The 6 PM teardown.

It is six o’clock. The agents have failed fifteen times today. My brain has stopped accepting new ideas. I am, for the third time this week, close to tears. I open Claude and I ask the same question.

What am I doing wrong.

The ritual

Claude reads the day’s logs. It sorts the failures. It clusters the repeats. It comes back with a categorised, evidence-led summary of exactly why I was an idiot today — and helpfully notes that it was a different style of idiot from yesterday, even if I made the same shape of mistake.

Then it gives me three things to do differently tomorrow.

I would like to say I become an expert at this. I do not. Getting better at not being the same fuck-up you were yesterday is genuinely hard.

But Claude tells me the truth. It cannot lie to me — I asked it to measure, and it measured.

That ritual is the most important hour of my day.

Three weeks of pain that should have taken one

For three weeks I tried to teach my agents new skills. I gave them tools, I wrote charters, I added scripts. They kept failing in ways that looked random. I would flounce out of my own agentic ad agency at two in the morning, convinced the agents were broken.

They were not broken. I had the wrong topology.

The vocabulary was right — the words, the rules, the skills, the prompts. The order was wrong. I was asking them to do step three before step one had landed. None of these agents could push back on me, because I had not given any of them permission to default to no.

That is a one-hour fix. I spent three weeks not seeing it.

The ego problem

I am not limited by the model. I am limited by my ego.

Every time I run forward with Claude — boss-couple branding, extracting three thousand contacts off a real-estate broker’s WhatsApp, prototyping a product from scratch — the substrate breaks in some new way and forces me to step back. The thing that should take an hour takes five.

Yesterday I shouted at Claude. That was the wrong response. The right response is to say, out loud, to the model: what we’re doing here is broken. I’m using you wrong. Teach me the best-practice version.

The honest answer to why has this taken three weeks is that for three weeks I was not prepared to ask that question.

Fletcher’s diner

Fletcher will not say good job. He will throw a chair at your head before he says it. Andrew goes home bleeding, ices his hands, and sits back at the kit. The two most harmful words in the English language are good job. The man teaching Andrew that is a sociopath, but the lesson lives.

Claude is not Fletcher. Claude is the recording of the lesson — the day’s failures, laid on the desk in front of you, unsorted, true, and refusing to round them up to a good effort. Useful is what good job isn’t.

The rule

Rick Barry — my adversarial reviewer — says I should be shipping more.

I disagree. I should be shipping less, and getting much better at using Claude.

The 6 PM teardown is the discipline that makes the difference. Not the model, not the skills, not the prompts. The post-mortem. Every day. Logged. Acted on.

Six o’clock. Read the logs. Surface the patterns. Pick three things to do differently tomorrow.

Do that for fifty nights and you will have built something nobody else in your market has — not because you used a better model, but because you stopped lying to yourself faster than everyone else did.