Kerber.ai: A Venture Studio Powered by AI

Kerber.ai started with a practical question: how small can a company become before it stops being able to build serious products?

I had spent a year working with Cursor, Claude Code, Copilot and more abandoned configurations than I want to admit. The useful discovery was not that AI writes code ten times faster. It often does not. The larger gain was in all the work surrounding the code: setting up a repository, tracing an unfamiliar system, preparing tests, writing documentation and keeping several small workstreams moving.

The setup

Kerber.ai is a venture studio with a deliberately small structure:

• me, with more than twenty years in product and design
• Henry, the AI operating layer I use for research, development and operations
• a set of workflows that make the work visible enough for me to review

Henry is not a virtual co-founder and I do not pretend it has judgement of its own. I define the objective, the constraints and what counts as done. It can break that work into smaller tasks, route them and bring back artifacts. I still decide what deserves to exist and what is safe to ship.

The workflow

The exact tools change, but the working loop has stayed stable:

1. Start with the outcome, not a list of functionality.
2. Review the plan before paying for implementation.
3. Inspect the diff and run the checks.
4. Turn repeated failures into instructions, tests or constraints.

Honest truths

The setup cost is real. Useful agent workflows take time because the hard part is not prompting. It is deciding where autonomy ends, what evidence a task must return and how a bad result fails safely.

Experience helps, but not because senior people know secret prompts. It helps because they recognise when a plausible solution is pointed at the wrong problem. That judgement is still the scarce part of the studio.

I do not care who typed a line of code. I care whether the product should exist, whether the implementation can be understood and whether the evidence supports shipping it. A small studio can now execute more ideas than before. The discipline is learning which ones not to execute.