Most review prompts produce a wall of preferences. Constraining output to defects is what makes the result actionable — and telling the model it may say…

Asking for the contract before the code forces the model to read the function instead of pattern-matching it. The 'riskiest line' question surfaces exactly where a…

The instruction to stop and ask one question is the important part: a wrong reproduction is worse than no fix, because it launders a guess into…

Explaining before editing is the cheapest way to avoid a confident rewrite of code nobody understood. The 'disagrees with its own names' item catches the traps.

Item 4 is the useful one: it tells you how to sanity-check the number before reporting it upward, which is where most analytics mistakes actually surface.

Restricting to 'expensive to change after launch' filters out the style opinions that make API reviews exhausting and produce no change.

The 'say so and stop' escape hatch catches the cases where migration is impossible incrementally — which is exactly when a confident plan is most dangerous.

Writing tests from the implementation reproduces the implementation's blind spots. Deriving them from the spec is the only way to catch the requirement that was never…

Scoping to security prevents the model from padding with general advice, which is the usual failure mode of 'review this for security'.

Asking for the earliest non-consequential signature is the whole prompt: it prevents the model from confidently naming the loudest error, which is usually a downstream symptom.