How I work
The same method ran a $23M portfolio, a $35M integration, and three AI systems, it does not much care what the subject is.
- 1
Find the constraint that does not move
Every problem has one rule that survives all negotiation, in accounting it is that debits equal credits, in insurance it is that you cannot move fast by breaking things, and with a model it is that the model can be persuaded. Start there and most of the design finishes itself.
- 2
Scope is a decision
What I will not build gets written down first, the Sabit page for example lists what it is NOT. At Rock Wallet, due diligence said no to deals as often as yes.
- 3
Interrogate the work
Outputs get questioned until they give evidence, and a claim without proof gets treated as debt. The delivery framework at Empire Life stuck because I kept running it and correcting it until it held.
- 4
Put a name on it
Every override and every refusal carries a name and a record next to it. The AI builds run this way, however so did every program before them.
About the tooling
The three systems on this site were built with AI writing the code under direction, the specs, the acceptance criteria, and the refusal to claim what is not demonstrated stay with me. If that division of labor sounds familiar, it just so happens to be how a program office already works.