The companion piece asked when to let the model draft. This one asks the harder question: how do you know when it is time to close the laptop. Not out of purity, out of practice. The loop only teaches you something if you stay in it.
Under an active-inference reading, learning happens where a prediction meets an observation and updates the model that made the prediction (Class E, Parr, Pezzulo, and Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, 2022). Reach for a draft-completer before you have made your own prediction, and there is nothing for the update to land on. You have not shortened the loop. You have skipped it. The signals below are the ones we watch for in workshops to catch that skip early. If you notice one, put the model down. If you notice two in the same session, close the laptop and take a walk.
Six signals the loop has collapsed
One. You cannot say the sentence out loud without reading it. If the paragraph on your screen is one you could not deliver from memory to a colleague, you did not write it. You approved it. Approval is a fine act; it is not learning, and it is not a claim you can defend on Monday.
Two. You feel relief when the output arrives, not curiosity. Relief is the feeling of avoided work. Curiosity is the feeling of a prediction meeting a surprise. If the tool has become the thing that keeps you from feeling stuck, the stuck is where the update was going to happen.
Three. You are reprompting instead of rethinking. Three reprompts in a row on the same task is a tell. You are treating the tool as a slot machine for a phrasing you have not yet earned. Stop reprompting. Write one sentence in your own hand about what you actually want to say. Then decide whether the tool still has a role.
Four. Your log has gone quiet. The tool-use log is the small artifact each participant keeps (Class C, the log is the integration point between your practice and any later review). A quiet log during a busy day means the loop closed without you noticing. When the log is empty and the deliverables are full, the model is doing the thinking and you are moving the mouse.
Five. You are defending the output before you have understood it. If a colleague pushes back on a paragraph and your first move is to explain what the tool meant, you have already handed the seat over. The right move is to say, honestly, that you have not yet made the sentence your own, and to rewrite it in front of them.
Six. You cannot remember what you learned this week. A cohort that ships drafts all week can look extremely productive on Friday and be measurably less capable on Monday than a month ago. We have seen enough of this pattern in early cohorts to name it (Class A, empirical-in-session), and the cognitive-offloading literature is worth reading on its own terms (Class U, we have not yet run our own controlled study).
The exit ritual
Putting the model down is a specific act, not a mood. Three steps, in order.
First, name what you were about to skip. One sentence in the log: "I was going to let the tool answer instead of thinking about X." No self-punishment. Just the observation.
Second, do the smaller version of the thinking by hand. Not the whole task, the smallest version of the reasoning step you were avoiding. Two sentences on paper. A three-line outline. A napkin sketch. The point is to give your model a prediction to make.
Third, decide whether the tool still has a role in this task. Sometimes yes, with a narrower prompt and a clearer job. Sometimes no, and the honest move is to finish the task without it. Either answer is fine. The unacceptable answer is drifting back into reprompting because the smaller version was harder than expected.
The falsifier
Falsifier (Class F). If workshop participants who apply this ritual report, in their own words at exit and again at ninety days, that the ritual did not change how often the loop collapsed, we redesign it. Not tune, not reframe. Redesign. The claim that a small named practice can keep the human in the loop earns its standing only by surviving that question in the field.
Where to go next
- When to let the model draft: the paired piece, the cases where the tool earns its seat.
- AI as a partner, not a replacement: the pillar this practice sits under.
- The IamHITL workshop: the practice held in a room with other people trying it at the same time.
Evidence classes cited in this post: A (empirical-in-session, cohorts to date), C (the tool-use log as the practice integration point), E (Parr, Pezzulo, and Friston, 2022), F (falsifier stated), U (cognitive-offloading study not yet run in-house). Corrections welcome. Bring the counter-evidence and the record updates.