llms.txt
One human: me.
The one doing the hands-on work was the AI (Claude Code). Drawing on my experience so far, I narrowed the target and renewed the site to suit it.
It took roughly a week to finish, and most of that was the planning.
We also generalized this process into a step-by-step guide: Building a website with Claude Code — 26 pages in 3 days
The seven days broken down: the AI renewal process
Day 1: Planning and setting the goal
Starting with the planning strategy. Where does the traffic come from, who is the target, what kinds of clients have there been so far? I took stock of the information and set the goal the renewal was to achieve.
Day 2: Migration and defining the design direction
I extracted the old site’s articles and portfolio entries by scraping and migrated them. I checked whether the information currently published matched the target, and pulled all the assets down locally. At the same time I defined the foundations of the design, setting out the colors, the placement of buttons, and the direction I wanted to express.
Day 3: Making 50 design variations
I created the new design in HTML. I probably made around 50 design variations. Weighting the work toward the top page, I told the AI what specifically I disliked about the places I disliked, and repeated the cycle of generating three variations at a time.
Day 4: The sitemap and the lower-level pages
Once the top page was finished, I drew up an explicit sitemap. The reason I did not make the sitemap first is that I did not want to put strange constraints on the AI. Generally speaking a sitemap is something you make first, so this was a slightly irregular step. Working from that sitemap, I built the lower-level pages in one go.
Once the design is settled on the top page, building the lower-level pages is just a matter of following it, so the design work can be shortened dramatically.
Day 5: SEO optimization and building in Astro
I optimized every page for SEO and for AI, migrated the CMS article pages and portfolio pages from the existing site, and first designed just one detail page in HTML. After that I moved on to building in Astro. The reason I chose Astro rather than Next.js is that its compatibility with Cloudflare is good, and I have confirmed that its loading and speed deliver the highest performance for an SEO property.
Day 6: Integrations and scrutinizing the text
I connected the Google tags, the forms, Cloudflare and the rest, and brought it to just short of going live. And then, from there, came scrutinizing the text of the whole site, checking it against the plan, and redefining the site’s direction using the numbers derived from GA4 and elsewhere. The pages the AI had built contained plenty of wording and figures I had not intended, so I went through them all together in one pass.
Day 7: Visual check of every page, then launch
I check every page by eye. The people actually visiting the site are staff at Japanese companies and heads of firms who want to commission a website, so there is no room for publishing a company’s information incorrectly. So I checked the whole thing through for mistakes. After the check, we launched.
That is the flow of the seven days.
Let me be honest, though. It took about a month before I could generate designs well, and along the way I produced a great deal of rubbish with AI.
Tuning the prompts and the way things are worded is extremely difficult. There were also decisions to throw away a site I had built halfway and start again from zero. When items requiring a backend came up, generation did not go well at all. That is how long it took to arrive at a way of running everything, server costs included, for free.
Even so, AI evolves by the day. This time I used Opus and Fable in Claude Code, and when it comes to AI building a site, I think it has already reached the summit. Give it clear instructions about what you want to make and it can make almost anything.
What information to publish, though, and what kind of design conveys trust — that is still very much a territory for humans to think about.
Building a website was, in truth, already something an amateur could do easily, back in the era of no-code tools. But what design to make, how to present it, and which target to deliver it to — there, I feel, work for the professional still remains.
On the other hand, while it is not free, I think that paying around ¥30,000 a month for AI and putting in about a month of effort will get you to the point where you can build a good-looking site. In my case, of course, I already had the design knowledge, the build knowledge, and the SEO knowledge, so I did not need that groundwork study.
Doing that groundwork study and then running your own site yourself for a long time to come will be an ideal approach for companies that want to keep costs down while carrying out their own publicity from here on. On the other hand, constantly chasing trends and continually adopting the newest technology is bound to be hard work for some people.
On that point, I chase trends constantly, almost as a hobby. And the sites we run at our company are sites where anything newly released is applied as soon as it appears.
As noted above, building a website takes about a month including the interview phase. The actual working time may be around seven days, but that assumes every premise is already settled.
What I (the human) was doing
To repeat: thinking and deciding what to carry out is entirely the human’s job. AI can only do what it is told.
- Deciding the concept and the structure (what the site is there to communicate)
- Aesthetic judgment on the design
- Fact-checking the copy, and the commercial judgments (pricing, how to present it)
- Reviewing each stage and giving the go-ahead
In other words, the AI does the hands-on work; the human makes the calls. That division is what decides the quality.
