Prompt Constraints

Prompt constraints

Constraints are the instructions models drop first, especially when several compete. Making them followed is a matter of specificity, placement, and reducing how many there are.

Method

1. Make constraints checkable. Three bullets of at most fifteen words is followed; be concise is not, because the model cannot tell whether it complied.
2. State the reason where it helps. A constraint with a rationale is followed more consistently, since it lets the model generalise to cases the constraint did not anticipate.
3. Put the most important constraint last. Recency is the strongest position, and it should hold the one that matters most (see prompt-structure).
4. Reduce the count. Beyond a handful, adherence falls across all of them; pick the constraints that matter and drop the rest.
5. Prefer bounds to prohibitions. Answer in under 100 words works better than do not be verbose (see negative-instructions).
6. Enforce hard limits in code. Truncation, validation, and rejection belong outside the model for anything that must not be exceeded.
7. Test the constraint under pressure. Adherence falls with long inputs and complex tasks, so it must be tested at realistic scale rather than on short examples.

Boundaries

Constraints shape output probabilistically and are never guarantees. Over-constraining produces stilted output that satisfies every rule and serves nobody. Conflicting constraints resolve arbitrarily, so consistency between them is the author's job.