Fadi88/one-man-startup

Build a software startup without writing code. Not vibe coding: a proven team workflow (requirements, design, test-first build, QA, release approval) run by AI agents

Summary

One-man-startup is a repository template for running AI coding agents as a software team. Instead of relying on improvised prompting, it defines a staged workflow covering product discovery, requirements, design, test-first development, QA and security review, and explicit release approval.

What it is useful for

The template is intended for founders without a technical co-founder, but the README also identifies solo developers who want the discipline of a team without hiring one. It is written for Claude Code, Codex, Gemini CLI, and Cursor.

A typical feature workflow starts with “grill me on this.” A researcher examines similar products, the project manager questions the idea, and a feasibility check considers whether it can be built and operated within the stated budget. Once the specification is confirmed, a designer defines the screens and developers write tests before implementation.

Separate review, QA, security, and legal agents then inspect the work. This separation makes the workflow useful for engineers who want plans, requirements, decisions, and code changes to remain reviewable in Git rather than relying on an opaque chat history.

The template also makes operational decisions explicit. Deployments, payments, customer messages, and deletions require approval. If a release causes problems, the README says the last known-good version is restored first, with the cause investigated afterward. Everything produced is stored as ordinary code and documentation, so the project is not locked to one AI tool.

How engineers can use it

1. Prepare the documented prerequisites: an AI coding tool, a GitHub account, and Git and Python. The README notes that the AI tool requires a paid subscription, while hosting the finished application is a separate cost.
2. Follow docs/START-HERE.md: the documented path is to copy the template on GitHub, install Claude Code, open the copy, and say hello. The README does not provide shell commands, so use its step-by-step guide rather than inventing a setup sequence. It also points to docs/ASSUMPTIONS.md for what has been tested with each supported tool.
3. Complete the setup interview: describe the product, target customers and countries, data it will hold, monthly AI budget, intended code licence, and technical level.
4. Use the staged feature process: confirm requirements, review the design and implementation, inspect the tests, and approve any release rather than treating the first generated version as finished.
5. Customize deliberately: AGENTS.md contains the agents’ operating manual, while docs/HOW-IT-WORKS.md explains how the agents, checks, and tool adapters fit together.

The README explicitly warns that agents can be confidently wrong: checks catch many mistakes, but not all, and cannot validate the business idea. Its legal output is a checklist or first draft, not legal advice; payments, health data, children, and financial data warrant professional review.