Why I do not ride an e-bike or my use of Claude.ai

Written By Christoph Hinske

A meta-reflection on my research practice with AI

Context note: In an earlier note, the editor put the writer on the hot seat. This time the writer looks at the tool. This post is the second boundary test in the PhD Notes: a reflection on my own research practice with AI. It reflects personal practice, not any organizational context; the only organization named is the vendor of a tool I subscribe to privately. The experience is mine; the drafting was done with Claude, the same tool whose limit this note examines. How I work with AI is described in my AI disclosure.

The signal

It keeps ending the same way. I am deep in a framework with the model, tracing how a field's literature connects and where it contradicts itself, at a pace I could not sustain on my own. Then the meter runs out. The conversation is replaced by a notice: usage limit reached. The window reopens in a few hours.

My first reaction has never matured. It is an annoyance every time.

The lockout usually catches me while checking. Not while building, and rarely while polishing. Checking: one more source against the frame, one more contradiction to chase, one more test the model is happy to run. Sometimes it catches me mid-flow instead, which feels worse.

The signal worth reporting is what I noticed over months of this: the work that follows a lockout is regularly better than the work that precedes it. Which leads to a sentence I did not expect to write. The usage limit on Claude's Pro plan, twenty-two euros a month, may be the most effective feature in my research practice. Nobody designed it for that.

Two speeds

A sprint with the model compresses conceptual work. With the effort set to maximum, blocks that used to take an afternoon now take minutes. The model surfaces more literature than I can absorb at dialogue speed. Leads accumulate. Distinctions accumulate. The frame grows.

And checking has no natural end. With a partner that always answers, there is always one more verification available, and it is always cheap. Herbert Simon made the underlying point long ago: search without a stopping rule does not stop; bounded actors need a threshold, or a clock (Simon, 1955). I did not set a clock. The plan set one for me: the session budget resets every five hours, and intensity empties it early.

What the pause contains

The lockout does not produce rest. It produces a mode change. I read the articles the sprints washed up, the ones I had been queueing behind one more check. I write memos by hand. I hold the frame against years of project experience rather than the next model answer. Then the window reopens, and the tempo picks up again.

So the pause is not the opposite of the work. It is the work at a different speed. There are two clocks in this practice: the machine speed of discovery and the human speed of absorption. The lockout is where the second catches up with the first.

The footprints

Following my own rule from the earlier meta-reflection, footprints over opinions. Three residues are checkable in my files.

First, reading lists that originate inside sessions and get read outside them. Second, pause memos, written while locked out. Third, and most telling: the concept is different when I return. The difference has a shape. Most often, something falls away or gets simpler. Second, most often, the structure gets rebuilt. New elements come only third. Renaming comes last.

One example I can place: agreement footprints, a concept I now use daily, settled during one of these pauses. The sprints had circled it, with the model as reflection partner. The settling happened away from the machine: some reading, a slow pass through years of project memory, and then a judgment. The simple image is the right one. Agreements leave traces. Mostly, the traces are not visible, and finding them is detective work. That ratification did not happen at dialogue speed. It happened at reading speed.

A working explanation, held loosely

A signal is not a diagnosis, and one practice is not a law. But the pattern lines up with what is known.

Additions come easily; subtractions are overlooked. Adams, Converse, Hales, and Klotz (2021) showed that people systematically miss subtractive changes and that subtractive ideas take more cognitive effort to reach. A sprint is high-load by construction. The pause is the only part of my week where the load drops far enough for subtraction to become available. That would explain why simpler is the most common thing I bring back.

Constraints can also serve learning directly. Bjork and Bjork (2011) call them desirable difficulties: conditions that slow performance while deepening retention and transfer. The friction I keep paying for behaves like one.

The classic incubation effect probably contributes at the margin. The meta-analytic evidence indicates that incubation helps, and that it helps most when the interval is filled with light activity rather than demanding work (Sio & Ormerod, 2009). My lockouts are filled with light reading. Still, I hold this strand loosely; the reading and the memos feel like the main event, not the resting.

The counterproof

The objection writes itself: I am rationalizing a paywall. Sweet lemons, praising a constraint I cannot remove.

Except I removed it. For a stretch, I paid for one of the larger plans. The lockouts thinned out, then stopped mattering. And the work degraded in a way that took weeks to see. Less fell away. The frames grew. The reading trail went unread. It hindered good work and learning. I went back to the small plan.

My own discipline has a name for that: revealed preference. What I pay for says more than what I praise.

Schelling (1978) described savers who accepted lower interest so that the bank would guard their Christmas money against their own withdrawals. I recognize the transaction. I accept lower capacity so that a counterparty guards my attention against my own appetite. Self-command usually fails at exactly the instance that can revoke it; binding holds better when it is exogenous, carried by an external arrangement rather than by resolve (Elster, 1984). What keeps my limit in place is not willpower. It is a one-tier decision a month instead of a contest in every session, with enforcement contracted out. The most credible form of self-binding is an agreement with a counterparty.

I notice this is the claim my research keeps making about coordination in general. Agreements shape behavior in ways intentions do not. And they leave footprints whether anyone designed them or not.

What this opens

I hold all of this as a finding about one practice, mine, under conditions I have described. It is not a recommendation, and it does not generalize on its own; a limit that gives one researcher a subtraction phase may simply block someone else's deadline.

The question it opens sits one level up. Organizations are currently engineering the opposite condition: frictionless, unlimited, always-on access for everyone. If a usage limit functioned in one research practice as a learning architecture, what does the removal of all friction do to absorption, to subtraction, to judgment, at the level of a team or an organization? I do not know. It is the next thing I want to watch.

A cyclist I once saw in the mountains wore a shirt. The back asked: "Why don't I ride an e-bike?”

The front said: “Because I can.”

References

Christoph Hinske
Previous

How Credit Instruments Automate Their Own Failure
Next

The Replication Question