Ten rules, and each one cost me something.
I manage AI the way a decent engineering manager manages a team: spec first, verify everything, automate the guardrails, and never accept "looks fine". Below is what you are actually buying. Seventeen of these rules now run as automatic checks, so I cannot skip them on a tired day.
Plan and build never share a room
One session researches the problem, asks every question it has in a single batch, and refines a written plan until it is genuinely solid. A different session builds from that plan. The planner never gets to touch the code, so the plan cannot hand-wave.
You approve once, at the start
Approvals go at the front and the back, never in the middle. You say go, then it runs to the end while you do something else. The only interruptions are things that cannot be undone.
Anything irreversible gets queued, not asked
Going live, deleting data, sending a message, spending money. Those get written into a file for you to approve in one sitting, and the work carries on around them.
Files are the memory, not chat
Every project gets the same four files: what's pending, the plan, the current state, the report. If a conversation is lost, nothing is lost.
Never guess — verify, then cite
If I am not sure, I say so out loud. If I got something wrong, I say that too. A confident wrong answer costs a client far more than an honest "I need to check."
Read before write. Back up before anything destructive.
Twenty-five dated backups sit in one client's project. None of them were ever needed. That is entirely the point.
A pass mark before the test
Before anything is built, we write down what "it worked" means as a number. Otherwise you only ever get a story afterwards.
A mistake gets a systemic fix, not a patch
Ninety-one written rules and seventeen automatic checks exist in my workspace because each one burned me exactly once. The check runs by itself now, so I cannot forget.
Live customer systems never get pushed to directly
A preview link first. You look at it on your own phone. Only after you say yes does it reach the thing your staff use.
Three tries maximum
If an approach fails three times, the approach is wrong. Change the strategy, don't repeat the attempt louder.
1 · One call, then a written plan
Thirty minutes. You describe what is broken; I ask the annoying questions. Then you get a plan in plain language with a number attached to each promise — no jargon, no estimate ranges hiding a guess.
2 · You approve once
One approval at the start. Then I build, and you get a preview link you can open on your phone. Nothing touches your live system until you have looked at it.
3 · A report, not a status update
At the end you get what was built, what the numbers say, what went wrong, and what I would do next. Including the parts that did not work — those are the useful ones.
Ask me what I got wrong last month. I keep a list.