A CyanLabs Prototype from Scratch in 30 Minutes for Only $0.20

A note before you read: this is not a technical tutorial, and it won’t even teach you how to write a single line of code. It’s about something else — how someone who genuinely enjoys tinkering with AI, with several AI tools installed and yet never really put to work, decided to seriously build a website of his own and, for the first time, let AI actually roll up its sleeves: 30 minutes and RMB 1.36 ($0.20), and a working website prototype stood up.

---

It all started with a question

Let me be upfront about who I am: I use AI every day. Gemini, ChatGPT and the like are old friends. I’m curious about new tools, and I even have agent frameworks like Hermes installed on my machine — but honestly, they had never been put to any real use. Looking things up, asking questions, writing small snippets — chatting solved all of that. I simply never ran into a “serious task” that justified letting an agent loose.

Building a website was the first serious task I actually wanted to see through.

My mental model back then was simple: AI is for asking, not for doing. So the first thing I did was wonderfully plain — I asked Gemini: if I want a website of my own, where do I start?

---

Step one: Let Gemini help me pick a theme and a name

The hardest part of building a website is usually not the technology, but one question: what website do I actually want to make?

I had three or four completely different ideas floating around: writing about photography, recording ordinary bits of everyday life, occasionally digging into personal finance and trading. They were too far apart. I couldn’t pick one, let alone figure out how to squeeze them all into a single site.

So I dumped every fragment into Gemini. It didn’t rush to hand me a plan. Instead, it helped me sort things out and came back with a surprisingly smart structure: rather than “one big mixed bag of everything,” make it “a knowledge garden made of several clearly defined columns.”

It organized my scattered interests into four sections:

• AI & Local Compute
• Photography & Optics
• Quant & Alpha Trading
• Everyday Saving

The moment those four names appeared, I got it instantly — this is what my website should look like. Each stands on its own, yet they share a spirit of “hands-on tinkering, doing it yourself.” Riding that spirit, Gemini also helped me come up with a name I loved:

CyanLabs.

“Cyan” is the color — AI, calm, and techy. “Labs” is where you run experiments: a place built for tinkering. A cyan-toned, lab-flavored name for a site that records my experiments — settled, and I never gave it a Chinese name. Through this whole phase, I never wrote a line of business code; I just kept talking to Gemini. Consulting is where chatbots shine.

---

Step two: Figuring out what the “website-building environment” actually is

Theme and name settled, the next question was practical: with this ordinary Windows laptop, can I actually build a website without spending money or buying a server?

Back to Gemini I went. It taught me one crucial concept — building a website doesn’t have to mean buying a server. There’s something called a “static website”: fixed web page files you drop onto a free hosting platform, and people can visit them — no server costs at any point.

It also walked me through a lot of “basic environment knowledge.” I grabbed the few points most useful to a first-time site builder:

• Install some things on the computer that “run website-building tools” (like Node.js) — follow the steps and it’s done
• Once the files are ready, throw them onto a free platform called Cloudflare Pages, and the site goes live
• Everything can be previewed locally first — see how it looks, publish only when happy

Looking back, these are basics — but back then it felt like a window suddenly opening: I didn’t need to master all the jargon upfront. I just needed to follow the pattern and take the first step.

---

Step three: Meeting a new species called “Agent”

Environment concepts sorted, but for a “chat AI veteran” like me, the real gap in understanding was this: I had always assumed AI’s job was answering questions — until Gemini introduced me to the Agent: AI that can actually do things.

It explained the biggest difference this way: with a chatbot, you ask and it answers; with an Agent, you tell it to execute. You can say “build me a website in this folder,” and it will genuinely go create files, write code, and run commands on its own — like hiring an assistant who gets things done, not just a consultant who talks.

That concept hit me hard: I had Hermes installed yet kept feeling it had “no real use” — precisely because it had never occurred to me to let AI execute a complete task. Building a website was big enough and concrete enough — it finally gave my agents a real job.

So my focus shifted from “what can AI answer?” to “how do I give an agent clear instructions and be a good client?”

---

Step four: Why I ended up with a “dual Agent” setup

One AI assistant should be enough, right? I thought so too, at first. But once I actually got started, I hit a very real pain point:

Chat AI is superb at brainstorming and reasoning — a phenomenal consultant. But ask it to actually sit down in the file system, dutifully create dozens of files, and run commands over and over until they succeed… it tends to feel “not quite wired for that.”

A dedicated agent that can genuinely operate your computer and working directory, on the other hand, excels at head-down execution — but it can get so focused on details that it loses the big picture.

