The full chain
As soon as the code touches money, customer data or a system already in use. All six steps, no exceptions.
// The AI plumber · how I work
Someone builds fast.I make sure it holds.
An AI writes code in minutes. It also gets things wrong, with the same confidence as when it is right. My job: get that code into production without breaking anything. I do not ask the AI to be right, I surround it with checks that do not fail in the same place it does.
Six steps · tests that prove · nothing in production without human sign-off
The principle
An AI is confidently wrong: nothing in its answer flags the mistake.
What proves that code works is the tests it passes, not the score an AI gives itself.
Nothing goes into production until a human has signed it off.
The six steps
Each step has its exit gate. Until it is cleared, we do not move on to the next one.
- 1Need
- 2Plan
- 3Critique
- 4Code and tests
- 5Review
- 6Sign-off
1 / 6
The need, with its success criterion
You describe what you want in one sentence. Together, we set how we will know it worked.
Exit gateA criterion that can be checked. Without one, I do not start.
2 / 6
The plan
Before the first line of code: what will change, the split into small batches, the risks spotted. Fixing a plan costs less than undoing code.
Exit gateA written plan you can read.
3 / 6
The plan critiqued by another AI
The plan is reviewed by a second AI model, from another vendor, trained differently. Two different models do not necessarily fail in the same place: what one lets through, the other may catch.
Exit gateEach objection is accepted or rejected in writing, with its reason.
4 / 6
The code and its tests, in batches
The code is written in small batches, and each batch comes with its tests, written alongside it. A test is a small program that checks on its own that an expected result is actually produced.
Exit gateThe tests pass on the machine, and their output is kept.
5 / 6
The review
A review on four points, always the same. Security: passwords in the code, exposed data, admin access, backups. Accuracy: money calculations, rounding, edge cases, dates and time zones. Compliance: where personal data goes, who can access it, how long it is kept. Rollback: if it breaks, how we return to the previous version.
Exit gateNo serious defect left open.
6 / 6
Human sign-off
I read the result, the test output and the review, then I say yes or ask for a fix. When the change affects your business, the decision is yours too.
Exit gateA written yes. Without it, nothing ships.
What I do not do
Four rules I keep on every project.
- Treat the score an AI gives itself (“9 out of 10, all good”) as proof.
- Sign off on the strength of a summary: I read the test output.
- Skip the plan critique because the plan looks fine.
- Promise defect-free code. I promise that defects are looked for before you find them, and that we know how to roll back.
The full chain, or a short version
The short version
An internal script, a web page, a one-line fix: the plan, the code with its tests, then sign-off. Human sign-off is never skipped.
For you, in practice
You may already have a tool written with AI, by an employee, a contractor or yourself. It works on demo day.
What remains to be seen is whether it holds the day a customer types an unexpected date, the server restarts, or someone else has to change it. I take over what exists, run the chain on it, and tell you in writing what holds and what does not.
A tool that needs to hold?
The offers and prices are public. For a specific case, write to me: we will look together at where to start.