How I Build Websites Using AI

I am not going to walk you through how to set up Hermes Agent with an LLM. Most of what I describe here works in a normal chatbot. What I am going to do is walk you through how I build a website using AI, from start to finish.

My Setup and What It Costs

Here is what I use:

• Hermes Agent — free, open-source AI agent that runs on your computer
• DeepSeek V4 Pro — approximately USD 20/month, the cheapest LLM that can handle coding tasks without a VPN
• GLM-5.2 via OpenRouter — approximately USD 50/month, more reliable for coding tasks than DeepSeek

Beyond the AI tools, there are other costs involved in running a website:

• Domain name — around USD 10-15/year (I use Porkbun)
• VPS (Virtual Private Server) — around USD 5-10/month (a rented computer on the internet to host your website)

Understanding What Makes Up a Website

Normally a website is made up of a frontend (what is shown to visitors), a backend (the logic that makes the website operational), a database (to store any information), and everything else that makes a website work (version control, analytics, security, deploying, DNS, hosting, backup, logging).

Asking AI to build a website and it will likely give you the frontend code (HTML, CSS, and JavaScript). It loads fine and works but it is living on your computer and no one can see it. If you have a contact form, it is not working yet because it needs the backend logic to run.

If your website has payment options or membership features, you will likely need a database to keep track of users and transactional information.

No matter what your website is about, you will need to deploy it to a VPS (a rented computer on the internet). It is like a digital storefront for your business. The same concept as renting a physical place to do business — you would not run your business from home. Hosting, DNS, and a bit of networking go in here. You can ask your AI agent to run these for you, but personally, I do not trust it enough to run these. There are nuances to different service providers and not every AI model is trained to handle these properly.

Come Up With a Rough Plan

Now you know what goes into a website, we can start writing a plan of what we want to achieve, what the technical specifications are, what pages there are, and what I do not want my agent to do.

I will leave my comments, that I do not include in the plan. These comments explain jargon and why I do certain things.

Here is how I build my portfolio website:

Goal: Build a one page portfolio website for me as a freelance web developer serving SMB, SE clients in Hong Kong.

Tasks:
- Break down the project into small, manageable tasks
- QA test at the end of each task

Technical specifications: # These are what I use to build a website
- Django for backend
- Tailwind and HTMX for frontend. Minimal Javascript usage
- Contact form to send out email.
- Rate limit the sending of the contact form
- Sanitize user inputs in the contact form
- Create a log for everything
- Use Plausible for analytics
- Ping me on Telegram if it encounters an error # I usually forget about it after the website is deployed and only by pinging me on Telegram will I notice any errors

Do not do the followings: # It's as essential to tell your AI what not to do as it is to tell it what to do
- Do not touch git
- Avoid using technical jargons
- User authentication and paywall are not needed

Follow the layout: # This is a layout that maximises building trust and conversion
- Hero with headline, subheadline and two action buttons
- Portfolio section showcasing the 33 industries I have in my portfolio
- Service. Display as cards. Use layman terms.
- FAQs
- Contact form
- Footer

Design
- Load the frontend-design skill and use the brutalist/industrial design # AI agent has skills, which are pre-defined instructions to carry out different tasks. It's perfect for me as I am more of a backend guy who has no taste when it comes to design

Deploy
- I will handle version control (github), deploying and managing domains on my own

I would then ask my AI agent for feedback.

Separate the Planning Phase From the Developing Phase

Start a new session when the AI agent understands what you want and you are happy with the plan.

Starting a new session allows your AI agent to not be confused with what has been said in the conversation. From my own experience, this helps carry out the development process efficiently, and therefore, saves on costs.

Picking an LLM Model and the AI Agent

Claude and ChatGPT are known to be expensive and inaccessible without a VPN in Hong Kong. If there are any changes from Claude or ChatGPT, you might lose your VPN connection and not be able to run the agent. Also, they are extremely expensive. If left unattended, they can run up the costs of development.

I opt for the cheapest and easily accessible options. In my opinion, DeepSeek V4 Pro does a great job and it is one of the cheapest LLM models available without a VPN. It is not perfect but it can handle simple coding tasks.

An alternative LLM model for coding is GLM-5.2, which is slightly more expensive but it can handle coding tasks much more reliably than DeepSeek.

Both LLMs are easily accessible via OpenRouter.

Getting Your Website Online

Right, now that the build is finished. You have checked the website, happy with it. It is time to deploy.

Here is a checklist to go through:

1. Get a domain name from anywhere. I use Porkbun. Easy to use. A domain name is yourcompany.com. Cost: around USD 10-15/year.
2. Find a VPS to host your website. VPS stands for virtual private server. It is basically a computer to host your website. Cost: around USD 5-10/month.
3. Save your work — commit your files to GitHub. This way, if mistakes are made and you want to go back to an earlier version of your website, you have the code on GitHub. It is like saving your game before the boss fight.
4. Upload your website to the VPS.
5. Configure the DNS (A/AAAA, CNAME, NS, TXT records). This is so that when someone types "yourcompany.com", the computer converts it to an IP address of where your VPS is.
6. Get an SSL certificate. This is the "https" in front of your domain. It means the transmission is now encrypted. This is vital if you accept user credit cards or any other sensitive information but it is a standard practice in 2026.
7. Get your website indexed on Google Search Console, so that when people look up your business, or certain words, it will show up.

Conclusion

These are the exact steps I take to build websites using AI. The key is to separate the planning phase from the developing phase as different sessions. It is how I managed to build 132 demo websites overnight.