Paste anything. Get a blueprint.
Not limited to code. UnBuild analyzes repos, websites, products, APIs, and plain-text ideas — then produces everything an AI needs to build it.
GitHub Repos
vercel/next.js
Full codebase analysis — file tree, dependencies, architecture patterns, entry points. Produces implementation blueprint with exact commands.
Any Website
stripe.com
Renders the page with a real browser. Extracts design system, navigation, components, and infers the full backend architecture.
Product Pages
adafruit.com/product/5613
Component teardown, bill of materials, assembly instructions, and open-source alternatives for hardware and physical products.
Ideas & Descriptions
Build me a Duolingo clone...
Describe any app, tool, or product in plain text. Get a researched blueprint with tech stack, data model, and phased build plan.
How UnBuild thinks
Four phases. Chain of Verification. Every claim evidence-based. The output works with any AI coding tool.
Fetch & Extract
Real browser renderingFetches your input using Playwright (a real browser). For repos: reads source code, configs, schemas. For websites: renders JavaScript, extracts CSS, crawls pages. For products: scrapes specs and components.
Understand the Product
Chain of VerificationThe AI doesn't just describe what it sees — it figures out what the product DOES, what backend it needs, what data it stores, and what open-source alternatives exist. Every claim must be backed by evidence.
Architect the Build
Multi-pass LLM analysisThree separate AI passes: architecture analysis (JSON), implementation blueprint with phased agent decomposition, and a standalone CLAUDE.md. Each pass builds on the previous one's decisions.
Generate the Prompt
800-1100 word IBPA dedicated Initial Build Prompt — the single document you paste into any AI coding tool. Instructs the AI to research first, plan the architecture, build in phases with parallel agents, and test with Playwright after each phase.
Four deliverables, one analysis
Everything an AI coding tool needs to build from scratch. Works with Claude Code, Cursor, Codex, Windsurf, or any other.
2000-4000 words
- •Product DNA — what it IS and DOES
- •Multi-phase build plan with parallel agents
- •Data model, API endpoints, auth flow
- •Data seeding strategy (where initial data comes from)
800-2000 words
- •Complete design system (colors, fonts, spacing)
- •Dev commands, build order, test strategy
- •Component specs with states and behavior
- •Product-specific gotchas and warnings
800-1100 words
- •The single prompt you paste into ANY AI tool
- •Research phase: find similar open-source projects
- •References CLAUDE.md instead of embedding it
- •Agent decomposition for parallel work
Structured
- •Evidence-based tech detection (not guessing)
- •Component graph with dependencies
- •User journeys with inferred API calls
- •Reproduction challenges specific to this project
Works from your terminal
Analyze repos, websites, or anything — right from the command line.
MIT Licensed · Open Source · Contributions Welcome
View on GitHub