Glossary
Articulation gap
The distance between what you meant and what you managed to hand over. It sounds like something you would have to take on faith. You can count it.
Articulation gap (noun). The share of the decisions a piece of work requires that the system had to resolve by guessing, because neither your instruction nor anything it had captured from you supplied an answer.
The gap is not a weakness in how you communicate. It is a structural fact about handing work to anything that cannot read your mind, and it exists for every operator, including the ones who write very well. What varies is its size, and whether anyone has ever looked at it.
The instrument
Twenty minutes, one real job, no tooling. The trick is to interrupt the work before it happens, because once the output exists it is too late to see what went into it.
- Take an instruction you would genuinely give. Not a test prompt. Something from this week, at the length you would actually have typed it.
- Tell it not to do the work. This is the whole move. Instead, have it write out the full specification it would have worked from, expanded to every decision it would have had to make along the way.
- Have it tag every one of those decisions with where the answer came from: from your instruction, from what it already knows about you, or guessed.
- Count the third pile.
Everything in it was decided on your behalf, without you.
Most people are surprised twice. First by how many decisions a job they thought was simple actually contains. Then by how few of them their instruction covered, and how few came from anything the system knew about them.
Why the output cannot tell you this
The obvious objection is that you already have a way to assess this: look at whether the work came out well.
That signal does not carry the information you need, and the reason is worth being precise about. A guess that happened to be right and a guess that happened to be wrong look identical in a finished piece. Both are just text. So a good result is compatible with the system having invented most of the decisions behind it, and you have no way to tell the two situations apart by reading.
Worse, the reading is done by you, and you close gaps automatically while you read. You look at the output, silently supply the context that makes it make sense, and rate it fine. That works right up until the moment a job runs without you, and then every gap you were quietly covering is covered by an invention instead.
Which is why the count has to be taken before execution. Afterwards there is nothing left to count.
Reading the three piles
| Pile | What it means | What to do with it |
|---|---|---|
| Guessed | Decided for you, silently. The gap itself | Each item is a call you could make once and have kept |
| From your instruction | You paid for this in words, this time and every time | Anything recurring here is a capture candidate |
| From what it knows about you | Already carried. This is the part working | Nothing. This is what you are trying to grow |
The second pile is the one people skip, and it is half the diagnosis. Every decision you spelled out that the system should have been able to carry on its own is a recurring tax you are paying in words. It does not show up as a problem because the work comes out right. It shows up as instructions that never get shorter.
The two ways to close it, and only one of them stays closed
Write more. Every decision you specify is one the system no longer invents, so the gap genuinely narrows. It also costs you the time you brought the system in to save, and it narrows nothing for the next job, because the next instruction starts from the same place. This is the expensive close and it has to be paid again every time.
Capture the calls. Take items out of the guessed pile, make the decision once, and have the decision kept with the reason. Now the system resolves those from what it knows rather than from invention, and the next instruction can be shorter without the guessing going up. This is the close that holds.
Most operators only have the first move available, so they conclude the gap is a permanent cost of working this way. It is a permanent cost of working this way with nothing captured, which is a different claim.
What the number is for
Not a grade. A work list.
The value of running this is not the score, it is the guessed pile, which is a list of decisions your business is currently making without your judgment, enumerated, on one page. That list did not exist before you ran it and it is not visible from any other angle.
Rerun it on the same kind of work a month later and the useful signal is the shape of the change: fewer guesses at the same instruction length, or the same result from a much shorter instruction. Either one means the gap moved. Both at once means it is closing the way it is supposed to.
Frequently asked
What is an articulation gap?
The distance between what you meant and what you actually managed to hand over. It is not a communication weakness. It is a structural fact about handing work to anything that cannot read your mind, and it becomes visible the moment you count the decisions the system had to invent for itself.
How do I measure mine?
Give the system a real instruction and tell it not to execute. Instead have it write out the full specification it would have worked from, and tag every execution decision as coming from your instruction, from what it already knows about you, or from a guess. Then count the guesses. That count is the gap, and it takes about twenty minutes.
Isn't a guess sometimes right?
Often, and that is what makes it dangerous rather than harmless. A correct guess and an incorrect one are indistinguishable in the output, so a good result tells you nothing about how much of it was invented. The count is the only signal, and you only get the count by asking for it before the work runs.
Does a longer instruction close the gap?
It closes part of it and charges you for the rest. Everything you spell out is a decision the system no longer has to invent, which is real, and you have also just spent the time the system was supposed to save. Length is the expensive way to close a gap. Capture is the way that stays closed.
What should I do with what the tags show?
Read the guessed pile as a work list. Each item is a decision that was made on your behalf without your judgment, and each one is capturable. The over-specified pile is a second list: those are decisions you spelled out that the system should have been able to carry itself, which means they are candidates for capture too.
Where this sits
The asset you are building as the gap closes is articulation density. What the calls become once captured is at Imprint. The category is defined at What is Imprinted AI.
Related terms
- Articulation density. The asset you are building when you close the gap.
- Mirror Score. The other measurement in this glossary. Person-level rather than task-level.
- The Final Twenty. What the gap costs you per job while it stays open.
- Imprinted AI vs a Custom GPT. Why closing it by writing more instructions has a ceiling.
Where the term comes from
This is one entry in the vocabulary of a longer argument. The full glossary has fifteen terms. The case they belong to runs about 23,000 words, it is free, and there is no email gate on it.
Read The A.I. Business Manifesto
Nothing on this page is for sale. Quote it, argue with it, or pass it on.
Last updated: 28 July 2026