Build a site of any complexity in one evening, no code and no builders – with voice messages in Telegram

Section 01What you get – a picture of the finished result

You're standing in line at a coffee shop. You open Telegram, tap the microphone, and say: "build a one-pager for my photography course, audience is moms 30 plus, goal is to book a free consultation." Three to five minutes later an HTML file lands in the chat.

You double-click it – and there's a live site on your screen. With all the sections, colors, buttons. A good one. Not perfect, but working – the kind you can already collect leads with.

Then come the edits, right there in the same chat. "Swap the photo on the second screen, make the button green, add a question about refunds to the FAQ." For every message the bot sends back a new version within a minute. Ten or fifteen iterations – and the site is built exactly the way you pictured it in your head.

At the end – one message: "publish it to the server, domain fotomom.com, I already bought it." Ten minutes later the site opens in the browser on that domain. With a green padlock. From any country.

That's it. You never opened a laptop. You never touched a terminal. You never typed a single command by hand.

Section 02What the system is built from

The system has four parts. Each one costs pennies or is free altogether. The main thing is understanding how they connect to each other.

Four building blocks

• Claude Code – the brain. It thinks, writes HTML, handles servers, manages files.
• VPS – a 24/7 computer in a data center. Claude and your bot live on it.
• Telegram bot – the interface. You give it commands by voice or text.
• Domain – the site's address on the internet. Without it the page only opens on your own computer.

How they connect

The logic is simple. You message the bot in Telegram – the bot passes your command to Claude on the server – Claude builds the HTML – the bot sends it back to you as a file. If you ask to publish – Claude puts the file in a folder on the server, which serves it on your domain out to the internet. That's the whole scheme.

Next we'll go up this chain from the bottom. First we assemble the building blocks themselves: sign up for Claude, rent the server, wire it up with the bot. Then we buy a domain and build the first site.

Section 03Signing up for Claude

If you're in a region Anthropic restricts, signing up directly with a local number and card won't work. But there's a path that works, and it takes about 30 minutes. You need three things: a VPN, a foreign number for SMS, and an account on claude.ai. If you're in the US, you can skip the VPN and number tricks and sign up directly.

Section 04Paying for Claude

For our task you need the Pro plan – $20 a month. Free Claude in the browser is just chat, with no agent features. And we specifically need Claude Code, which can crawl around the server and publish files. It only works on paid plans.

If your local cards don't go through on the Claude site, here are three working ways around it.

Section 05Renting a VPS

A VPS is a rented computer on the internet that runs 24/7. This is exactly where Claude and your bot will live. You pay $20–30 a month – and you get a server that never shuts down.

Why the server should be in the US or EU

Anthropic restricts connections from some regions. If Claude tries to reach their servers from a flagged IP – the account can get blocked instantly. So we take a VPS in a well-supported region: the US, Germany, the Netherlands, or Finland. Then all requests to Anthropic come from a clean address – legal and unflagged.

Where to get one

The best balance of price and quality for this task is a mainstream cloud provider with a simple dashboard and fast support. DigitalOcean, Hetzner, and Vultr are all excellent choices – clean control panels, servers in the US and the EU, and support that answers in minutes. Pick whichever has a data center closest to your audience.

Which server to order

Sign up with your provider by email. In the top menu look for "Droplets," "Cloud Servers," or "Deploy," then "Create server." You pick the specs to fit our task:

This is enough for both the bot and all the sites you'll be building – dozens of landing pages live comfortably on a single server at once. If you need more later – the plan scales up with one button, without losing data.

Once the server is created, the provider shows three values: the IP address, the login (usually root), and the password. Save them in Notion, your notes, or a password manager. We'll use them in five minutes.

Section 06Installing Claude Code on the server

This is the part where it gets magical for the first time. Instead of typing five commands by hand, you log into the server once through the browser console – and from there Claude sets itself up. You just watch.

Opening the web console

In your provider's dashboard, open your server. There's a button labeled "Console" or "Web terminal" – click it. A black window opens right in the browser. That's the server's terminal; you don't need to install any separate programs.

curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash - && sudo apt-get install -y nodejs

npm install -g @anthropic-ai/claude-code

claude

These three commands are the only place in the whole guide where you type anything into a terminal by hand. After that, everything happens through the Telegram bot.

Section 07Creating a Telegram bot

Telegram has a "bot for creating bots" – it's called BotFather. Through it you get your own personal bot with a unique token in three minutes.

Section 08Connecting the bot to Claude

Here's where the whole point of this begins. You have Claude on the server and a Telegram bot token. Now we connect them – and you switch to controlling the server by voice straight from the messenger.

