Intent0 - Intent to Development Logic Contract

Gemini-powered MVP that turns vague ideas into dev-ready logic contracts. AI researches your domain, asks smart follow-ups, outputs specs, risks, tests & tasks. Export to MD for any dev workflow.

Comment

Inspiration

Every application/vibe coding project starts with an idea—but turning that idea into something developers or AI agent can actually build is surprisingly hard. Miscommunication between stakeholders and developers costs companies billions in rework every year. We asked: what if AI could act as a "Technical Co-Founder" that deeply understands what you actually need, challenges your assumptions, and produces a precise development contract?

When Google released Gemini 3 with Google Search grounding, we saw the perfect opportunity: an AI that can research your domain in real-time before asking smart questions, ensuring the output is grounded in industry best practices, not just generic suggestions.

What it does

Intent0 transforms vague ideas into dev-ready Logic Contracts through a structured pipeline:

  1. Requirement Input: User describes their idea in plain language
  2. Domain Research: Gemini searches the web to understand industry context, best practices, and common pitfalls
  3. Progressive Questioning: AI asks up to 5 targeted follow-up questions, each building on previous answers
  4. Contract Generation: Outputs a complete development spec including:
    • Clarified requirements with v1 scope
    • Risk assessment with mitigation strategies
    • Acceptance criteria with test cases
    • Task breakdown for implementation
  5. Export: Download as Markdown for use in any dev workflow

The final output is a "Logic Contract" that any developer, team, or AI coding tool can execute.

How we built it

Tech Stack:

  • Frontend: React + TypeScript + Tailwind CSS
  • Backend: Express.js + Node.js
  • Database: PostgreSQL with Drizzle ORM
  • Auth: Clerk
  • Deployment: Railway

Gemini 3 Integration:

  • Google Search Grounding: Domain Research Agent performs 5-round iterative web searches to gather industry insights before questioning
  • Structured JSON Output: All agents use responseMimeType: "application/json" with Zod schema validation for reliable, type-safe results
  • Multi-Agent Architecture: 5 specialized Gemini-powered agents work in sequence:
    • Domain Research Agent (web search grounding)
    • Progressive Diagnostic Agent (context-aware questioning)
    • Clarification Agent (requirement expansion)
    • Risk Review Agent (technical/business risk analysis)
    • Acceptance Testing Agent (test case generation)

Challenges we ran into

  1. Prompt Engineering for Consistency: Getting Gemini to output consistent, schema-valid JSON across all agents required extensive iteration. We solved this with Zod validation and fallback handling.

  2. Progressive Questioning Logic: Designing an AI that asks follow-up questions based on previous answers—without being repetitive—was tricky. We limited it to 5 questions max and built context chaining between rounds.

  3. Balancing Research Depth vs. Speed: The Domain Research Agent could search forever. We optimized it to 5 focused rounds covering: domain identification, best practices, case studies, pitfalls, and synthesis.

  4. Grounding vs. Hallucination: Ensuring AI outputs are grounded in real research (not invented facts) required careful prompt design with explicit anti-hallucination instructions.

Accomplishments that we're proud of

  • End-to-end working MVP deployed on Railway with real Gemini 3 integration
  • Google Search Grounding that provides real industry context—not just generic AI responses
  • Multi-agent pipeline where each agent builds on the previous one's output
  • Progressive questioning that adapts to user answers in real-time
  • Export to Markdown for seamless integration with any development workflow
  • Multi-language support: Automatically detects and responds in user's language (English/Chinese)

What we learned

  • Gemini's Google Search grounding is a game-changer for building AI apps that need real-world context
  • Structured JSON output with schema validation is essential for reliable LLM integrations
  • Progressive, context-aware questioning produces much better results than asking all questions upfront
  • The "Technical Co-Founder" persona helps frame AI as a collaborator, not just a tool

What's next for Intent0

  • Integration with dev tools: Direct export to GitHub Issues, Jira, Linear, Notion
  • Version history: Track how requirements evolve over multiple iterations
  • Team collaboration: Multiple stakeholders contributing to the same Logic Contract
  • AI follow-up: Let Gemini ask clarifying questions about specific sections after generation
  • API access: Enable other tools to use Intent0 as a requirements-gathering service

Built With

Share this project:

Updates