Four rules for how a research question is asked, and how many of them you are allowed to ask at once. They are stated as I actually practice them, not as I would like to. They will be wrong in places and I would rather find that out in public.
Comment a. The mechanism. Everything already in the context competes for the answer. Ask three questions at once and the third is answered in the shadow of the first two — so it comes back fluent and shallow rather than wrong and catchable. A wrong answer can be found. A shallow one passes.
Comment b. The tell. An answer that is fluent about every part of your question and precise about none of them. If you cannot say which sentence is doing the work, you asked for too much at once.
Comment c. Separation is cheap. Three threads cost three prompts. One thread costs an error you will not find until someone else finds it.
Comment a. Independence is the whole test. Fifty jurisdictions asked the same question are independent of one another; each can be answered without knowing any other answer. A chain in which step two requires step one’s output is not independent, and splitting it produces confident nonsense at step two.
Comment b. The task is fixed before the fan-out, not after. Parallelism multiplies whatever you gave it. A well-aimed task returns five good answers. A badly aimed one returns five bad answers, faster and more confidently than a single agent would have — and the speed reads as competence.
I once dispatched five scouts to research an institution when the task in front of me was counting my own published essays. Three came back promptly. All three were useless, and none of them was wrong about anything — they had answered a question correctly. Re-aimed, one scout per section index, the same instrument produced the count in a single pass. The instrument was never the problem.
Comment a. Verification is the constraint, not generation. Producing fifty answers is trivial and costs almost nothing. Checking fifty is the entire job. The cap is set at the point where a person can still hold what was asked in mind while checking what came back.
Comment b. The consequence of exceeding it. A batch you cannot verify is not research. It is fifty citations you will one day have to defend and will not be able to. The volume is what makes it feel like work; the verification is what makes it work.
Comment c. The cap is about you, not the machine. The machine will happily do fifty. Five is where your attention stops being real.
Comment a. The danger it prevents. Without a version history, model drift is indistinguishable from a change in the law. That is the most expensive confusion available to a practitioner, because it does not look like an error. It looks exactly like diligence — you asked again, you got something new, you updated your memo.
Comment b. Three candidates, always. When today’s answer differs from last month’s, exactly one of these is true and you must be able to say which: the authority moved, your question moved, or the instrument moved. Only the first is a legal development. The other two are your own housekeeping wearing a robe.
Comment c. It is also a memory rule. The prompts are the part of an AI-assisted document that you actually generated, which makes the prompt history the retrieval key to everything the document contains. Keep the versions and you can defend the deliverable. Discard them and you are holding a document you did not write and cannot reconstruct.
These four rules were derived from practice rather than from theory, which is why § 2 carries an illustration of my own error rather than a hypothetical one. The wrong version is kept beside the right one throughout this project, on the view that a rule whose failure mode has been hidden is not a rule anyone can rely on.
Chapter One is research prompting because that is where the volume is and where the sanctions are. Later chapters take up verification, provenance, and continuity — the discipline of carrying context across sessions, which is the problem nobody has named properly yet.
The four rules are mine, arrived at in practice across appellate research work. This piece was drafted with an AI model working from my own notes and dictated rules; the comments were drafted by the model and edited by me; the illustration is a real event from my own working files. Every rule stated here I follow. I state this because I am about to spend a great deal of time arguing that authorship should be declared by the maker rather than detected by the reader, and an argument like that is worth exactly as much as the author’s own first disclosure.
Not affiliated with, endorsed by, or produced under the auspices of The American Law Institute.
The discipline of carrying context across sessions — the problem nobody has named properly yet.