It works. Until you touch it.
You built it with AI tools, and now every change breaks something that worked yesterday. Or there’s no software at all and the whole job lives in one person’s head. Either way, there’s nothing you can put in front of a customer and charge for.
Not a mistake you made. These tools are good at a first version and bad at a twelfth, and nobody mentions that at the start.
Why a week beats another prompt
A prompting tool builds a good first version in an afternoon, then starts from scratch on every ask. Week twelve here inherits what the first eleven taught us. A fresh prompt inherits nothing.
Five systems in production, in five industries that share nothing
Each one began as judgement somebody carried in their head, usually the person everything else was waiting on.
Two other engagements are left off deliberately — both were work on someone else’s platform rather than systems built here.
Those four numbers are SignalsAPI’s, which is our own product. Client systems run in the clients’ own accounts, so their figures are theirs to publish rather than ours.
“They can hand the codebase to Claude now and add features without breaking anything.”
Client, KIMA — real estate, Germany. Quoted with permission. Our description of that project, not his: it arrived vibe-coded on top of a spreadsheet, and was rebuilt on a queue.
Every option is either too small to ship or too big to start
Everyone arrives by the same route, and the whole route is already priced. Pick any of them.
€5,000 buys a version. A version is one week.
Stop after any of them. No notice, no setup fee, no minimum term.
Book twenty minutesFour answers: what’s buildable, what we’d build first, what’s hardest, and what we’d refuse.
Working software every week, and you keep it if we stop
No spec to write, no roadmap to approve, no discovery phase to pay for. The specification is what comes out.
One client bought forty weeks, one at a time
A company founded by a former UK cabinet minister needed one executive assistant’s judgement about meeting slots turned into software. It lived nowhere but in her head. They had interviewed seventy people off Upwork and stopped believing anyone existed; one of the seventy could show, during the interview itself, that it was buildable. That is still exactly what these twenty minutes are.
- Interviewed before this engagement
- 70
- First working beta
- 1 month
- Scheduling round-trips afterwards
- One, on average
The assistant whose judgement the software absorbed didn’t lose her job. She ended up running the company, which tends to be how it goes: the person holding the knowledge usually understands the business better than anyone else in it.
That is still exactly what these twenty minutes are.
Paused rather than churned — the product was good enough to raise on. Those forty weeks were bought at the rate then in force, so multiplying by today’s €5,000 gives a number nobody paid.
Three things we talked a client out of buying
“Don’t build this” is one of the four answers on the call. Here is what it has looked like when the answer was given.
- A vector database. The corpus fitted in Postgres. pgvector did the job, and the money went on parsing instead.
- Bulk outreach over a shared consumer channel. One flagged number takes it down for every firm on it, so it went out staged, per firm.
- A second CRM integration in month one. The first write path was two weeks old.
We don’t ask for a spec, because the people who hold the knowledge can’t write one
Ask a domain expert how the software should work and you get the manual version described back, which overconstrains it before a line is written. So the first version is the question.
Owns the architecture standard everything here is built to. Twenty-five years putting a stable interface in front of chaotic reality: payment providers in 2010, job boards in 2021, model vendors in 2025. Whatever you describe on the call is unlikely to be the strangest thing.
Don’t book if any of these is true
- You’re technical. You’ll do this yourself, or hire developers you can lead.
- You want a fixed price against a frozen spec. Not a preference — the method scopes by building.
- Nobody on your side knows the domain. We can build it. We can’t know what it should do.
- One prompt would finish it. Go and finish it. We’d rather lose the five thousand than be the reason you spent it.
- You need someone to run developers you’ve hired. This replaces a team rather than managing one.
- The money isn’t there yet. Go and sell the presentation first. We’ll still be here.
The honest concession: hiring leaves the capability inside your company permanently and this doesn’t. If what you need is an engineer who’s still there in three years, hire the engineer. We’d say so on the call anyway.
We work 08:00–17:00 CET, remotely, and don’t take engagements more than two hours off that.
Twenty minutes, and you’ll know what we’d build first
Describe the thing you want to exist. “Don’t build this” is one of the four answers, and you get it in twenty minutes rather than in month four. If someone else is the better fit, we’ll say so and tell you who.
Book twenty minutesRather write first? [email protected]