Most advice here is some version of "use AI less". It does not work, for the same reason most abstinence advice does not work: it asks you to give up something valuable in exchange for a benefit you cannot observe, and it collapses the first busy week.
Here is a framework that survives contact with real work.
The rule that actually matters
Delegate execution. Keep judgement.
That is the whole thing. Everything below is elaboration.
The research distinguishes beneficial from detrimental offloading. Beneficial offloading frees attention for higher-order work — you hand off something that was never developing you, and spend the recovered capacity on something that is. Detrimental offloading bypasses the thinking itself.
The trap is that both feel identical while you are doing them, and both produce output that looks fine. Fluent output is not a signal of anything except fluency.
Four questions before you prompt
Ask these of any delegation you are about to make. They take about five seconds.
1. Would doing this myself have taught me anything?
If no, delegate freely. Formatting a table, converting units, recalling a syntax you look up every time anyway — there is no encoding to lose. If yes, that is your answer.
2. Am I asking it to do the work, or to decide?
"Draft three options" is execution. "Which of these should I pick" is judgement. The second is where capability goes, because deciding is the skill and you have just outsourced the rep.
3. Do I have a position before I ask?
Prompting before you have formed a view means the first output becomes your anchor. You will edit it, feel like you contributed, and never notice you did not start from your own thinking. Write two sentences of your own position first — badly is fine — then prompt.
4. Could I defend this in a room without the tool?
The MIT participants could not quote essays they had submitted minutes earlier. If you could not reconstruct the reasoning under questioning, you have output you do not own.
Reserve some domains entirely
Pick two or three things you are unwilling to lose and do not delegate them at all. Not "less often" — not at all, so there is no threshold to negotiate on a bad day.
The reserved domains should be things where your judgement is the actual value: the thing you are known for, the thing you would want to still be able to do in five years, the thing you would be embarrassed to be caught outsourcing.
Everything outside that list is fair game. That is the point — a rule you can hold is worth more than a rule you break by Wednesday. This is exactly what IMPRINT's Skill Vault is for: naming the capabilities you are protecting, because an unnamed skill cannot be observed slipping.
Keep the difficulty that develops you
The learning-science term is desirable difficulty: conditions that make a task harder in the moment produce more durable learning than conditions that make it easy.
This is counterintuitive and well established. Struggling to retrieve something strengthens the memory more than re-reading it. Being wrong before being corrected beats being told first.
Applied to AI, it gives you a practical test. When you delegate, ask whether you are removing friction or difficulty. Friction is overhead that teaches nothing — remove it all. Difficulty is the resistance that builds the capability. Remove that and you get the output without the development.
The blank page is difficulty. Formatting citations is friction.
Practise unassisted, on a schedule
Skill-retention research puts measurable decline in an unused capability at roughly 60 to 90 days. That is the actual clock you are working against, and it is slower than the panic suggests — but it only runs in one direction if nothing interrupts it.
The interruption does not need to be large. Some deliberate unassisted work at regular intervals is enough to keep a capability live. What matters is that it is scheduled rather than aspirational, because unassisted work is by definition the harder option and will lose every unscheduled contest against the easier one.
Measure, or you are guessing
Here is the uncomfortable part. Everything above is unfalsifiable without a comparison point.
You cannot tell whether your thinking has changed by consulting your memory of how you used to think, because memory of your own past capability is one of the things that degrades. It is the same reason you cannot judge your own writing from a year ago accurately — you have no access to the version of you that wrote it.
So capture a fixed sample of unassisted work now, while you still recognise it, and compare against it later. That is the entire idea behind a cognitive baseline, and it is the one part of this you cannot improvise after the fact.
You can do that with a text file and a calendar reminder. IMPRINT does it more systematically and produces a Drift Score from the comparison — and publishes exactly how, including what the measurement gets wrong, because a number you cannot inspect is not worth acting on.
What not to bother with
- AI blockers. Trivially circumvented, and a capability protected only by a blocked website was not being protected.
- Usage quotas. They measure the wrong variable. Heavy deliberate use beats light careless use every time.
- Total abstinence. Rarely sustained, and it leaves you with no practice at the skill that actually matters now — delegating well.
The goal is not less AI. It is knowing which part of the work is yours, and still being able to do it.