Why a human voice guide is not enough for a model
A typical voice section says the brand is warm, plain-spoken and expert, gives a "we are, we are not" table and offers a few example lines. That works for a copywriter who has read the brand's emails for a year, sat in the meetings where a headline was rejected and knows which customers complain about jargon. The adjectives are shorthand for a body of shared judgement, and the writer supplies the judgement.
A model receives only what is in the request and the material attached to it. "Warm" with no example gets the model's default idea of warmth, which tends towards exclamation marks, compliments and phrases such as "we're so excited". "Plain" gets short sentences, but not necessarily the brand's choice of which details to leave out. "Expert" often comes back as confident claims, which is the opposite of the honest, bounded expertise most voice guides mean. The result is a draft that ticks every adjective and reads like nobody in particular.
Models follow demonstrated patterns more reliably than described ones. That is the whole premise of this tutorial: every instruction about voice should come with a passage that shows it, ideally a pair that shows the same content written the wrong way and the right way. Paired examples carry the judgement the adjectives leave out, and they give a reviewer something concrete to compare a draft against.
| What a human guide says | What a model does with it | What the voice pack adds |
|---|---|---|
| "Warm but expert" | Produces its default idea of warmth, often gushing | A before-and-after pair for each trait |
| "Avoid jargon" | Removes some technical words, keeps its own stock phrases | A word list of avoided terms with the brand's replacements |
| "Adapt tone to the channel" | Guesses what each channel needs | A short adjustment per channel, with an example line |
| Example headlines | Copies their surface pattern | Labelled reference passages explaining what each demonstrates |
| Nothing about format | Adds headings, bullets and sign-offs by habit | Explicit format rules per output type |
The AI for content guide places drafting in the wider content process; this tutorial deals only with making the drafts sound like the brand.
Before you start
- An approved voice guide. The voice principles, tone by situation and vocabulary should already exist, ideally in the shape the brand guidelines template sets out. If the brand has no agreed voice, define it first; a model cannot discover a voice the business has not decided on. House rules on spelling, product names, numbers and dates belong in the separate editorial style guide, which a pack can share.
- The positioning. The positioning statement explains who the brand serves and why it is different, and the traits should trace back to it.
- Approved writing to learn from. Published pieces the brand team considers on-voice, across channels, plus a few that were rejected or rewritten and the reason why.
- An approved AI tool. One your organisation permits for this use, with terms the team has read, since the pack will contain unpublished material.
- A small group of reviewers. At least one person who owns the voice and one person who writes for the brand every week.
- Real briefs. A handful of recent briefs across different channels and audiences, to test the pack against.
Building the voice pack, step by step
Gather approved writing
Collect a set of passages the brand team agrees are on-voice: product pages, emails, social posts, support replies, a long-form article. Choose variety over volume, because the aim is to show the voice across situations, not to give the model one channel's habits. Add a smaller set of passages that were rejected or heavily edited, each with a one-line note on what was wrong. Strip out anything confidential and any facts that may have changed, such as prices and offers, since the model will treat everything in the pack as usable material. Label each passage with its channel, audience and purpose so that it can be retrieved for the right brief later.
Name the traits
Take the voice principles from the existing guide and reduce them to a short list of traits, each with a name, a one-sentence definition and a boundary that says where the trait stops. "Plain: we use the everyday word and say the price; plain is not curt, so we still explain why." Three or four traits are usually enough, which matches the advice in the social brand voice guide. Check each trait against the approved passages: if no passage demonstrates it, either the trait is aspirational and should wait, or the passages need replacing.
Write paired examples
For each trait, write pairs: the same content written off-voice and on-voice, with a line explaining the difference. Make the off-voice version realistic, the kind of draft a model or a hurried writer actually produces, not a caricature. Cover the situations where the trait is hardest to hold: a price rise, an apology, a technical explanation, a call to action. Pairs should hold the facts constant so the only difference is voice, which teaches the model what to change and stops it learning content from the examples. These pairs are the most valuable part of the pack and deserve the most review time.
Set word lists
Write three short lists. Preferred terms: the brand's words for its products, customers and actions, and how it writes the company name, dates and figures. Avoided terms: the industry jargon, cliches and stock phrases the brand never uses, each with its replacement. Banned claims language: words that imply guarantees, superlatives or outcomes the business cannot support, which the review step will also catch. Models have favourite phrases, so after the first test round add the ones that keep appearing in drafts to the avoided list with a replacement.
Adjust by channel
Voice stays the same across channels; tone and format change. For each channel the brand writes in, record the tone adjustment, the format rules and one example line: sentence length, whether headings or lists are used, how a reply opens and closes, how much humour is allowed. Include the situations the existing guide covers, such as complaints and outages, and say which traits are dialled down in each. Tie each channel to the Buyer PersonaA research-based profile of a typical customer type: goals, problems and buying behaviour. it mostly serves, because the same trait reads differently to a first-time buyer and to an experienced customer.
