I have no engineers. No product manager. No full-time designer. No co-founder.

I’m building Nebula — an AI assistant that gives you focus time back — and my entire team is nearly exclusively me and a collection of AI systems.

I wrote “nearly” — after being completely alone for 6 weeks, three former colleagues volunteered to help me, and this is important as AI can’t do user interviews as well as humans, won’t challenge you like humans, and is currently not capable of creating extraordinary designs. Thank you so much Vishal, Theresa and Malte!

This is not a “look how productive I am” post. This is an honest account of what it’s like to build a real product in 2026 with AI doing the work that used to require five to ten people.

The Background

I’ve spent over 30 years in the software industry in various roles. I know what a well-run product team looks like.

And then I decided to build without one.

Not because I think teams are unnecessary. But because the economics of early-stage building have fundamentally changed, and I wanted to see how far one person could push it. And one person can only push it far for a product of Nebula’s complexity if the “Dark Factory” pattern — a fully automated coding setup where I never write or review code — is used.

Here’s the org chart of my startup I’m aiming to have, and I mostly have. My next blog post will show where this still breaks.

Product Manager

New features are discussed with Claude, written into a markdown file, and then reviewed by Claude Code before implementing. It’s interesting how much more clarity comes from a peer review from pretty much the same person ;)

Existing features live in a 40-page specification document. Every design decision, every edge case, every architectural trade-off — written down, versioned, and used as the primary interface between my product thinking and AI-assisted implementation. A CI run checks every week if the spec and code match.

Engineering Team

Claude Code with custom skills, automated workflows, and a CI/CD pipeline that runs on self-hosted GitHub Actions runners. The code gets written, tested, linted, and reviewed without me touching it. Velocity is crazy — by the time I’m writing this (10:30 am), over 25 deploys to production have happened already.

Designer

I started with a design system Claude helped me create, and we are now migrating into a new one created by Theresa. Every night, compliance with the design system and a general UX review is performed by CI.

QA Department

This is the most important part for the Dark Factory pattern and is worth a separate blog post. Here are the basics:

  1. A multi-agent test harness. Six synthetic personas working at a fictional company called Crestline Analytics, generating realistic communication patterns across 10 channels, DMs, and emails.
  2. Around 1,000 unit tests run after every commit.
  3. And I make heavy use of the Claude QA skill to verify end-user behavior.

This is Part 1 of a series on building a startup with AI as your only team. Next week: what actually works well now, and where either AI or my setup have limitations.