I read 123 of my own work sessions. 42% contained a correction.

I have been building software with AI for about four months. Four hundred and twenty commits on one client project alone. It felt like it was going well.

Then I did something I recommend to everyone and nobody does: I read my own transcripts.

All of them. 123 work sessions, sitting in a folder on my machine, complete with every message I typed and everything the AI sent back. I wrote a script to count what was in there, and then I read the ugly ones by hand.

The numbers

Work sessions analysed	123
Tokens processed across them	6.99 billion
Messages I typed by hand	1,333
Sessions containing at least one correction signal	42%

A "correction signal" is me saying some version of no, that's not what I meant — "no", "wrong", "I said", "not that", "why did you", "undo". Four sessions in ten contained one.

Sit with that for a second. In 42% of my working sessions, the work went in a direction I had to pull it back from. That is not an AI problem. If I hired a person and had to correct them in four out of ten tasks, I would not conclude that the person was broken. I would conclude that I was briefing them badly.

What was actually in the corrections

I sorted them into buckets. Three of them cover almost everything.

1. I asked for a thing without saying what "done" meant (about half)

"Make the dashboard better." Better how? Faster? Prettier? Fewer clicks? I would get back something that was genuinely better along an axis I did not care about, and then I would correct it — as if the wrong answer had been given to a clear question.

2. I let planning and building happen in the same breath (about a quarter)

I would start a session with a half-formed idea, think out loud, and start building somewhere in the middle of thinking. The build then encoded a decision I had not actually made yet. Every one of those corrections was me discovering my own opinion by watching it get implemented wrongly.

3. I asked a question that only I could answer (the rest)

Product decisions. Should this be one screen or two? Is this feature worth it? Those are not research questions and they are not code questions. They are mine. Handing them to an AI produces a confident answer with nothing behind it, and I would correct it because it was not what I wanted — which was never written down anywhere.

Notice what is not on this list: the AI hallucinated. It happens, and it is real, but it was nowhere near the top. The dominant failure was mine.

What I changed

Four things. All boring, all mechanical, all still running.

Plan and build never share a session

One session does nothing but research the problem and ask questions. It produces a written plan and an execution instruction — and then it stops. A different session, opened fresh, does the building.

This sounds like ceremony. It is the single highest-leverage change I made. A planner that cannot start building has no way to hide a vague decision behind a plausible first attempt. It has to write the decision down, and the moment it is written down I can see it is wrong.

Every question I have goes into one batch, at the front

Instead of a build being interrupted eleven times, all the questions get asked at once, before anything starts. I answer them in one sitting with my full attention. Then the work runs to the end without me.

The rule that makes this work: anything irreversible gets queued, not asked. Going live, deleting, sending a message, spending money — those go into a file for me to approve later, and the work carries on around them. I come back to a finished job and one short list of decisions.

A pass mark, written before the work starts

Before anything is built, we write down what "it worked" means as a number. Not "the site should be fast" — "the site loads in under two seconds, at 1440 pixels and at 390 pixels, with zero console errors."

You cannot correct something into being good. You can only measure it, and the measurement has to exist before the work does, or you will invent one afterwards that the work happens to pass.

Every mistake becomes an automatic check, not a resolution

This is the one I would defend hardest. When something goes wrong, "I'll remember next time" is not a fix. It is a promise made by a tired person about a future tired person.

I now have seventeen automatic checks running in my workspace. One of them refuses to let Vietnamese text ship if it has lost its diacritics, because that mistake went out once on a public page. One warns me when I mix planning and building in the same terminal. One blocks a claim about a user interface if there is no citation behind it. Each one exists because something burned me exactly once.

Ninety-one written rules sit behind those seventeen checks. The written ones I forget. The seventeen, I cannot.

What did not change

I did not switch models, buy a tool, or adopt a framework. Nothing in this post required a purchase. Every fix was a change to how I brief and how I verify — the two things that were always mine.

I also did not get the correction rate to zero, and I do not expect to. The point was never a perfect number. The point is that 42% was invisible to me until I counted it, and I had been telling people things were going well for months on the strength of a feeling.

The part that transfers to your business

If you have put AI into your company and it is "going okay", you almost certainly do not know your correction rate either. It is measurable. Export the conversations, count how often a human had to say no, not like that, and read twenty of the worst ones.

I would bet most of what you find is not the model being stupid. It is briefs with no definition of done, decisions nobody made, and mistakes that were resolved rather than automated.

Those are all fixable in a week, and none of them cost anything.

---

I audited 123 sessions across roughly four months of building — one live retail CRM, two games, an events automation stack and a dozen smaller tools. If you want the ugly version of this conversation about your own AI usage, email me.