Comparison
Imprinted AI vs a Custom GPT
A Custom GPT is you, describing yourself, in a text box, before any work has happened. That sentence contains every limit this page is about.
A Custom GPT is configured. You write instructions describing how it should behave, optionally attach files, and it applies that description to every conversation.
An Imprinted AI is accumulated. It is built from rulings you made on real work, one at a time, and it holds material you never wrote down and could not have.
That was already the easy part to hand off.
The ceiling is set by introspection, not by effort
Writing your instruction block is an act of self-report. You sit down, think about how you work, and put it into words.
Whatever comes out is the sayable portion of you.
Now notice what the sayable portion is. It is the part you could already explain to a new hire on their first afternoon. Formatting preferences. Tone. Things you always do and things you never do. Genuinely useful, and it is useful precisely because it was already transferable by ordinary means.
The reason you are hard to replace is the other portion. The calls where you cannot fully account for the decision even to yourself, where the reason arrives after the ruling rather than before it. Michael Polanyi named this in 1966 in a sentence nobody has improved on: we can know more than we can tell.
So a better-written instruction block does not raise the ceiling. It fills the same room more neatly. The material that would raise it is not accessible to introspection at all, which is why a questionnaire cannot reach it and why the only route to it runs through decisions you actually made.
There is a second-order version of this that costs more. What you write about yourself is not merely incomplete. It is a theory. People are consistently wrong about their own decision rules, and confidently so. A document about how you think is a document about how you think you think, and the system then applies your theory of yourself rather than you.
It is written before the work, about situations you imagined
An instruction block is authored in the abstract. You picture the kinds of things that will come up and write rules for them.
Imagined cases are where people write rules they do not actually follow. It is not dishonesty. It is that in the abstract you reach for the principle, and in the moment you apply the exception, and the exception is where your judgment lives.
Capture never leaves the moment. The ruling gets recorded against real output with real stakes, after you have seen the specific thing in front of you. That is the difference between what you believe you would do and what you did.
Instructions compete, so the block degrades before it fills
Everyone runs into the capacity limit eventually. The more interesting problem shows up well before it.
Instructions are not additive. Each one you add competes with the ones already there for the model's attention. Past a certain size, some rules get followed and others get dropped, and nothing tells you which. You notice a behaviour broke, you add a rule to fix it, and something you fixed two months ago quietly stops working.
That is a bad growth curve for anything meant to accumulate. A year of judgment cannot be written into a text box, not because the box is too small, but because the mechanism gets less reliable as it fills.
A retrieved ruling behaves differently. Only the rulings that govern the case at hand are brought forward, so a body of ten thousand of them puts no more pressure on the system than ten. Growth improves it instead of crowding it.
Nothing improves unless you go and edit it
When a Custom GPT gets something wrong, here is the repair path. You notice. You work out which instruction caused it. You open settings. You rewrite the text.
Every step depends on you having the patience for it that day. Which means the system's picture of you is frozen at the last time you were willing to sit and edit a box, and the hundred corrections you made in conversation since then are gone. They lived in chat windows and died with them.
This is the reason the same fix gets made forever. The correction happened. It just had nowhere to go.
Side by side
| Custom GPT / instruction block | Imprint layer | |
|---|---|---|
| Source | Your description of yourself | Decisions you made |
| Written | Before the work, in the abstract | During the work, on real stakes |
| Reaches | What you can articulate | What you could not articulate |
| Growth | Degrades as it fills. Rules compete | Improves as it fills. Only relevant rulings surface |
| Corrections | Manual edit, when you get around to it | Captured at the moment you make them |
| Copyable | Yes. Whoever sees it has it | No. It is your decisions or it is empty |
| Best at | A defined task with a defined output | Undefined work that needs your standard |
When a Custom GPT is exactly right
When the job is shaped. A specific task, a known output, a repeatable process. Turn transcripts into summaries in this format. Draft replies in this tone. Run this checklist against this kind of document.
For shaped work, configuration is fast, cheap and sufficient, and reaching for anything heavier is wasted motion.
The limit shows up when the work is unshaped. When the right answer depends on the particular case and you would have decided it differently than the reasonable default. No description of yourself covers that, because if you could have described the rule in advance, it was not judgment. It was a policy, and you should absolutely put a policy in the box.
The check that separates them
Read your instruction block and ask how much of it you could have handed to a capable new hire on day one.
If the answer is most of it, that is not a criticism of what you wrote. It is the ceiling doing what it does. Everything in there was sayable, and sayable knowledge was never the part that made you difficult to replace.
Frequently asked
Is a Custom GPT the same as an Imprinted AI?
No. A Custom GPT is configured from a description you write about how you want it to behave. An Imprinted AI is built from decisions you actually made, captured against real work. The first is limited to what you can articulate about yourself in advance. The second reaches the part you could not have articulated, which is where most of the value in expert judgment sits.
Why can't I just write really good custom instructions?
You can, and you should, and there is a ceiling on it that better writing does not raise. Introspection returns the part of your thinking you can put into words. That part is, by definition, the part that was already transferable. The reason a person is hard to replace is the portion they cannot state, and no amount of effort spent writing about yourself converts that portion into text.
What happens as I add more instructions?
Quality falls before capacity does. Instructions compete: as the block grows, later rules dilute earlier ones, some get followed and others quietly get dropped, and there is no signal telling you which. This is why long instruction blocks often behave worse than short ones, and why adding a rule to fix a problem sometimes breaks two things you had already fixed.
Can someone copy my Custom GPT?
Anyone who can see the instruction block has everything in it, and instruction blocks leak routinely. That is worth thinking about, because it means an instruction block is not an asset in the ordinary sense. Captured judgment cannot be obtained the same way, since its value is in a body of decisions that only exist because a particular person made them.
Should I stop using Custom GPTs?
No. They are the right tool for a defined job with a defined output, and configuring one is often the fastest useful thing you can do. The mistake is expecting a configuration file to carry judgment. Configuration sets the task. Capture supplies the standard applied inside it, and the two work together rather than replacing one another.
Where this sits
Read the category definition at What is Imprinted AI. The capture mechanism is public: Imprint is local-first and preserves provenance on every entry.
Related comparisons
- Imprint vs an SOP. Describing the work rather than describing yourself.
- Imprinted AI vs an AI clone. A machine writing the self-description instead of you.
- Imprinted AI vs an AI agent. Configuration against authority, once the system starts acting alone.
The full argument
This page is one distinction pulled out of a longer case. The whole thing 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. If the distinction is useful, the useful thing to do with it is pass it on.
Last updated: 28 July 2026