Voice is taught with examples
Collect three approved texts and explain why they work. Define the target audience, register, preferred words, and phrases to avoid. Asking for just a professional tone leaves too much room for interpretation. Do not upload private conversations or client materials without verifying that the tool is authorised.
Turn preferences into criteria
A useful criterion can be checked: concrete opening, few sentences, no superlatives without proof, a CTA. Distinguish the stable voice from the tone of the individual occasion. A comment on research and a response to criticism require different choices even if they belong to the same brand.
Request a justified revision
Ask for two versions to be proposed and request which choice differentiates them. Keep specific information and remove adjectives that add nothing. Do not insert invented experiences, emotions or testimonials on behalf of the organisation. If the model rewrites a quotation, restore the original text.
Build a repertoire of decisions
Note recurring corrections in a brief card. Update examples and criteria when the project changes. The result to aim for is a clear and recognisable message, not a text capable of passing an AI detector or concealing use that should be declared.
Sources and documents
Document date or reference period: Documentation consulted on 25 October 2026.
- WIPO — Generative AI: Navigating intellectual property ↗
- Meta — Transparency on AI-generated content (2026) ↗
IAnews24 desk guide prepared with AI assistance. Methods, prompts and examples are editorial team’s operational proposals, to be adapted to the brand and verified before use. They do not guarantee growth, coverage or commercial results.
