“Write me a marketing email” has no wrong answer, which sounds generous until you realize that's exactly the problem. No length, no tone, no audience, no scope. The AI has to guess at all four, and it will default to the safest, most average version of each: medium length, neutral tone, generic audience. Add one line, “under 100 words, casual tone, for existing customers, don't mention the price increase,” and the guessing stops. Same request, one sentence of boundaries, a completely different range of possible answers.
That's what a norm actually is: not a restriction bolted onto a prompt, but the thing that narrows an impossibly wide range of acceptable answers down to the one you actually want.
What a constraint is doing for you
Anthropic's own prompt engineering guidance frames it plainly: constraints tell the AI what to leave out, not just what to include. Without one, the model can drift off-topic, pad its answer with filler and disclaimers, or default to generic content because nothing told it not to.
A widely cited study on verbosity bias backs the length point specifically: the ratings models get trained on, human and AI alike, systematically favor longer answers even when a shorter one says the same thing, which is exactly why a model defaults to verbose unless told otherwise. A plain word limit forces it to prioritize instead. Two or three well-chosen constraints beat five competing ones, stack too many pulling in different directions and the model can't fully satisfy any of them.
The three constraints that do the most work are the same three almost every unconstrained prompt is missing: length (how much is too much), tone (formal, casual, technical), and scope (what's in bounds and what isn't). Leave all three open and you haven't given the AI freedom, you've given it three separate guesses to get right at once.
Boundaries make the AI freer, not more limited
This feels backwards until you've seen it work. A writer given a blank page and told “write anything” tends to freeze or default to cliché. A writer given a specific brief, a word count, a required rhyme scheme, an audience, produces something sharper, because the constraint gives the effort somewhere to go. The same shows up in prompting: a boundary doesn't shrink the AI's options, it removes the options you never wanted in the first place, leaving the model free to spend its effort on the part of the answer that actually matters to you.
The same logic, higher stakes
The principle scales up further than a single prompt, just through a different mechanism. Once AI moves from a one-off chat answer into an autonomous agent making decisions in production, unconstrained instructions stop being merely generic and start being genuinely risky. Forbes Tech Council draws a sharp line here: a prompt is an instruction the model can choose to follow, a guardrail is code that enforces the rule regardless of what the model decides. Telling it, in a prompt, never to auto-approve expenses above a set threshold is a suggestion; the same rule enforced in code is a fact. Dickson Lukose and QED42 both cover what that looks like in practice for production systems: safety limits, task-specific boundaries, and honestly acknowledging what a system doesn't know rather than guessing past its limits.
Worth being precise about the difference: none of that is what Norm does inside CORDON. Those three sources are about enterprise agents operating with real consequences, where a rule has to be enforced in code, not just typed into a prompt. The underlying logic still transfers, though: a system with no stated limits, whether it's a single chat answer or an autonomous agent, defaults to guessing at boundaries nobody bothered to set. A prompt is lower stakes than an agent in production, but the same gap is doing the damage either way.
Norm closes the cycle
Context sets the situation. Role sets the voice. Norm sets the edges: what to avoid, how long, how formal, what's out of scope. Norm is the field people most often leave blank, not because it doesn't apply, but because “no constraints” sounds like giving the AI more freedom. It's the opposite: a blank Norm isn't the same as a Norm that genuinely doesn't apply, it's usually just a boundary nobody got around to setting. What role should you give AI? 10 examples. covered the Role piece, this one closes the cycle. Why you still have to decide every time. covers the other side of this: CORDON's own four-of-six minimum means a field with nothing real behind it stays blank rather than forced, Norm usually isn't one of those fields though, it's one of the cheapest, sharpest ones to answer, and the prompt that leaves its edges open is the one most likely to come back wrong.
Try CORDON and see how much a single line of Norm changes the answer.