Skip to main content
PET BRAINS

// tool · 01 / 03

MVP-builder

Build MVPs with an AI agent that verifies its own work. A Document-Driven Development framework for Claude Code with TDD cycles and self-review.

// what is it

A Document-Driven Development framework for Claude Code.

MVP-builder is a workflow for building greenfield projects with Claude Code. It treats specifications as the input and code as the output: a structured chapter list, generated from a PRD, drives the agent through the build — section by section, with TDD cycles guaranteeing tests and a self-review loop catching the agent before it commits broken work.

Session memory keeps context across runs. The agent doesn't lose track of decisions made three commits ago; the chapter list is the persistent plan, the running diff is the persistent state. Open source, MIT-licensed, no platform lock-in.

// the problem

AI demos stop at the prompt.

Most AI-coding content stops at the demo: the agent generates a feature, the screen recording ends, the second-day work never appears. The verification, the rework, the regression a junior would have caught — none of it on camera, none of it shipped.

MVP-builder is built for the second-day work. Specs generate code, not the other way around. TDD cycles run before any commit lands. Self-review loops check the agent against the chapter list before the diff is final.

The result is a build the agent can defend in code review — not a screen recording.

// how it works

Five steps from PRD to shipped code.

  1. // 01

    Define the PRD.

    Write a product requirements doc — what's being built, why, for whom. MVP-builder reads it as the input to every later step.

  2. // 02

    Generate the chapter list.

    The agent decomposes the PRD into a structured chapter list — features, user stories, acceptance criteria. The list is the persistent plan.

  3. // 03

    TDD cycles per chapter.

    For each chapter, the agent writes failing tests, then implementation, then runs the tests. No code lands without a passing test that wasn't written first.

  4. // 04

    Self-review before commit.

    A separate review loop checks the diff against the chapter list and the acceptance criteria. The agent rejects its own work before it asks for human review.

  5. // 05

    Memory across runs.

    Session memory persists: chapter status, open questions, prior decisions. Restart the session and the agent picks up where it left off.

// who it's for

Indie hackers, technical founders, senior devs shipping with AI.

MVP-builder is built for builders past the demo phase — the people already using Claude Code or Cursor day-to-day, looking for a structured workflow that produces production-grade code instead of prototype-shaped output.

Best fit: greenfield projects. Worst fit: one-shot scripts and existing-codebase refactors — the chapter list is overhead unless the project is large enough to need a plan.

// compatibility

Compatibility

// get started

Read the README, run the workflow.

Get notified when MVP-builder ships a major update.