Keeping brand voice when a model writes the first draft

A client sent us forty generated subject lines to review. Not one of them was wrong. Grammar fine, length fine, offer stated accurately, no legal problem anywhere. Read three in a row and they were unobjectionable. Read all forty and they could have come from any brand in the category.

That is the failure mode worth planning for. Generated copy rarely produces the disaster people fear. It produces competence without character, which is harder to argue against in a review meeting and does more damage over a year.

The brands keeping their voice are not the ones with the best prompts. They are the ones who wrote down what their voice is in terms specific enough to be checked, and who kept a person at the one point in the process where a message becomes particular.

Write the voice down as rules, not adjectives

Most brand guidelines describe voice with adjectives: warm, confident, human, never corporate. Those words mean nothing to a model, and not much more to a new copywriter. Every brand in the category has the same four.

Rules are different. A rule can be followed or broken, and you can tell which by reading the output.

• Sentence length: no sentence over twenty-five words, and at least one under eight in every paragraph.
• Person: second person throughout, never third-person references to "customers" in customer-facing copy.
• Forbidden constructions: no rhetorical questions as an opening, no "we're excited to", no sentence beginning with "Introducing".
• Numbers: prices always with the currency symbol, delivery times as a date rather than a duration.
• Claims: no superlatives without a source line, no comparison to competitors by name.
• Humour: allowed in subject lines and social, not in service messages or anything about money.

Twenty rules of that kind do more than three pages of adjectives. They also let you check output mechanically, which matters when the volume is high enough that nobody can read everything.

Write the rules from your own best work, not from theory. Pull the fifteen campaigns your team is proudest of, read them together, and write down what they do that other brands do not. The rules are already there in the copy.

Give the model your material, not your adjectives

A brief that says "write in a warm, confident brand voice" produces the average of everything the model has read. A brief that includes eight of your own approved messages, with a note on what makes each one right, produces something much closer to yours.

Build a voice pack and keep it in version control rather than in a prompt somebody pasted from a document.

1. The rule list above, verbatim.
2. Six to ten approved examples per format: subject line, push, hero paragraph, service message.
3. A short set of rejected examples with a one-line reason for each. Negative examples do more work than people expect.
4. The product and category vocabulary, including the words you deliberately do not use.
5. The legal and claims boundaries, kept as the same text your review team uses.

Refresh the examples quarterly, and always by adding recent work that was approved. A voice pack made of two-year-old campaigns will hold you at a version of the brand you have already moved past.

Keep a human where the message becomes specific

The useful split is not between human-written and machine-written. It is between the parts of a message that are generic and the parts that are particular.

Generic parts scale well: structural variants, length trims, translating one approved message into the house tone for a second channel, producing ten options for a line that already has a clear idea behind it.

Particular parts do not. The angle, the reason this message exists now, the single sentence that makes a reader feel understood, the decision to say less than you were asked to. That is where a person earns the salary, and it is also where the model has no information, because the reason lives in a conversation it never heard.

The practical consequence for dynamic content is that you should generate the variants and hand-write the frame. Blocks that change per recipient can be produced at scale from a template with tight rules. The message-level idea that holds them together should not be.

Where the variable part is a product recommendation, keep the model out of the merchandising decision entirely. Let the recommendation logic choose the products and let generation handle only the wording around them. A model asked to both pick and describe will happily write a persuasive sentence about the wrong item.

The approval and disclosure side of this is separate, and we set out the process we use in last week's piece on generative content and the rules. Voice work assumes those controls are already in place.

Review the batch, not the message

Message-by-message review misses the actual problem, because each message passes. The characterlessness only appears in aggregate, which means the review has to happen in aggregate too.

Once a month, take everything sent in the period and read it in one sitting, in send order. Look for four things:

• Repetition of structure. If eleven subject lines open with a question, the rule list needs a new rule.
• Drift towards the category average. Would a competitor's logo fit on this without anyone noticing?
• Vocabulary collapse. Models reach for the same intensifiers. Count them, and ban the top three for a quarter.
• Sameness across channels. Push and email should not read identically; the constraints are different and so is the moment.

Keep a rejection log with a reason against each entry. Over a few months that log becomes the most accurate description of your voice anyone has written, and it feeds straight back into the negative examples in the voice pack.

Test voice as a variable, not as a matter of taste

Voice arguments in review meetings are unwinnable because both sides are stating preferences. Move the question into a test where you can.

An A/B test between a tightly on-voice variant and a neutral one is straightforward to run on a broadcast, and the read should be the click or the order rather than the open. Expect small differences on any single send. Voice is a long-run asset, and a single subject-line test is not the instrument for measuring brand equity.

What a test does reliably tell you is whether a specific rule costs you anything. If the rule against rhetorical questions turns out to be free, keep it for consistency. If a particular construction consistently underperforms across several tests, the rule list should change and the voice pack with it.

Hold the rules as a living document with an owner, a version and a date. The failure we see most often is a voice pack written carefully in month one and never touched again, while the copy quietly drifts back to the average it came from.

What to do this week

• Pull your fifteen best recent messages and write twenty checkable rules from what they do, replacing the adjectives in the current guideline.
• Build the first voice pack with approved and rejected examples per format, and put it somewhere versioned rather than in a chat thread.
• Book a monthly batch review, read a full month of output in send order, and start the rejection log.