There are two dominant ways to make a model produce output in a target register, and teams tend to adopt one religiously without realizing it is a choice. The first is explicit rules: tell the model exactly what to do — no contractions, no exclamation points, lead with the conclusion. The second is few-shot exemplars: show the model one or two samples of the target voice and let it infer the pattern. Both work. They fail differently, cost differently, and scale differently, and the right answer depends on what kind of register quality you are protecting.
This article lays out the competing approaches, the axes along which they trade off, and a decision rule you can apply per task. There is also a hybrid that captures most of the upside of both, which is what experienced teams converge on. But you cannot use the hybrid well without understanding why each pure approach behaves the way it does.
The framing matters beyond this one decision. The same tension — precise-but-rigid versus flexible-but-fuzzy — recurs across prompt engineering. Understanding it here transfers.
Approach One: Explicit Rules
Rules state the target register as a set of instructions. "Use contractions. No exclamation points. Reserve qualifiers for genuine ambiguity. Address the reader as 'you.'"
Where rules win
- Auditability. Each rule is checkable against the output, which makes review and automated linting straightforward.
- Hard constraints. When a marker is non-negotiable — a banned word, a required politeness level — a rule enforces it deterministically. An exemplar only suggests.
- Teachability. A new teammate can read the rules and understand the voice. Tacit pattern-matching does not transfer in writing.
Where rules struggle
- They cannot easily express soft qualities: a dry rhythm, a habit of opening with the reader's situation, the particular cadence of a brand. Describing these in rules becomes a long, brittle list that still misses the feel.
Approach Two: Few-Shot Exemplars
Exemplars show rather than tell. Paste two samples of the target voice and the model matches their pattern.
Where exemplars win
- Soft qualities. Rhythm, cadence, and structural habits transfer through examples in ways rules cannot capture. The model imitates what it sees.
- Speed to a good result. For a voice that is hard to articulate, two good samples often beat an hour of rule-writing.
- Holistic consistency. Exemplars carry the whole register at once, avoiding the gaps a partial rule set leaves open.
Where exemplars struggle
- Unpredictable enforcement. The model may imitate the wrong feature of an exemplar — its topic instead of its tone. There is no guarantee a specific constraint holds.
- Hard to audit. You cannot point to "the rule that was violated"; the spec is implicit in the samples.
- Drift on edge cases. When the task diverges from the exemplars, the model improvises and register can wander.
The worked cases in Six Annotated Prompts Where Tone Either Landed or Backfired show both behaviors: some fixes were rules, some were better exemplars.
The Axes That Separate Them
Precision versus expressiveness
Rules are precise but cannot express soft qualities. Exemplars are expressive but imprecise. This is the core trade-off, and it maps to what you are protecting: hard constraints favor rules, holistic feel favors exemplars.
Auditability versus speed
Rules are auditable and teachable but slower to write for a subtle voice. Exemplars are fast to assemble but resist review and automated checking.
Stability versus flexibility
Rules hold constraints stable across edge cases. Exemplars adapt fluidly to the in-distribution case but wander when the task drifts from the samples.
The Hybrid and the Decision Rule
Combine rules for constraints, exemplars for feel
The approach experienced teams converge on uses rules for hard, auditable constraints — banned words, contraction policy, politeness level — and one or two exemplars to carry the soft rhythm. This is exactly the structure of the Vocabulary and Exemplars layers in The Anatomy of a Reusable Brand Voice Prompt. Rules fence the boundaries; exemplars fill the interior.
A decision rule
- If the register is dominated by hard constraints (legal, regulated, brand-protected words), lead with rules.
- If the register is dominated by a hard-to-describe feel, lead with exemplars.
- If both matter — which is most real work — use the hybrid: rules for what must not vary, exemplars for what is hard to name.
- If you need to audit or scale across a team, weight toward rules, because tacit pattern-matching does not transfer or instrument well. The instrumentation case is in Scoring Whether Generated Tone Actually Fits the Reader.
For teams just beginning, starting with a couple of rules and one exemplar is the fastest credible path, as laid out in Your Fastest Route to a First Reliable Tone Spec.
A Worked Comparison
The same task, two approaches
Consider a brand that wants warm, plain-spoken support replies. The rules-led version specifies: contractions on, no corporate filler, acknowledge the customer's frustration in one sentence, state the resolution directly. It is auditable and consistent, but the replies come out slightly mechanical — correct warmth, missing the easy human rhythm a real agent has.
The exemplar-led version pastes two of the brand's best historical replies and says "match this voice." The output captures the rhythm beautifully on typical cases, but when an unusual request arrives — one unlike the exemplars — the register wanders, sometimes turning oddly formal because the model improvises without a constraint to hold the line.
Why the hybrid wins here
The hybrid keeps the rules as a floor the output cannot drop below — contractions, no filler, acknowledge frustration — and adds the two exemplars to supply the rhythm. On typical cases the exemplars give the human feel; on edge cases the rules catch the output before it drifts formal. This is the general pattern: rules guarantee the boundaries, exemplars enrich the interior, and the combination degrades gracefully where either pure approach fails.
Cost of the hybrid
The hybrid is not free. It requires both writing the rules and curating the exemplars, and it makes the prompt longer. For genuinely low-stakes output, a couple of rules alone may be enough, and the exemplar curation is not worth it. The decision rule still applies: match the effort to what is at stake, and reach for the hybrid when both hard constraints and a distinctive feel genuinely matter.
Frequently Asked Questions
Should I use rules or examples to control tone?
It depends on what kind of register quality you are protecting. Use rules when hard constraints dominate — banned words, required politeness levels, anything that must be auditable. Use examples when a hard-to-describe feel dominates. Most real work needs both, so the hybrid is usually the right answer.
Why do examples sometimes produce inconsistent register?
Because the model may imitate the wrong feature of an exemplar — its subject matter or structure instead of its tone — and there is no guarantee a specific constraint holds. Examples carry the whole register implicitly, which is powerful for feel but unpredictable for any single rule you care about.
What is the main weakness of pure rules?
They cannot easily express soft qualities like rhythm, cadence, or a brand's structural habits. Trying to capture those in rules produces a long, brittle list that still misses the feel. That is precisely the gap exemplars fill.
How does the hybrid approach work?
Rules handle hard, auditable constraints — contraction policy, banned words, politeness level — while one or two exemplars carry the soft rhythm. Rules fence the boundaries; exemplars fill the interior. This maps directly to combining the vocabulary and exemplar layers of a structured voice spec.
Which approach scales better across a team?
Rules, because they are explicit and teachable. Tacit pattern-matching embedded in exemplars does not transfer well in writing and resists automated checking. When multiple people must produce one voice, weight toward rules and use exemplars as a supplement.
Does the choice change with model quality?
Somewhat. Stronger models infer register from exemplars more reliably, making example-led approaches safer. But hard constraints still belong in rules regardless of model quality, because even a strong model offers no guarantee that an implicit constraint holds on every generation.
Key Takeaways
- The two dominant approaches to register control are explicit rules and few-shot exemplars, and most teams adopt one without realizing it is a choice.
- Rules are precise, auditable, and teachable but cannot express soft qualities like rhythm and cadence.
- Exemplars capture holistic feel and reach a good result fast but enforce constraints unpredictably and resist auditing.
- The core axes are precision versus expressiveness, auditability versus speed, and stability versus flexibility.
- The hybrid that experienced teams adopt uses rules for hard constraints and exemplars for soft feel.
- Lead with rules when constraints or team-scale auditing dominate; lead with exemplars when a hard-to-describe feel dominates.