Put your own photo into a website template. Watch what happens.
The hero that looked right in the demo now cuts a face off at the chin, because the template was built for a wide image and yours is a portrait. Your headline is four words longer than the one in the demo and wraps onto a third line. The font that looked elegant under "Elevate your brand" looks cramped under "Plumbing and heating in downtown London".
That is the moment a template starts fighting you. Today we are introducing Seeds, a website template alternative built for a workflow where an AI agent does most of the building. A Seed is not a site template with 20+ pages and variants to decide on. It is the guidance an AI agent needs to create something truly unique.
Want to try it yourself? Go ahead and try seeds
Why Seeds?
A template is a finished design. Every choice in it was made for the demo content: the image ratios, the headline lengths, the number of services, the colour of the buttons. Your business did not make those choices. The template designer did, for a company that does not exist.
The result is a lot of sites that look like each other.
Templates are marketed as a one-way ticket to a quick launch, not to standing out. When you find yourself fighting the template instead of progressing, that is the signal that you have already outgrown it.
So a template is easy on day one and expensive on day three. The friction starts the moment you want to make it yours.
Wireframes worked better
For ten years, our fastest workflow did not use templates at all. We used a library of wireframe sections. A navbar, a hero, a feature grid, a footer: plain blocks in black and white, with placeholder text and grey boxes for images. We dropped them into a page, filled them with the client's content, then styled them to the brand.
This works because a wireframe section makes no design decisions for you. It only gives you structure.
But a section is not a site. You still need to add styling and content, and this is where wireframes stopped working for us in AI workflows. AI-generated layouts often end up looking the same, and when you feed the model a common wireframe, it never escapes that look.
So wireframe sections sit at one end and templates at the other. One gives you structure with no design. The other gives you a designed site with little room for change.
Introducing Seeds
A Seed sits in the middle. It is a subset of a home page: a header, a hero and a few content sections. The goal is only to set the stage: the design and the style. All of it is filled with placeholder text and placeholder images. The colours are variables, not fixed values, so a brand change is one edit, not a hundred.
Nothing in a Seed is designed to be kept. Everything in it is designed to grow. That is the difference between a Seed and a template.
Take the first one, Design studio. Open the live demo: a header, a hero, selected projects and a contact section, and nothing more than that. Projects are a content collection, so your twelfth project is a new item in Content, not a new page you build by hand. Every project in the demo is a placeholder, ready to be filled by you or the agent.
Compare that with the direction AI builders have taken. Lovable now sells templates sorted by type: law firm, wedding, tech conference, editorial. The pitch is "Start from something that works" (Lovable Academy). That is a real improvement over a blank prompt. But a Lovable template is a finished design, like every template before it.
Why an AI agent needs a Seed
If you have asked an AI tool for a website with a vague prompt, you know what came back: Inter, a purple gradient, three columns, rounded corners. We wrote about why most AI generated sites look the same. The short version is math. The model gives you the most likely design, and the most likely design is the least distinctive one.
Anthropic calls this "distributional convergence". Without direction, the model "samples from this high-probability center". Their advice is the part that matters here: "the more specialized the task, the more context you need to provide" (Anthropic).
TechRadar reached the same conclusion from the user's side. "Most outputs look generic, templated, and unmistakably 'AI-generated'", and the usual cause is "generic instructions". The fix: "You need to feed your AI website builder with rich context and specific direction" (TechRadar).
A Seed is that context, in the form an agent reads best: real files. When you load a Seed, the agent does not have to guess what a design studio site should look like. The type, the spacing and the first sections are set. The agent grows every other page from that start, and spends its effort on the part that should be yours: your photos, your words, your story.
That is how you make an AI website not look generic. You do not ban the defaults one by one. You give the agent a better starting point than the average.
How it works
- Pick a Seed on the Seeds page. Each one has a live demo.
- Load it. You now have a starting point in Splitleaf: a folder on your computer, with real HTML, CSS and content, and nothing the Seed does not need.
- Your agent (Claude Code, ChatGPT, or any model on OpenRouter) now uses this well-defined Seed to create the other pages and grow the project into something unique.
- Refine it together in the Designer. Change the type, change the palette, remove a section, add a page. Nothing in a Seed is locked.
We have three Seeds today: a design studio, a conference and a SaaS product. More are on the way.
Templates were built for people who had to fill in a site by hand, so they shipped finished. A Seed is built to be grown by an agent with you, so it ships small and ready to change.