In a sense, I think a demanding company president should build the site themselves.
The beautiful image in your head and whether it is beautiful once actually given form are two completely different things. Cram in everything while it is still glittering in your head and you will end up with a site that looks pretty but is a mess. My role is to tidy that up.
The AI will tidy it up for you as well, but humans are inefficient creatures, and we reject the AI’s suggestions and make things messier and messier. I was like that at first too. That is why deciding the direction matters.
At first I intended to customize the site for large-company work.
But from the pageviews and the routes clients arrive by, I judged that the target really should be startups and small and mid-sized companies.
The business itself has been running since 2020, and we have taken on work for JTB and Nippon TV. Those are low in repeatability, though. They came to only about 10% of the whole, and the remaining 90% were startups and people about to start a business or wanting a renewal. I also could not bring myself to take out the insurance needed to pass a large company’s credit screening.
What I learned by doing it
Three conclusions first.
- Fine adjustments are slower than back when I muscled through with no-code
- The wall of languages and technologies I do not know has all but vanished (I have even shipped App Store apps)
- Publish with 100% left to the AI and every site ends up with the same face, which lowers brand value
Honestly, whether time was shortened is something I still do not really know.
You cannot fix the spot you are looking at directly. For example, when I wanted to widen the gap between the About and Blog sections a little, in the past I would just click that spot and widen the margin. With AI, telling it to widen only that does not go well, and the design stops looking clean. In other words, a small adjustment develops into an adjustment of the whole.
That does of course produce a clean design, and I think it is a good thing, but honestly I suspect various tasks finished faster back when I was forcing adjustments through by brute strength. Widening a piece of whitespace a little like that is nothing more than a small preference of mine, and the visitors who arrive will almost never look at it and think it odd.
Pointless work like that piles up, and it is hard to get rid of however much you try. That, I think, is the trap of building with AI.
Of course, publishing what the AI makes without correcting it is perfectly possible. Do that, though, and the AI-ness gets stronger and everyone’s site becomes much the same. In the end it lowers brand value and leads to the result of “would it not have been better to build it in no-code?”
On the other hand, I feel as though I can now handle any code, even code I do not know. For example, I currently have several apps released on the App Store. Registering an app on the App Store meant having to use a language called Swift, but as long as you can think through the plan, the strategy and the design, you can now build anything whatever code it uses.
Languages that are not supported, and 3D games in something like Unity, are still somewhat difficult, I think, but that too is only a matter of time.
Frequently asked questions
Q. Can a website be built with AI alone?
The implementation can. The planning, the aesthetic judgment on the design, and the fact-checking are the human’s job. Abandon those and you get an AI-looking site you have seen somewhere before.
Q. How long does it take?
This site took 3 days of building plus 4 days of planning, 7 in total. For client projects, allow about a month including the interview phase.
Q. What does it cost?
AI website production starts at ¥1,000,000 (excl. tax, 1–2 weeks at the fastest), and the post-launch Site Care Plan starts at ¥15,000/month.
Q. Which AI did you use?
Claude Code (Opus and Fable). How to use it is written up in the article on AI vibe coding.
We have made this process into a product
AI website production is a way of building websites in which humans handle the planning, the judgment and the quality checks, and AI handles the coding and the page generation. Alongside the publication of this article, we are opening applications for AI website production, built the same way as this site.
- AI website production: from ¥1,000,000 (excl. tax, 1–2 weeks at the fastest)
- Post-launch Site Care Plan: from ¥15,000/month (every month we chase the latest AI and SEO information and apply it)
A detailed record of the production process is in the full record of the AI build. If it interests you, please do get in touch.
FAQ
Can a website be built with AI alone?
The implementation can. The planning, the aesthetic judgment on the design, and the fact-checking are the human’s job. Abandon those and you get an AI-looking site you have seen somewhere before.
How long does AI website production take?
This site was finished in 3 days of building plus 4 days of planning, 7 in total. For client projects, about a month including the interview phase is the guide.
How much does AI website production cost?
AI website production starts at ¥1,000,000 (excl. tax, 1–2 weeks at the fastest), and the post-launch Site Care Plan starts at ¥15,000/month.
Which AI did you build it with?
Claude Code (Opus and Fable). The division of labor is that the AI does the hands-on work, while deciding the concept, aesthetic judgment, fact-checking and the go-ahead to publish are the human’s.
