Why can a cosmetic reword of a chain-of-thought prompt shift accuracy on the same task?

## The uncomfortable property You change "Reason through this carefully before answering" to "Think through this carefully before answering". Nothing about the task changed. Accuracy on your evaluation set moves by three points. This is not a bug you can fix; it is a property of conditioning a generative model on text. The prompt is an input that shapes a distribution, not a contract that pins down behaviour. ## What moves the needle - **Trigger phrasing.** Different reasoning instructions elicit measurably different chain quality on the same task, and which phrasing wins is model-dependent and not predictable from how sensible the phrasing sounds to a human. - **Exemplar order.** With worked examples, the sequence matters — recency and primacy effects are real, and a label pattern across examples can be picked up as a cue in its own right. - **Placement.** Whether the reasoning instruction precedes or follows the question, and whether it sits in a system role or the user turn, changes results. - **Formatting and delimiters.** Markdown headings versus XML-style tags versus bare newlines, list markers, whitespace, casing of field names — all of these carry weight the author did not intend to assign. - **Output-format demands.** Requiring a strict schema alongside free reasoning constrains the chain, and tightening the format can trade reasoning quality for parseability. - **Incidental content.** Examples drawn from one domain can bias reasoning on inputs from another. ## Why this wrecks naive prompt iteration The usual loop — tweak, run once, keep whatever scored higher — is a machine for accumulating noise. On a fifty-item set, a two-item swing is a four-point accuracy change that is entirely ordinary sampling variation. Iterating this way for a week produces a prompt that fits the evaluation set's noise, in a process indistinguishable from overfitting, and the improvement does not transfer to production traffic. It also produces false attributions. If you changed the trigger phrase, reordered exemplars and tightened the output schema in one edit, and the score moved, you have learned nothing about which change mattered — and you will carry all three forward, including the one that hurt. ## The discipline that survives contact with reality **Fix the set before you start.** Frozen, representative of real traffic, large enough that the effect size you care about is distinguishable from noise. A handful of examples cannot support the conclusions people draw from them. **Change one variable at a time.** Phrasing, or ordering, or format — never all three in one commit. **Repeat runs and report spread.** Run each variant several times and look at the distribution, not a single number. If two variants' ranges overlap heavily, you have no result, and the honest conclusion is to keep the simpler prompt. **Version prompts like code.** They belong in the repository with a change history, an owner, and a regression check that runs on change. A prompt edited in a console and pasted into production is an untracked deployment. **Re-baseline on every model change.** Surface-form sensitivities are properties of a specific model. When the model moves, a prompt tuned to the old one's quirks can regress, and the tuning that earned two points may now cost three. Re-running the suite on upgrade is not optional maintenance; it is the only way you find out. **Prefer clarity to incantation.** Edits that make the task genuinely clearer — an explicit output schema, an unambiguous definition of the labels, a stated tie-breaking rule — tend to survive model upgrades. Edits that found a magic phrase tend not to. When two variants score the same, keep the one a new engineer can read. ## The senior framing The interviewer is checking whether you know that prompt tuning is an empirical activity with all the usual statistical traps, not a craft of finding the right words. The strong answer names the specific surface features that move results, then spends most of its time on the measurement discipline that stops you from chasing noise — and admits plainly that some measured differences between reasonable prompt variants are not real.