Gemini made a clever suggestion: why not try a “dual Agent” division of labor?

• One agent acts as the “commander-in-chief” — overall strategy, task assignment, running the pipeline, watching quality. The project manager.
• The other acts as the “senior welder” — brought in whenever the interface code demands fine, detail-obsessed craftsmanship. The senior engineer.

That made total sense to me — one person acting as both project manager and welder easily drops the ball on one or the other; split them up, and each stays focused. So the combo became: Hermes leads as commander-in-chief, orchestrating the whole build; when a complex UI component shows up, Codex — the “welder” — is brought in to handle it.

That “commander + welder” pattern later became my go-to way of working — the full methodology lives in a separate write-up; this story is just about the first time I tasted the payoff.

---

Step five: Giving the agents a “brain”

Tools chosen, the agents still needed a “brain.” Gemini helped me sort out a clean mental model: the Agent is the hands and the skeleton; the Model is the brain. The agent operates files, runs commands, and drives the process; the model does the actual understanding and reasoning behind it. Pair the right model with the agent, and the agent becomes capable.

I equipped my dual-agent setup with DeepSeek models, and made a deliberate split:

• Hermes (commander-in-chief) mostly runs on DeepSeek V4 Flash with thinking mode off — for one reason: speed. Orchestration and routine operations don’t need deep reasoning; faster and cheaper wins.
• Only for those complex component codes did I bring in Codex, running DeepSeek V4 Pro to think carefully and craft the details.

A “fast grunt work + slow fine work” combo. At first I was just following the recipe; it took me a while to appreciate the logic — hand different difficulty levels to different “intelligences,” and you get both speed and savings.

---

The climax: in the final half hour, a website prototype actually appeared

And then came the part that amazed even me.

After I had laid out theme, name, environment, agent, dual-agent split, and model configuration, the actual “building” took only about the final half hour.

In that half hour, the Hermes agent was off on its own, grinding away — creating files, tweaking configs, running commands, fixing errors. When it hit a genuinely complex UI component, I handed it to Codex for fine work; when Codex came back, Hermes stitched the whole site together.

I basically watched it work, answering a question here and there — “should this button be emphasized?” “what copy goes in this section?” I never wrote a line of business code myself.

When it showed me the homepage in the preview, I stared at the screen for a long while: the cyan logo, the four column cards, an article entry point — all there. A real, visitable prototype of a website, ready to be filled with more content.

I kept confirming one thing in my mind: this is me — someone who usually just chats with AI and had agent tools installed without ever putting them to work — standing in front of a website an AI built from scratch.

---

The ledger: 8M tokens, RMB 1.36 ($0.20)

You might be wondering what all this cost.

I kept careful records, because the number was so small it surprised me. From both agents doing their work, running commands, and handling components, the entire build consumed roughly 8 million (8M) tokens — a considerable amount for AI, since the whole site had to be built from scratch with plenty of trial and error.

And what does that translate to in money?

RMB 1.36. At that day’s exchange rate, it came to $0.20. Yes: two dimes.

Not RMB 136, not RMB 13.6 — RMB 1.36. Cheaper than my breakfast. That number genuinely made me question reality for a moment: I’d always assumed building a website costs real money, and yet a live-ready prototype costs less than a single breakfast.

Of course, I’m aware: partly it’s because DeepSeek’s models are priced very low to begin with, and partly because the “fast model for grunt work, slow model for fine work” split cut out unnecessary expensive reasoning. Either way, from that day on, “even one person can own a website at near-zero cost” became real for me.

---

Ending: from a single question mark to a website of one’s own

Looking back, all I did was something very small and ordinary: opened a chat window, asked a serious question, and followed through on letting AI finish the job.

I didn’t become a programmer — and I didn’t need to. But I now own my very own website, CyanLabs, and I know: the website is just the beginning — every piece I want to write from here on has a home.

If you’ve always wanted a website of your own but kept thinking “that’s a programmer’s thing” — then I have one piece of advice from someone who’s been there:

Don’t let the word “technology” scare you. What actually matters isn’t the ability to write code — it’s the itch to build something, and the patience to spend an afternoon asking the right questions, one at a time. Leave the rest to agents that keep getting more capable.

The website is built. As for the stories inside it — like that trade, or that time the air conditioner got fixed — well, that’s what comes next.

Website diary series: A video and an old blog post · The 30-minute prototype · The dual-agent method · Launch and indexing

---

Told by an AI enthusiast, written with the help of Hermes Agent (DeepSeek V4 Flash), and recorded on the day CyanLabs was built.