PromptUI's core is a LangGraph state machine — a directed graph of AI agent nodes, each a specialised function with a clear job. Here's what happens the moment you hit "Build".
Step 1: Manager Node
The Manager receives your raw prompt and calls the AI model to produce a structured task plan: app type classification (dashboard, landing page, game, form, etc.), a list of files to create, which UI patterns to use, and a system prompt tailored to the app category.
This is why PromptUI handles "build me a Kanban board" differently from "build me a data dashboard" — different category blueprints, different component patterns, different quality checks.
Step 2: Developer Node
The Developer is the longest-running agent. It processes every task in the plan using your chosen model (Claude Sonnet 4.6 by default), writing files that appear in real time in the file tree via Server-Sent Events. If a task fails, the Developer retries it once before moving on — this gives transient API errors a second chance without blocking the whole build.
Step 3: Validator Node
The Validator runs a two-pass check. First: deterministic static analysis (syntax, missing imports, forbidden patterns). Second: an AI-powered semantic review using a fast model (Grok 4.1 Fast) to catch logic errors the linter can't see. This two-pass approach catches common mistakes cheaply without burning premium compute.
Step 4: Fixer Node (up to 3 passes)
If the Validator flags issues, the Fixer node (powered by Kimi K2.5) attempts targeted repairs — patching only the affected files rather than regenerating the entire project. We cap at 3 fix iterations to bound cost while giving complex bugs enough attempts to be resolved. Each attempt is accurately tracked and logged.
Step 5: QA Node
The QA agent does a final holistic review: does the output match the original prompt? Are all referenced components actually implemented? Is the UI coherent? QA acts as the last gate before the session is marked "done".
Step 6: Deployer Node
If you hit the Deploy button, the Deployer calls the Vercel API, creates the project, uploads all files, and saves the live URL back to your session. Your app gets a permanent .vercel.app URL you can share immediately.
Why Multi-Agent Beats Single-Prompt
Each specialist node operates with a focused system prompt and smaller, cleaner context window. The Manager doesn't need to know JSX. The Validator doesn't need the app category. Specialisation reduces hallucination and dramatically improves reliability on complex multi-file builds.