How to start
Six steps from a bare server to your own code in production. Everything below is already installed — you are switching it on, not building it.
- Open the control panel — everything about this server is configured there.
- Pick the languages your application will ship in.
- Use the settings to describe your project: name, description, logo, SEO.
- Connect GitHub and push the server's code into your repository.
- Clone that repository onto your own machine, develop there, and push back.
- Press Deploy in the panel — the server takes your commit and rebuilds itself.
Recommended before you start
Neither blocks anything. Both save rework: the first switches on the thinking half of the product, the second changes the address of every page.
- An OpenAI key. Without a key the Quiz asks no questions, and without the Quiz there is nothing to describe your user cases with — so the coding agent refuses to build. That is why the panel treats the key as a RED requirement until the first cases exist, and as an amber suggestion afterwards: the site works without it, only vector search and the knowledge graph stay empty. The key is entered once and the cost goes straight to your model provider.
- Your own domain. While the site lives at a numeric address it has no certificate and no installable app — a browser grants those only over a secure connection. Moving to a domain changes every page address, so it is cheaper to do before they are indexed.
Before any code
Quiz — seven questions instead of a blank page
A project's most expensive mistake is made before the first line of code: the wrong thing gets built. Not through poor building, but because «where do I start» is hard to answer alone. Quiz turns it into a conversation: you answer, the model asks further, and out of it grows the list of scenarios the project is then built from.
The seed
Seven short questions: what the product is, who it is for, what a person should walk away with. Answer in your own words — dictation works. Everything after this grows from here, so a couple of sentences yields a markedly better result than a couple of words.
The conversation
Then one question at a time, in your language. There is an auto-quiz: the model asks five new questions and answers them itself, deepening the description — but anything it invented on your behalf is marked «Assumption», and you correct it. A guess passed off as fact would surface later, inside the finished scenarios.
The scenarios
The conversation is synthesised into numbered cases: who arrives, what they do, what must be true at the end. You read and confirm each one separately. An unread case is still the model's guess.
“And this is a product rule, not advice: while a single case is unconfirmed the panel keeps its alarm lit and the coding agent refuses to build. Building on an unread guess costs more than not building at all.”
Quiz — seven questions instead of a blank page
What this project is, technically
This is not a finished site but the Fractera architecture: one skeleton carries a landing page, a large SaaS and multi-level automation alike. Growth needs no rewrite — the data, authorization and panel layers are already separate, and each is built for load you do not have yet.
Code is not written here. A developer clones the repository to their own machine and works with Claude Code, which reads the instructions and skills that live inside the project: they state the rules, and machine checks refuse to let them be broken. The server only receives the result and rebuilds.
The skeleton is built for a project that will outgrow a million lines: every entity owns its folder, the shared layer does not grow with their number, and routes and permissions are declared where they are enforced. Stability here is not a promise but a consequence — a new page adds nothing to a central spine.