Build an AI Writing Voice Brief—and Test Whether It Works

A voice brief tells an AI writer how you prefer to communicate. Supplying examples and instructions is not necessarily model training: a tool may place them in the prompt, retrieve them as context or use a separate tuning process. Ask which mechanism your system uses before assuming it learns automatically.

Choose samples with permission and purpose

Use writing you are entitled to share with the tool and that represents the voice you want to publish. Remove confidential customer or colleague information. An internal email may sound natural while being inappropriate to upload.

Start with a varied handful of approved examples: an explanation, an opinion and a practical update. There is no universal sample count that guarantees a match. Add examples to resolve observed gaps rather than treating volume as the objective.

Describe observable choices

A useful brief for a fictional operations consultant might say:

• Explain the operational problem in the opening paragraph.
• Prefer “client” to “customer” for agency work.
• Use connected paragraphs; use a list when the steps are genuinely sequential.
• State the tradeoff without forcing a contrarian conclusion.
• Avoid sports metaphors and invented personal scenes.
• End with a practical action when the piece teaches a process.

Attach a short example beside an instruction when the meaning could be ambiguous. “Conversational” can mean different things to different writers.

Keep facts in a separate section

Writing style does not establish product truth. Store the current offer, approved claims and source links separately from vocabulary and rhythm preferences.

If a past sample says “we onboard everyone in a day,” that sentence should not become a current capability claim unless it remains accurate. Tell the writer that samples demonstrate style, not permission to reuse every fact.

Test on material the system has not seen

Hold back one approved sample or use a new source brief. Ask for a draft, then compare it against the voice instructions. If you test only by rewriting the supplied samples, you may be measuring imitation rather than transfer to a new task.

Record concrete mismatches: an opening that delays the point, a phrase you would not use, or a conclusion stronger than the evidence. Ask for a revision addressing those issues.

Save the correction deliberately

Turn recurring edits into brief updates. Check where the updated brief is stored and whether a later task uses it. An edit is potential feedback; it is not proof that the product updated a model or saved a preference.

FeedSquad separates these jobs: Stash holds business context and voice guidance, while Ghost prepares LinkedIn drafts. Review proposed memory changes before confirming them. Confirming a fact does not authorize its public disclosure or approve a post.

Measure review time and the kinds of corrections still required. Fewer edits can indicate improvement, but only if you are maintaining the same quality bar.

Keep voice subordinate to truth

A draft can sound convincing and still contain an invented story. Run the evidence and quality review before publication. Use the ghostwriter brief to keep the author's actual position and experience visible.

To reuse those preferences in another assistant, build a portable brand brief with approved examples, source dates and an owner for corrections.