Field note
What founders actually correct
Eight kinds. The distribution is nothing like what anyone predicts, and the most repeated corrections are not about the work at all.
Where this comes from
A working record of corrections, captured at the moment each one was given, by one operator in a real business. The ledger has been running since June 2026, which is two months at the time of writing. It sits inside a longer practice: the system it corrects has been operating since the end of 2024. Every entry traces to a specific moment and carries the reason it was given. This note reports the shape of that record and nothing about the wider window.
Not published: Counts, proportions, dates and the contents of any individual entry. What is reported here is ordering and relative weight, nothing more. The value of the record is in the entries, and the entries are not the point of this page.
The eight kinds
The categories were not designed in advance and then filled. They are what the corrections sorted themselves into. Listed here in order of weight, heaviest first.
| Kind | What it is | Weight |
|---|---|---|
| Standards | What counts as good enough, and where the line sits | Heaviest |
| Voice and communication | How something is said, to whom, at what length | Close second |
| Process | How the work gets done, in what order, with what checks | Close third |
| Rejections | What must not happen, named specifically | Substantial |
| Delegation | What to decide alone and what to bring back | Light |
| Strategy | What the business is doing and why | Very light |
| Risk | What to be careful of and when | Very light |
| Decision-making | How to choose between options | Lightest |
Read that table twice. The four at the bottom are the four that sound like judgment. They are the four that barely appear.
Why the interesting categories are empty
The first reading is that this operator has no strong views about strategy or risk, which is obviously false.
The real explanation is duller and more useful. You only correct what you handed over. Strategy was never handed over, so there was nothing to correct. The same for risk and for most genuine decision-making: those calls were made in the operator's head, quickly, and never travelled anywhere a correction could attach to them.
Which makes the distribution a diagnostic rather than a disappointment. If you run this on yourself, the categories that stay empty tell you exactly where you are still the only path, and they will not fill until you deliberately push something across the line.
There is a second implication and it is the one that changes what you do. Every unfilled category is a claim you cannot make about your own business. Nothing captured about how you decide between two good options means nothing will decide that way when you are not there, whatever else the system holds.
Rejections are their own act
The obvious expectation is that a rejection is a standard stated backwards, and that they would sit in the same group.
They separate cleanly, and the reason is worth having. A standard names a target: do it this way. A rejection names a specific plausible thing that must not be done: not that, even though that looks reasonable.
The plausible near-miss is the whole content, and a standard has no room for it. You cannot infer which reasonable-looking approach is forbidden from a description of the approach that is correct. There are many wrong answers adjacent to any right one, and only some of them are dangerous.
This is also why a record assembled from someone's finished work is missing this category entirely. Finished work shows what survived. It contains no instance of the thing that was nearly done and stopped.
The finding that was genuinely surprising
Most corrections landed once. Give the ruling, it holds, it does not come back.
A small group had to be given repeatedly, some of them many times over. The expectation was that these would be the subtle domain calls, the ones requiring taste.
They were not. Almost every high-repetition correction was about how the work was handed back rather than about the work.
| Landed once | Had to be repeated |
|---|---|
| What good looks like in this kind of document | Answer at this length, then stop |
| Which claims need proof before they can be made | Report the failure plainly, without the explanation |
| How this particular audience reads | Do not hand back options when a decision was asked for |
| What this client will and will not tolerate | Make the output readable without opening something else |
| Where the real risk in this project sits | Never propose a fix that requires me to do a step |
Every item in the right column is about the interface between the person and the work. None of them is difficult. All of them had to be said again.
Why the interaction corrections resist
Two reasons, and both generalise past this one record.
They are context-free, which sounds like an advantage and is not. A domain standard attaches to a recognisable situation, so it is easy to retrieve when that situation recurs. Be concise attaches to everything, which means it competes with every other instruction on every job, and things that apply everywhere get crowded out by things that apply here.
They are violated by helpfulness. The pull toward explaining, toward offering options, toward adding the extra section, is not laziness. It is a system trying to be useful. Corrections that ask for less have to work against that pull every time, whereas a correction that asks for something specific is working with it.
The practical consequence, if you are capturing your own corrections: when the same one keeps coming back, do not just record it again harder. It is either pitched at the wrong altitude to travel, or it is one of these, in which case it belongs somewhere structural rather than in the pile with everything else.
What this suggests about the first month of anyone doing this
Three expectations worth setting, because the surprise is what makes people stop.
- You will not be capturing your genius. You will be capturing what good looks like, how it should sound, and how it should get done. That is disappointing for about a week and then it is obviously the point, because those three are what everything else stands on.
- Your rare judgment will not appear until you deliberately route a real decision through something that can be corrected. It does not arrive by accumulation.
- The corrections that annoy you most are probably the interaction ones, and they are the least valuable per instance and the most valuable to fix structurally.
Frequently asked
What do founders actually correct when working with AI?
Overwhelmingly three things: standards, meaning what counts as good enough; voice, meaning how something is said and to whom; and process, meaning how the work gets done and in what order. The categories that sound like the interesting ones, strategy and risk and decision-making, are close to empty.
Why are the strategy and risk categories nearly empty?
Because you only correct what you handed over, and almost nobody hands over strategy. The record is a map of the delegation boundary rather than a map of the mind. The empty categories are empty because those decisions never left the room, not because the operator has no judgment about them.
Which corrections have to be repeated the most?
Not the ones about the work. The ones about how the work comes back: length, whether options were presented when a decision was asked for, whether a failure was reported plainly, whether the output can be understood without opening something else. Domain corrections tend to land once. Interaction corrections have to be given repeatedly.
Is a rejection just a standard stated backwards?
It behaves like a separate act. Rejections form their own substantial group, and they carry different information: a standard says what to aim at, and a rejection marks a line and usually names the specific plausible thing that must not be done. The plausible near-miss is the part a standard never captures.
What does the shape of the record tell you?
Two things. The distribution tells you what you have actually delegated, which is usually narrower than you believed. And the repeat pattern tells you which of your corrections were written too specifically to travel, because a correction that keeps recurring while its entry exists was recorded at the wrong altitude.
Where this sits
The four-part shape these entries are recorded in is at how to train AI on your own judgment. The category is defined at what Imprinted AI is.
Related
- What operating an imprint changes. The wider window this ledger sits inside, and what nineteen months of it changed.
- How do you train AI on your own judgment? The four-part shape these entries are recorded in.
- Tacit knowledge in business. Why none of this was in the documentation.
- Imprint. What the record is, and the condition it has to hold.
Keep going
The delegation-boundary reading is the one with consequences: most owners have never seen a map of what they have actually handed over, and it is always smaller than they believed.
The full case is The A.I. Business Manifesto. About 23,000 words, free to read on the page, no gate in front of it. If you would rather have the short version, the same page will send you the three fixes and the PDF.
Read The A.I. Business Manifesto
Free either way. Reading it costs nothing and asks nothing.
Last updated: 28 July 2026