AI-powered development means the model is not just autocompleting code — it reads your prompt, decides which files to create or change, scaffolds a full-stack project, and reasons about updates over the entire life of the app. Understanding how that works helps you get better results with fewer tries.
What happens on a generation
- The platform classifies your prompt — a full build, a tiny one-file task, or an update to an existing project — so it can pick the right approach and model.
- It routes to the provider tied to your selected model (Claude, Grok, Gemini, Groq, or Ollama).
- The provider streams back an explanation into the chat and a set of files onto disk.
- Files land under
GeneratedApps/{userId}/{projectGuid}, dependencies install, and a version is saved so you can roll back.
Updates vs. new projects
- Inside an open project, a follow-up prompt is an update: the model receives the existing code as context and changes only what is needed, keeping the rest intact.
- With no project open, a prompt produces a fresh scaffold instead.
- This is why you build conversationally — each message refines the same app rather than starting over.
Why context matters
Because each update carries the current file state, the AI stays coherent across many turns:
- Rename a field, then add a page that uses it, then wire validation — without contradicting earlier work.
- The model sees your real files, so it matches your existing naming, structure, and style.
- It knows what it already built, so it won't duplicate a screen or re-create a model you have.
Getting the best results
- One focused change per message. "Add pagination to the orders list" converges faster and cleaner than a paragraph bundling five requests.
- Show, don't just tell. Paste an error, a sample payload, or a screenshot description so the model works from facts.
- Point precisely — name the file or feature; in the desktop editor, use @-mentions.
- Iterate, don't restart. If a result is close, refine it; a fresh prompt throws away the context you built.
Common mistakes to avoid
- Vague prompts. "Make it better" gives the model nothing to aim at — say what "better" means.
- Bundling unrelated changes. Split them so a bad one doesn't force you to roll back the good ones.
- Ignoring the version history. If an update goes wrong, roll back and re-prompt rather than fighting it forward.
FAQ
Why did it change files I didn't mention?
Usually the request implied them (e.g. adding a field touches the model, API, and form). Be specific, or in the desktop editor @-mention the exact file to keep the change contained.
Can I undo a generation?
Yes — every generation saves a version, so you can roll back and try a different prompt or model.
New chat or follow-up message?
A follow-up in an open project updates that project. Starting a chat with no project open creates a brand-new scaffold.