Use scales, not adjectives
Place the voice on four scales from 1 to 5: formal to casual, serious to playful, sober to enthusiastic, brief to expansive. “Casual 4, playful 2, sober 4, brief 5” is a description a model can actually follow, and one two colleagues read the same way.
Show three examples you are proud of
Find three texts that sound like you at your best: a post, an email, a product description. Paste them into the profile. Style spreads from examples in a way descriptions never manage.
Show one example you would never write
The negative example is often the most useful. Rewrite one of the good examples into something you would reject, with exclamation marks, clichés and sales pressure, and mark it clearly as “not like this”.
Make the list of banned words
Every brand has words it avoids: “revolutionary”, “solution”, “unique”, “we are proud to”. Write them down. It is the easiest rule to enforce, and the one that most often gives AI text away.
Decide the small things
- You, singular or plural? We, or the company name?
- Exclamation marks: never, rarely or freely?
- Emoji: which, where, never?
- Jargon: avoid, sparingly, freely?
- Humour: dry, warm or none?
Test it
Give the description to a colleague and to an AI, and ask both to write the same short text. If the results resemble each other, the description is good enough. If not, it is missing an example or a rule.
In Marshi
In Marshi the voice lives in the brand as fields, not as a document. Everything Marshi writes passes through it, and you can check any draft against it before it goes out.
