How the V2 engine actually works

Asking one model to generate a complete app from a paragraph gets you something that runs about half the time, drifts in style every retry, and struggles to keep multiple files internally consistent. Models are brilliant at language but don't have a stable sense of project structure.

Virex pulls structure out of the model and into the surrounding system. Three lanes (Standard Engine, Deep Engine, Engine Plus) handle different kinds of work. The engine knows what a project should look like, what sections live in our curated library, what the design system rules are, and how to assemble the pieces. AI gets used only where it does the most good — language and creative section code — and stays out of the parts that need to be predictable.

Result: same prompt twice gives different but coherent output. No drift, no retry roulette, no half-shipped repo.

Claude Haiku parses your prompt into structured intent: project type, vibes, audience, tone, formality, energy, color hints, features. Works in any language — Dutch, Spanish, French, German, English — same quality each time.

The engine picks one of three lanes. Site (1) for marketing pages and landings — flat structure, curated sections. App (2) for apps with auth, dashboards, and routing — sits on top of a foundation template. Creative (3) for prompts that don't fit the curated library — full LLM section generation.

A preset is picked based on intent (corporate-navy, warm-amber, neon-cyber, etc.) and customised with your brand color hint. Sets up the entire --virex-* CSS variable namespace so every section renders consistent across the build.

Curated section library, scored against your intent. Same prompt twice produces different — but coherent — combinations, not the exact same output every time. Sections are real, hand-tuned, and use the design tokens from stage 3.

Claude Sonnet 4.6 rewrites copy and props per section, conform the tone the intake captured. Section structure stays from the curated source; only the text/image/icon slots change.

For prompts the curated library can't handle — “asteroid rain hero with 3D globe”, “particle field background with custom CTA card” — the engine generates an outline, then each section from scratch with validation. Failed sections retry; on Premium and Pro, hard failures escalate to a stronger model automatically (Engine Plus).

Foundation tree (path 2) gets copied; sections layer onto a route group (`(virex-marketing)/`); mobile screens auto-wire into the navigator (Expo Router or React Navigation). Static preview HTML, build report, and download zip get prepared.

Every section passes a validation check (parses, has required pieces, sane length, no banned vars). The engine emits live progress events you see in the workspace. Final output: a downloadable Next.js or Expo repo, real and self-contained.

The engine isn't static. The curated library grows month over month, foundation templates expand, selection logic gets smarter. For users that means builds become faster and better over time without you doing anything.

We aim for the best output at the lowest cost so we can keep delivering more value at fair pricing.

AI handles language: parsing prompts in any language, rewriting copy with tone, generating creative section code when the library can't help. Everything else is engine: section structure, design tokens, foundation templates, validation, assembly.

That split is deliberate. Language is hard to get right deterministically and easy for AI. Project structure is the opposite. By drawing the line where AI is strongest, the engine produces output that feels both polished AND reliable. Same prompt twice gives different copy and different section combinations, but always a coherent, deployable repo.