A style preset is your default “sounds like us” instruction. It is broader than a rule: it sets tone, rhythm, and attitude. Rules handle constraints; presets handle voice. Weak presets read like marketing adjectives (“friendly, engaging”)—strong presets include examples and boundaries.
What belongs in a preset
Sentence length preference: short and punchy versus flowing.
Second person: how often you say “you” versus “we.”
Humor level: none, light, or central—be explicit.
Signature phrases you love or hate.
Audience expertise: beginner definitions on first mention or not.
Include two or three example sentences copied from your own best content—not generic samples. Models learn faster from real lines than from adjectives.
Starting from a system preset
Pick the closest system preset as a base.
Rewrite the description in your words—do not leave default boilerplate.
Add one negative constraint (“no rhetorical questions in openings”) if needed.
Generate a short paragraph, compare to your samples, adjust wording.
When outputs drift
Drift usually means ambiguity: the model guessed tone. Tighten the preset with a counterexample (“Do not sound like a motivational poster”) and reduce conflicting rules elsewhere.
Preset versus rules
Preset: voice and rhythm.
Rules: must-include disclaimers, banned claims, structure requirements.
If you have twenty rules, your preset is doing too little—consolidate.
Update the preset text before adding more rules—often the model needs clearer intent, not more guardrails.