Working with AI tools — plan, execute, audit

Three steps, in this order, and the order is the lesson.

Plan. Decide what you are building and how you will know it works, before any tool is involved. A model will implement the wrong thing very quickly and very convincingly, and you will not notice if you never wrote down the right thing. In my own work this is literal: PaySahbi has an engine design document covering tenancy, ledger rules, state machines and jobs, and the rule in that repository is that the document changes before the code does.

Execute. Use the tools properly. They are genuinely fast at work that is well specified and tedious, which is a large share of real engineering, and pretending otherwise is its own kind of dishonesty.

Audit. Read every line. Run it. Try to break it. If you cannot explain why a piece of code is there, it does not ship — not because a model wrote it, but because nobody understands it, and that is equally true of code a human wrote at midnight.

The students who skip the first step produce more code and learn less. The ones who skip the third produce code they cannot defend in review. The method exists to keep the speed without losing the engineer.