People think working with AI is about finding the clever prompt and watching something appear. Mine doesn’t look like that. On a normal day I have two or three sessions running at once, each one chewing through a big piece of work I have handed it, while I do the part the machine still can’t: decide whether any of what comes back is actually right.
The typing is the easy part. The judging is the job.
I delegate in big chunks. Not “write me this one function” but whole pipelines, end to end, with a clear outcome and room for it to plan the route itself. Then I let it run. I don’t hover. The sessions usually end the same way: a clean commit, and I move on to the next one. That part genuinely feels like magic, and it is useful.
But here is what I have learned never to skip. The machine is a brilliant junior. Fast, tireless, has read more than I ever will, and every so often confidently, completely wrong. The errors that matter are not the obvious ones. They are the answers that read beautifully and are hollow underneath.
So I have built my way of working around catching those.
When it hands me a number, I stop and ask where the number came from. More than once a clean, confident figure has dissolved the moment I pushed on it. Now I make it show its working before I believe anything.
When it diagnoses a problem and proposes a fix, I make it tell me its hypothesis and how we would prove it wrong, before it writes a single line. A confident wrong diagnosis is the most expensive thing in the room. It looks like progress, right up until it ships.
None of this is me being precious about craft. It is simpler than that. The quality of what AI produces tracks almost exactly with how well the person using it can judge the output. If you cannot tell good from impressive, the machine will happily give you impressive. Hand it your judgement and you get confident mediocrity, faster than ever.
So I do not let it forget. Every decision that comes up twice becomes a written rule. Every useful pattern becomes a small, reusable tool. I have built myself a whole set of little skills that each do one thing well. The machine remembers nothing about yesterday, so the memory has to live in the system, not in the machine.
And it works the other way too. I am not only giving the work a memory, I am building my own. I keep a second brain, a vault I have grown for years where everything I work out gets written down and linked to everything else. One of those little skills does nothing but help me learn: when Claude suggests an approach, or when I make a decision and it pays off or blows up in my face, the skill helps me pull the real lesson out and file it where I will trip over it again when I need it. The machine forgets every session. I make sure I don’t.
This matters more than it sounds. There is a quiet trap in all of this: if the machine always does the thing, you slowly stop knowing how to do it yourself. You keep the job and lose the craft. I would rather go the other way. Kept like this, with the judging and the learning firmly in my own hands, these tools are the fastest way to get better at something I have ever found. I am skilling up, not down. On purpose.
That is the reason I can sit with a company and tell them whether their AI content actually holds up. I am not guessing. This is how I work all day: building these systems from the inside, getting caught by the failure modes, and learning where to check. The tools are extraordinary. They will also let you down quietly, in the places you stopped paying attention. My job is to keep paying attention.
If that is the kind of help you need, here is how we can work together.