What the model may not decide for you
Lesson 6 of 6 · Statistics you can defend
1 · Learn the move · Constraints + negative instructions
Some choices in an analysis are computations, and the model can draft and check them. Others are scientific commitments: alpha, the primary endpoint, one-sided versus two-sided, which samples get excluded and by what pre-stated rule, how missing values are handled, what counts as an outlier. Leave those open and the model fills them with defaults that look authoritative and are arbitrary - and a default nobody chose is indefensible in review. The move is a standing negative instruction: draft the analysis plan with every scientific commitment as an explicit DECISION line naming an owner, and forbid silent defaults. The plan becomes a contract - the model drafts within it, checks against it, and never quietly amends it. Where a commitment is missing, the output says DECISION REQUIRED and stops.
Draft an analysis plan for [experiment]. Constraints: every scientific commitment appears as its own line, formatted DECISION: [choice] - OWNER: me, including at minimum alpha, primary endpoint, sidedness, exclusion rules, missing-data handling, and outlier policy. You may propose a value for any DECISION line, but it must be tagged PROPOSED until I confirm it. Never fill a commitment silently. Where a needed decision has no proposal, write DECISION REQUIRED and leave it open. Computations and checks you may draft freely.
2 · Your turn. You write the prompt
Your dose-response analysis is about to run, and you realize nobody ever wrote down what alpha is, what the primary readout is, or what happens to the two wells the intern flagged as bubbled. Before the model touches data, make it draft the plan that pins every one of those - with you as the signatory.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.