Question
How do you train AI on your own judgment?
By intercepting the corrections you are already making, in four parts: the case, the call, the reason, and the boundary. Almost everyone writes the first three and skips the fourth, which is the one that makes it usable.
First, the approach that does not work
The natural plan is to sit down and write your judgment out. Block a Saturday, open a document, get it all down.
It fails, and not because of discipline. It fails because a standard does not exist in your head as a sentence waiting to be transcribed. It exists as a reaction. You do not walk around holding the rule never let a promise get ahead of the proof. You hold nothing, until you read a paragraph that does it, and then something in you objects. The rule becomes visible at the moment of violation and not before.
So the Saturday produces the part you could already say, which is the part that was already transferable, which is the part that was never the problem.
The move: intercept, do not schedule
You are already producing the raw material, every day, and throwing it out. Every time you take something the machine produced and fix it, you have just applied a standard. The fixing already happened. The only thing missing is that nothing catches it.
So the method is not to add work. It is to put something under the work that is already happening.
Practically: when you correct something and feel that flicker of no, not like that, stop for two sentences and write down what just happened. Not later. Later, the irritation is gone and with it the specificity.
The four parts
A capture that will still be useful in six months has four parts. Three of them are obvious. The fourth is where this succeeds or fails.
| Part | What it holds | What breaks without it |
|---|---|---|
| The case | What the situation actually was | The ruling floats free and gets applied to situations it never governed |
| The call | What you decided, plainly | There is no ruling, only a mood |
| The reason | Why. The part that generalises | You have recorded a fact about one Tuesday instead of a standard |
| The boundary | When this stops applying | The rule is applied everywhere, including where it is wrong, and starts colliding with your other rules |
Why the boundary is the whole game
Here is what happens to people who capture the first three parts diligently and skip the fourth.
For the first few weeks it works well. The rules are few and they are all correct. Then the collection grows, and the rules start meeting each other. Always lead with the strongest proof meets never open on a number the reader has no reason to trust yet. Both are yours. Both are right. In a given case, one of them has to yield, and nothing in either entry says which.
Now the system either picks arbitrarily, which is worse than having no rule, or it asks you, which puts you back in the loop you were trying to leave.
The boundary is what resolves this in advance. It is one clause, usually beginning with unless or except when, and it is the difference between a preference and something that can operate without you standing next to it.
Worked example, in the shape above:
Case. Sales page draft, cold traffic, claim about results appearing in the second paragraph.
Call. Moved the claim down, behind the mechanism.
Reason. A promise made before the reader has any reason to believe it does not read as ambitious, it reads as suspect, and the rest of the page then has to fight the suspicion.
Boundary. Does not apply to warm traffic who arrived because of that specific result, where leading with it is confirmation rather than a claim.
That entry can be handed to something else and applied correctly, including in a case that has not happened yet. Strip the last line and it cannot.
What to expect over the first months
Two things, and the second is the one to watch for.
Early on you will capture a lot, and much of it will be things you are slightly embarrassed to find you had never said out loud. That is normal and it is the point.
Then the rate falls, and the proportion of repeats climbs. You start encountering standards you have already recorded rather than new ones. That looks like the system running dry. It is the opposite. A falling discovery rate with a rising repeat rate is what a filling collection looks like from the inside, and it is the first real signal that the thing is working.
The failure signal is different: three months in, you are still making the same correction, word for word, and the entry exists. That means the reason was written too narrowly to travel. Rewrite the reason at the level where it would have caught this case too.
Frequently asked
How do you train AI on your own judgment?
Not by setting aside time to write it out, which fails because a standard is invisible to you until something violates it. You intercept the corrections you are already making. Each one gets recorded in four parts: the case, the call, the reason, and the boundary. The boundary is the part nearly everyone omits, and it is what makes the ruling safe to apply to a different case later.
What exactly is the boundary?
The conditions under which the ruling stops applying. A rule with no boundary is applied everywhere, including where it is wrong, and it will collide with your other rules. Writing down when it does not hold is what turns a preference into something a system can use without supervision.
How long does this take?
The capture itself takes a sentence or two at the moment you are already correcting something, so the marginal cost is close to zero. What it is not is a weekend project. Anything that requires you to block out time to introspect will be abandoned, and the material it produces is the least valuable material you have, because it is the part you could already articulate.
Do I need special software?
No. A text file works. What matters is that the four parts are present, that entries are dated, and that nothing inferred gets written in alongside things you actually decided. Tooling helps at volume by making retrieval automatic, but the discipline is what produces the value, not the tool.
How do I know it is working?
The corrections change. Early on you correct the same kind of thing repeatedly. Later the system clears the things you used to fix and you find yourself correcting further along, on finer points. If you are still making the identical correction three months in, the reason was recorded too narrowly to travel.
Where this sits
What the resulting body of rulings is called, and the standard every entry has to meet, are at Imprint and captured, not guessed. The category is defined at what Imprinted AI is.
Related
- What founders actually correct. What the corrections turn out to be, from an actual bank of them.
- How is an imprint different from a prompt library? Why the file you are building is not a prompt library.
- Captured, not guessed. The standard every entry has to meet.
- What should a founder automate first? Where to point this once you have some.
Keep going
The four-part shape is the small end of this. The larger argument is about what a business is worth once the calls behind its best work are held somewhere other than one person's head.
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