Write the instruction block
Assemble a single block of instructions that every drafting request can include. Start with a short statement of who the brand is and who it writes for, then the traits with their definitions and boundaries, then the word lists, then the format rules for the output requested. Attach the paired examples and two or three labelled reference passages chosen for the brief's channel. Keep the language direct and free of contradictions: if one line says "be concise" and another says "always explain the reasoning", the model will pick one. Store the block somewhere the team can reach and version, not in one person's chat history; a shared prompt library, set up as the prompt library tutorial describes, is the natural home.
Test against real briefs
Run the pack on the real briefs gathered earlier, with the facts each brief needs supplied separately. For each draft, reviewers score every trait as on-voice, partly or off-voice, note the phrases that broke it and mark any format rule that was ignored. Run the same briefs without the pack as a baseline, so the team can see what the pack is changing. Read the drafts blind where possible, mixed with passages a person wrote, and note whether the reviewers can tell them apart. Change one part of the pack at a time between rounds so that improvements can be traced to their cause.
Review and version
Give the pack an owner, a version number and a change log. Record what changed and why: a new avoided phrase, a replaced example, a new channel. Re-run the test briefs after every change, because a fix for one channel can break another. Review the pack whenever the brand guidelines change, when a new channel is added and on a fixed schedule otherwise. Keep the previous version available so a drafting problem can be traced to the pack version that produced it.
A voice pack (illustrative)
The structure below is what the steps produce. The content in each section is short; the examples carry the weight.
| Section | Contents | Owner check |
|---|---|---|
| Brand and audience | Who the brand is, who it writes for, the positioning in one paragraph | Matches the positioning statement |
| Traits | Name, definition and boundary for each trait | Every trait demonstrated by an approved passage |
| Paired examples | Off-voice and on-voice versions with the difference explained | Facts identical within each pair |
| Word lists | Preferred terms, avoided terms with replacements, banned claims language | Updated after each test round |
| Channel adjustments | Tone, format and an example line per channel and situation | Covers complaints and outages |
| Reference passages | Labelled approved writing, retrieved by channel | No stale prices or offers |
| Test set | Real briefs, scoring sheet, baseline drafts | Re-run after every change |
| Change log | Version, date, change, reason, approver | Owner named |
A voice pack for a loose-leaf tea retailer
Brackenhollow Teas, an invented online tea retailer, has a human voice guide with 3 principles: knowledgeable, unhurried and plain about price. Its first AI drafts open with "Calling all tea lovers!" and describe every blend as "exquisite". The marketing lead builds a pack from 24 approved passages and 6 rejected ones, then writes 4 pairs per trait. One pair for "plain about price" sets "An exclusive opportunity to invest in our finest harvest" against "This tea costs more because the leaves are picked by hand, once a year." The avoided list starts with 11 terms and grows to 19 after the first test round, when "elevate", "indulge" and "journey" keep appearing. The team tests 8 real briefs across email, product pages and social replies. In round one reviewers mark 9 of the 24 trait scores off-voice; after the pairs are rewritten to cover apologies and delays, round two has 3. The remaining failures are all in social replies, so the team adds a channel section for replies with 5 example lines and records the change as version 1.2 in the log.
Keeping voice and facts separate
A voice pack should contain nothing that the model might repeat as a fact. Prices, offers, product specifications, delivery promisesA store's claim about what delivery will cost and when an order will arrive, made on product pages, at checkout, in confirmations and in ads., awards and customer numbers change, and once they sit in reference passages the model will reuse them in drafts long after they stopped being true. Remove them from approved writing before it enters the pack, or replace them with obvious placeholders.
Facts for each piece come from a source pack attached to the brief: the product sheet, the current offer, the approved claims, the proof points the business can substantiate. The model is told to use only those facts and to mark anything it needs but does not have. This split keeps responsibilities clear. The voice owner maintains the pack; the product, legal or compliance owners maintain the facts; and reviewers check a draft twice, once for voice and once for every fact and claim against the source pack, using the tiers in the accuracy and review workflows guide.
The split also protects the voice. When a draft contains an invented claim, the temptation is to add another instruction to the voice block. That makes the block longer and less consistent, and it fails anyway, because voice instructions cannot make a model know what is true. Keep claim rules in the word lists only as language to avoid, and leave verification to review. Subjects work the same way: the brand's Content PillarOne of a small set of recurring themes that a brand's content is planned around. topics belong in the content strategy and the brief, not in the voice pack.
Common mistakes
- Pasting the human guide and calling it done. Adjectives without paired examples produce generic drafts that use the brand's words without its judgement.
- Writing caricatured off-voice examples. A pair only teaches if the wrong version looks like a real draft.
- Letting facts into the pack. Old prices and offers in reference passages resurface in new drafts.
- One huge instruction block. Contradictory lines make the model choose; keep the core short and add channel sections only when needed.
- Testing on invented briefs. The pack has to work on the briefs the team actually writes.
- Changing several things between test rounds. Without one change at a time, nobody knows what helped.
- No owner or version. A pack edited by everyone drifts, and drafting problems cannot be traced.
- Using the pack to define the voice. If the brand has not agreed its voice, the pack will encode whatever the model produced first.