One big prompt

Go back to the web console (where Claude is running). If Claude closed – relaunch it with the claude command. Then you write one big, detailed prompt:

I have a VPS. Access details:
– Server IP: 123.45.67.89
– User: root
– Location: Germany

I have a Telegram bot. Details:
– Bot token: 7234567890:AAH-jKlMnOpQrStUvWxYz1234567890abc
– My Telegram user_id: 265489854 (only this ID
  may use the bot, ignore everyone else)

The task – set up a full integration on the server:

1. Install the libraries python-telegram-bot, anthropic,
   requests, python-dotenv.

2. Store all secrets (bot token, my user_id)
   in /home/automation/secrets/bot.env with 600 permissions.

3. Create a script /home/automation/scripts/claude_bot.py
   with the following logic:
   – Connects to Telegram via the token.
   – Listens for incoming messages.
   – If a message is NOT from my user_id – ignore it.
   – If it's a text message – pass the text
     directly to Claude as a command to execute.
   – Don't handle voice messages at this stage yet,
     we'll add that in the next step.
   – Get the reply from Claude, send it to me
     in Telegram. Split long replies into parts
     of 4000 characters.
   – If the reply contains a file (HTML, image,
     document) – send it as a file, not as text.

4. Passing the command to Claude:
   – Via headless mode: claude --print "$COMMAND"
     --allowedTools "Read,Write,Edit,Bash,Glob,Grep"
   – Current working directory: /home/automation/
   – System prompt: "You are my personal assistant,
     you work with files in /home/automation/, you answer
     briefly and to the point, and before dangerous actions
     you ask for confirmation."

5. Logs (incoming, replies, errors) – to a file
   /home/automation/logs/bot.log.

6. Wrap the script in a systemd service claude-bot,
   autostart on boot, restart on crash
   with a 10-second delay.

7. After setup, send a test message
   to my user_id in Telegram: "Bot is up, ready
   to work."

At the end, show me all the code you created and the
commands you ran on the server.

Claude reads the task, confirms the plan, gets to work. It will ask clarifying questions – you answer "yes" or steer it. Within 10–15 minutes it sends a test message to your Telegram bot itself.

Checking the connection

Open your bot in Telegram – by the username you set in BotFather. Hit "Start," type "hello." If a reply comes back within a few seconds – the connection works.

From here on, all the control happens right in this chat. You can close the web console; you don't log into the server anymore.

Turning on voice messages

Right in the chat with the bot, send it this message:

Add voice-message recognition to yourself
using free Whisper right on the server.

When I send a voice message – you download the audio,
transcribe it to text locally (no OpenAI, no
paid services), then handle it
like a normal text command.

Pick a Whisper model that's good enough in quality
and doesn't lag.

After setup, let me know you're ready and ask me
to send a test voice message.

Within 5–10 minutes the bot installs Whisper itself, downloads the model, restarts. It writes "ready, send a voice message." You hold the microphone, say "show me your status" – the bot recognizes it, runs it, replies.

Section 09Buying a domain

A site without a domain only lives on your computer. No one on the internet sees it. So a domain is a required part. It costs pennies and takes five minutes to buy.

Where to buy

The simplest thing is to buy it from the same place you got your server, so everything stays in one dashboard. DigitalOcean, Hetzner, and most hosts have a "Domains" tab. If you prefer a dedicated registrar, Namecheap or Cloudflare are both great and cheap – connecting the domain to your server then takes two clicks.

Section 10Your first site by voice

The most fun part. You have a bot that listens to commands. You have a server with Claude. You have a domain. Next – we build the first working site. The cycle: brief → draft → iterations → publish.

Step 1. The brief through a dialogue

The main beginner mistake is to immediately ask "build a site for my course." You get a generic, faceless page. Before the brief, it helps to run the site through the exercise of narrowing your expert niche down to one sentence – then the headline and offer on the landing page appear on their own. The more detailed the brief – the higher the quality.

I do it more simply: I ask the bot to ask me questions. In the chat with the bot, write by voice or text:

I want to build a sales landing page.

First, ask me 12 questions to put together
a complete brief:
– What I promise the person (the main measurable result).
– The three strongest client pains.
– Five common objections.
– Case studies with numbers.
– Testimonials (names, niches, results).
– The site's tone (friendly, expert, provocative).
– What I like about competitors and what I don't.
– How I'm different.
– What's included in the product.
– Guarantees.
– Bonuses.
– The CTA button text and where it leads.

Ask one at a time, wait for my answer. After all
the answers, assemble the brief and ask me to approve it.
Only then start building.

The bot asks the questions one by one. You answer by voice. At the end it assembles the final brief – you say "approved" – it starts building the site.

Step 2. Getting the first version

Within 2–3 minutes an index.html lands in the chat as a file. You download it to your computer, double-click it – the browser shows a live page. All the colors, fonts, buttons. You can scroll, click, switch to mobile view (in Chrome – F12 → the phone icon).

This is a draft. Next we refine it through iterations.

Step 3. The edit cycle

The open HTML in your browser = the live version of the site. Right from there you give edits to the bot – by voice or text. Examples of commands that actually work:

• "Change the hero headline to 'Your goal in 30 days'."
• "In the case-study block add a fourth one: Andrew, designer, +$5K in 2 months."
• "Make the button red, it's washed out right now."
• "Drop the FAQ – rewrite it as 5 catchy questions instead of 10 dull ones."
• "Add my photo to the hero," and you attach a photo in the chat.
• "Make the testimonials block dark background with white text."

