I can’t read code. That turned out to be the qualification.#
I’m not a developer. I can’t read a diff and tell you whether it’s correct — not in Swift, not in Python, not in anything.
Since the summer of 2025 I’ve shipped software for iOS, macOS, watchOS and the web. And I’ve built and published a Claude Code plugin that enforces an eight-phase, spec-first development workflow: eleven hooks, test-driven development with real artifacts, an adversary verification stage, an architecture-decision gate, a scope limit on every change, and a release pipeline that refuses to publish when the changelog and the version file disagree.
I couldn’t write a single line of that by hand. I designed all of it.
Why not being able to read code changes everything#
A developer using AI has a shortcut available: read the output, judge it, move on. It works. It’s also entirely dependent on a skill you either have or don’t — and it can’t be handed to anyone else.
I never had that option. So I had to solve the problem differently: make correctness observable from the outside.
That constraint produced a method:
- Evidence, not assertions. A test that claims to pass proves nothing. A test log and a screenshot do. No placeholders.
- Hooks, not documentation. Three projects in, the finding was blunt: written rules get followed roughly 60–70% of the time. Enforced rules, 100%. Documentation is a suggestion; a hook is law.
- Specification before code, with a human approval gate. Nothing gets built until I’ve said what “done” means in testable terms — and explicitly approved it.
- Bounded change. A hard cap on how much can change at once. Not because large changes are wrong, but because I can only judge what I can still see the edges of.
- Adversarial review as a gate. A separate pass whose job is to find the failure, with a three-state verdict. Nothing commits without a clean one.
None of that requires reading code. All of it requires knowing how to run something you can’t personally do.
Which is to say: it’s management#
Specify the outcome. Approve before work starts. Limit scope. Demand evidence instead of status reports. Document why you decided what you decided. Have someone independent try to break it.
You already do this with people. It’s the same discipline, applied to an agent that works faster, never pushes back, and will confidently build exactly the wrong thing if you were vague.
That’s the part developers often skip — they don’t need it, they can just read the code. And it’s the part you already have.
Who this is for#
Founders who need a working product before the first engineering hire — and want the option of building it themselves.
Domain leads sitting on a process everyone agrees is broken and nobody will ever prioritize. You know exactly what the tool should do. That’s now the hard part, and you already have it.
Managing directors and executives who want to judge what’s actually happening here — not from a deck, but by building something real and watching where it breaks.
The common denominator: you want to act, not to delegate. And you don’t want to learn to code first — you don’t have to.
How we work#
We work on your actual project. There is no curriculum.
- Setup. Tools, environment, repository. The boring part that decides whether the rest holds.
- Specification. Turning what you want into something an agent can build and you can verify. This is where the leverage is.
- Guardrails. The gates that make bad output impossible to ship rather than merely unlikely.
- Building. You drive. I sit next to you, in the sessions and between them.
- Review. How to judge output you couldn’t have produced — and what to do when it’s wrong.
Sessions are remote, 60 to 90 minutes, as often as your project needs. Some people book weekly for three months. Some book three times, get unblocked, and leave. Both are fine.
What you won’t get#
No AI strategy paper. No tool comparison matrix. No transformation roadmap. And no claim that this replaces engineers — it doesn’t. If what you build matters, you’ll want them. What changes is that you arrive at that conversation with a working thing instead of a wish.
Details#
250 € per hour, billed monthly, no minimum commitment.
Remote, English or German.
First conversation free — 30 minutes to see whether your project fits.
The framework is public and MIT-licensed: agent-os-openspec on GitHub. The projects built with it are documented under AI native Software Development.