A personal website traditionally means a few weeks of briefing a designer, waiting on a developer, and compromises. I tried it differently — through AI. This site was built in a week without me writing a single line of code.
This isn't a guide on how to fire your vendors. It's a snapshot of what AI can actually do today, where its limits are, and how to approach it whether you're building a project alone or with a team.
In the article you'll find a case study and four lessons that apply any time you put AI to work on a project of your own.
Why I went for it
The market changes faster than companies can adopt. Development agencies and AI products are everywhere, but the processes inside companies remain stuck in the "vendor — client" era: long RFP cycles, missing context, waiting on iterations. Meanwhile, the technology has moved so far that the same outcome can be reached an order of magnitude faster.
Market need is shifting toward in-house solutions. Typical example: a company wants an internal ERP+CRM connected with Czech tools (accounting, e-commerce, warehousing). Process consulting runs into the hundreds of thousands of CZK for the design alone, a development studio another hundreds of thousands to millions for implementation. The reality: it's enough to have someone on the team who knows the company's data and processes. A weekend with Claude.ai plus a few days of work — and that person has most of the solution done. External help is left to polish, not build from scratch. If this is exactly what you're solving, it's a scenario for Solution Architecture.
My own positioning across services. I have the Utima software house, the Kanbu.ai SaaS, and I do advisory. Clients want to hear an objective recommendation, not a pitch for one of my own services. This site helps me communicate an independent position — that's why jakubkubista.com exists separately from Utima and Kanbu, in the spirit of AI Strategy Advisory without vendor lock-in. Background on /about.
How it worked
I set the target audience, took the Utima brand manual as the visual baseline, and in Claude.ai I created a dedicated agent — a structured project with reference material (business goals, content strategy, technical brief, the way of communicating with Claude Code). Claude Code then built the site to spec without me touching the source code. My job was to iterate: refine the brief, review outputs, fine-tune details.
What this can do in practice today:
- Build a production website. Complex structure, animations, responsive design. Stack: Next.js, TypeScript, Tailwind, deployed on Vercel. Supporting infrastructure (Resend, Cloudflare, reCAPTCHA, GA, GSC) is just a few lines of configuration.
- Iterate in real time. Instead of an RFP cycle and waiting, I tested variants on the spot. Don't like something? Rephrase, regenerate, compare.
- Consistency. The Claude.ai agent kept context across hundreds of iterations. The documents (positioning, brand, content, technical spec) stayed synced.
Time and cost comparison: a classic build for a website of this complexity takes a month or two and hundreds of thousands of CZK. The AI process cost me a week of work, primarily the price of Claude Pro and Vercel. Bonus: the project is now ready for content publication automation — content via a dedicated agent, later an n8n flow.
What I ran into
1. I rethought the CMS. I originally planned Payload CMS. Then it hit me: why? Teaching someone another admin UI takes time, and content updates are more naturally delegated to an AI content-marketing agent. Instead of a CMS dashboard I have a pipeline where I tell the agent "write a blog post on topic X", it drafts, I review, I deploy. Lesson: Before introducing a traditional tool, ask whether you even need it in the AI era.
2. We came back to waterfall (partially). It sounds controversial, but it's true. AI authoring isn't agile in the classic sense — it needs a solid brief upfront. Without a clear architecture and content strategy, you generate content that doesn't hold together. Lesson: Invest time in the spec before you let AI loose. What used to be the "pre-project phase" is now the biggest insurance policy of success.
3. An LLM fills in the blanks without a brief. Without instructions it slides into generalities — "a modern approach", "a comprehensive solution", "cutting-edge technology". Nobody reads that. Specific numbers, names, context = readable copy. Lesson: Feed AI specific context — internal data, client names, real-world numbers. Generic in, generic out.
4. Without methodology you go in circles. The first iterations were chaos — I changed my mind, AI rewrote things, context got lost. Then I set clear rules (decision log, single source of truth, version control in markdown) and suddenly we were moving in a straight line. Lesson: Set the methodology of how you'll work with AI at the start — and adjust it as you learn.
What to take away
AI today makes it possible to build things that used to be outsourced or take months. But AI without a brief = generic output. The investment into preparation and methodology pays off many times over.
If you're considering a similar project at your company (an internal tool, a website, automation), the key question isn't "can we do this with AI?". The question is "is our context ready enough for AI to do this well?"
This site is a first iteration, not the final version. I'm gradually adding more content, publication automation, integrations.
If at your company you're working on a similar question — where AI makes sense, where it doesn't, and how to do it efficiently — I'd be glad to talk it through. Schedule a consultation →
/ Terms in this article /