For each message the bot sends a new index.html within 1–2 minutes. You overwrite the old file on your computer, refresh the page in the browser – you see the result. The cycle repeats until "I like it."

Ten or fifteen such iterations – and the site on screen is exactly what you wanted. Including the little things that usually get lost with contractors: the exact phrase in the hero, the right button color, the specific testimonial in the cases.

Step 4. Publishing to the server

When you like the site – one message to the bot:

I like the site. Publish it to the server.

Domain: fotomom.com – I already bought it.

Do this:
1. Connect the domain to my server (set up the DNS
   in my registrar's dashboard – tell me which
   fields to fill in there).
2. Upload the site to the server at
   /var/www/fotomom.com/index.html
3. Configure nginx so the domain opens this file.
4. Connect free SSL via Let's Encrypt –
   so there's a green padlock.
5. Turn on http → https redirect and compression.

When it's all done – send me the link and a screenshot.
I'll open it and check.

Within 10–15 minutes the bot writes: "done, https://fotomom.com opens, here's a screenshot." You open it in the browser – the site is on the internet.

Section 11Six advanced tricks

The basic cycle already gets you 80% of the result. The remaining 20% – the tricks that turn an "ordinary landing page" into a tool with double the conversion. All of them connect with a single message to the bot.

Trick 1 – five references instead of an abstract "make it nice"

Find 5 sites you like visually (Awwwards, Dribbble, landing pages of big brands). The same "give a reference – get HTML on-brand" approach works great in a sales webinar built with Claude too – one line about the palette and tone assembles a 30–60 slide deck. Send their links or screenshots to the bot with the task of analyzing each one – colors, fonts, spacing, rhythm – and assembling a single "design passport" for your site. From then on Claude uses that passport in all the edits. The design quality jumps several times over. The site stops looking like a "generic template."

Trick 2 – automatic checking on three screens

Ask the bot: "take screenshots of my site in three sizes – desktop, tablet, phone. Check whether any mobile text has overflowed the edge, whether the layout has shifted, whether the buttons are clickable. Mark the problem spots with a red outline." Within a minute you get three screenshots with markup – you immediately see what to fix.

Trick 3 – analytics in one prompt

Sign up for Google Analytics and copy your tracking ID. Which goals and metrics to track on the landing page so the funnel doesn't run blind – CAC, CR, LTV, ROMI – is covered in the expert sales funnel guide. To the bot: "connect Google Analytics with ID G-XXXXXXX, set up two goals – click on the lead button and a successful form submission. Publish the changes to the server." Within a minute the tracker is connected, and within a day you see visitor and lead stats in your analytics.

Trick 4 – dark mode with one command

"Add a dark theme to the site. A sun-moon toggle in the top right corner. If the person picks dark – remember it, and open in dark on their next visit. If their device is set to dark system-wide – default to dark. Publish to the server." In 30 seconds the site gets a second mode.

Trick 5 – an A/B test of the two main screens

"Take my current hero as version A. Make a version B with a different headline and button. Half the visitors see A, half see B, at random. Log every view to analytics. After 7 days remind me to look at the results." A week later the bot itself suggests: "version B brought in 30% more leads – switch?" You decide – it switches.

Trick 6 – automatic backups before every edit

"Before every publish to the server, first save a copy of the current version. Keep the last 10 copies. If I ever say 'roll back to yesterday's version' – you take the right copy and restore it." You ask once – after that Claude always has your back. Something broke – "bring back the last working one" – within 30 seconds the site is fine again.

Section 12Common mistakes

The main rule for any failures: you don't touch the terminal by hand. No systemctl, no journalctl, no editing files by hand. The bot fixes everything – you just describe what isn't working. The most common situations are below.

The bot doesn't respond

Open the Claude desktop app (or the server's web console) and write: "The Telegram bot isn't responding. Get into the server, check the claude-bot service, look at the last 100 lines of logs, check the Claude token and the Telegram token. If you found the cause – fix it, restart the service, send me a test in the chat." Within 2 minutes the bot works again.

Voice messages are recognized poorly

To the bot itself in Telegram: "your recognition is glitching. Force the language to English, switch to a more accurate Whisper model, and after transcription run the text through yourself with an instruction to fix recognition errors by meaning. Restart and confirm." The percentage of garbled recognition drops 5–10 times.

The site doesn't open on the domain

"Get in and check whether the domain is connected to the server, whether the SSL is valid, whether nginx is configured correctly. If something's wrong – fix it, check from your side that the site opens, and send me confirmation." Within 2 minutes the site works.

Building too long a site

30 screens before the button – the person won't scroll that far. The sweet spot is 8–12 meaningful sections. Better short and concentrated than long and watered down. If the bot built an overloaded site – ask: "cut the site down to 10 key sections, keep the strongest."

Not testing on mobile

70% of traffic comes from phones. Check every major edit through Trick 2 – ask the bot to take screenshots for three sizes. You'll save hours of rework.

Chasing perfection and never publishing

Better to ship a "good" site today and improve it as you go than to polish the "perfect" one in a drawer for three months. Faster test – faster feedback – faster results. Publish now, redo it in a week.