AI for a Tilda website: copy, HTML blocks and page structure

A neural network for building a Tilda website is useful, but not the way headlines like to describe it. Tilda is assembled from blocks in a visual editor, so "generate the whole site" will not work: the network does not click buttons in your builder. But it helps with three things: it writes copy for blocks, helps plan the page structure, and writes small HTML, CSS and JS fragments for embedding. Below are working techniques and limits. Model prices are as of October 18, 2026.

What AI does together with Tilda

On the general principle "AI writes, a human checks" read how to write texts with a neural network↗.

Step 1. Page structure

Start not with the copy but with a plan. Describe the business, audience and page goal to the network and ask it to propose a section order. A sample request:

Landing page for a bicycle repair shop in Kazan. Audience: city cyclists.
Goal: booking a diagnostic. Propose a page structure of 7-8 sections: what is in
each section, which main reader question it answers and which action it asks for.
No invented prices or reviews.

Once you have the plan, map the sections to blocks from the Tilda library: cover, benefits, steps, gallery, FAQ, form. See block names and codes right in the editor: the network may invent them, and you pick the real ones.

Step 2. Block copy

Give the model firm bounds: how many characters in a heading, which facts are allowed, what tone. Length limits in Tilda blocks are set by the layout, so it is better to ask for variants of different lengths right away.

Write for the "Cover" block: heading up to 60 characters, subheading up to 120,
button text up to 20. Give 3 variants. Use only facts from the list:
- diagnostics on the day of contact;
- we work by appointment;
- address: Primernaya St., 1.
Invent nothing beyond the list.

The rule: you insert all numbers, terms, guarantees and addresses yourself. If the network "made up" a term or price, that is an error, not creativity. To avoid copying one block at a time, you can run a list of pages or blocks through a script.

Batch copy generation with a script

A Python example with the official openai package. The fetch_... key is created on the API keys page↗, the address has /v1, and the request format is described in the chat completions documentation↗.

from openai import OpenAI

client = OpenAI(base_url="https://api.zukko.pro/v1", api_key="fetch_your_key")

facts = "Diagnostics on the day of contact. We work by appointment."
blocks = ["Cover", "Benefits", "How we work", "Questions and answers"]

for name in blocks:
    resp = client.chat.completions.create(
        model="claude-sonnet-5-5",
        messages=[
            {"role": "system",
             "content": "You are a website editor. Use only data from the FACTS block. "
                        "Do not invent prices, terms or reviews."},
            {"role": "user",
             "content": f"FACTS: {facts}\nWrite copy for the '{name}' block "
                        "of a bicycle repair shop website. Give 2 variants."},
        ],
    )
    print(f"--- {name} ---")
    print(resp.choices[0].message.content)

Copy the result into Tilda blocks by hand: that way you see at once how the text fits the layout.

Cost calculation (the volumes are an assumption). Say a page takes 1,500 input tokens and 1,000 output tokens, 30 pages in total: 45K input and 30K output. Formula: input in millions × input price + output in millions × output price.

• claude-sonnet-5-5: 0.045 × 0.4 + 0.03 × 2 = 0.018 + 0.06 = $0.078.
• gpt-6.1-sol: 0.045 × 0.27 + 0.03 × 1.32 = 0.01215 + 0.0396 = $0.05175.
• deepseek-v4.1-flash: 0.045 × 0.013 + 0.03 × 0.038 = 0.000585 + 0.00114 = $0.001725.

This is a teaching calculation. Which model writes best for your niche is unknown in advance; compare on a single block. Prices change; the current ones come from GET https://api.zukko.pro/v1/models.

Step 3. HTML, CSS and JS for embeds

Per the Tilda help, the T123 "HTML code" block adds any code to the page, including script and style tags. The block is in the block library, in the "Other" category. HTML can also be added as an element inside Zero Block. Code can go inside the page body or into the head of the site (Site settings, the "More" section, "HTML code to insert inside head") or of a single page.

Important notes from the help:

• in edit and preview modes the code is shown as text and works only after the page is published;
• Tilda support does not consult on code you added yourself;
• so styles do not break the interface and other blocks, do not use bare tags, universal classes or * in selectors: use unique classes or IDs;
• your own PHP cannot be inserted into the page, but data can be sent to your server through a Webhook receiver.

These notes are worth putting into the request. Example:

Write a fragment for a T123 block in Tilda: a "New" badge above a heading.
Requirements: only HTML and CSS inside <style>; a unique class with the zk- prefix;
do not touch tags or common classes; no external libraries.

A typical result:

<style>
  .zk-badge {
    display: inline-block;
    padding: 6px 14px;
    border-radius: 20px;
    background: #e8f0ff;
    color: #1a3d8f;
    font: 600 14px/1.2 sans-serif;
  }
</style>
<div class="zk-badge">New</div>

And a small script: the current year in the footer.

<span class="zk-year"></span>
<script>
  document.querySelector('.zk-year').textContent = new Date().getFullYear();
</script>

Check after publishing, on a phone and on a computer. The network cannot see your page, so it does not know which styles Tilda already sets.

What not to do

• Do not put an API key in page code. Everything in T123 is open to any visitor. If you need a neural network on the site (for example an assistant widget), requests must go through your server, and the key stays there. On a chatbot: a chatbot for a site on a neural network↗.
• Do not promise clients what the model wrote. Check terms, prices, guarantees and legal wording.
• Do not rely on SEO texts "as is". Proofread long search-oriented texts and add your own facts: the model easily produces repetition and generic phrasing.

Vibe coding on Tilda: an honest boundary

The word "vibe coding" implies that you describe the result and AI writes the code. On Tilda, code is only needed for embeds, and everything else is done in the visual editor. So a realistic scenario looks like this: structure and copy from AI, blocks assembled by hand, custom details through T123 with a check. If you need a site where all the code is written, see what vibe coding is↗.

The scripts in this article need a Zukko balance: it is kept in dollars and can be topped up from Russia from $1, via SBP with the payment page's 8.5% fee, MasterCard or cryptocurrency.

FAQ

Can a neural network create a whole Tilda site?

No. It helps with copy, structure and embed code, but the site itself is assembled in the Tilda editor.

Which neural network to take for site copy?

Any from the catalog: claude-sonnet-5-5, gpt-6.1-sol, deepseek-v4.1-flash and others. Compare results on one block and look at the cost in the calculation above.

Why does the pasted code not work in the editor?

Per the Tilda help, in edit mode code is shown as text. Publish the page and check.

Do I need to know HTML?

For embeds it helps to at least read the code the model returns. If an embed breaks the page, turn off the T123 block and go back to the code's author, that is, to your own request.