How to Write a Brand Voice Guide an LLM Can Actually Use

Your brand voice guide is probably a beautiful PDF. It has a page about your “tone spectrum” with a slider between Playful and Professional. It says you’re “confident but never arrogant” and “human, not corporate.” Somebody spent three weeks and £8,000 on it.

Paste that into Claude and ask for a 600-word blog intro. You’ll get something that reads like every other SaaS blog intro published since 2019.

This isn’t the model failing. It’s a translation problem. A brand voice guide for AI needs to do a different job from one written for humans: humans infer from vibes and examples they’ve absorbed by osmosis, while a language model needs explicit constraints it can check its own output against. The good news is that converting one into the other takes about half a day, and the improvement in first-draft quality is the single biggest lever most small content teams have.

Why “confident but never arrogant” produces nothing

Adjectives without referents are noise to a model. “Confident” has been used to describe Patagonia, Goldman Sachs, Oatly and the Ministry of Defence. When a model reads it, it falls back on the statistical average of everything labelled confident in its training data, which is corporate LinkedIn English.

Here’s the test I run on any voice guide before rewriting it. Take every adjective in the document and ask: could a competitor in my category claim the opposite and still sound sane? If “we’re clear, not jargon-heavy” has no plausible opposite (nobody markets themselves as deliberately confusing), the line is doing zero work. Delete it.

What survives that test tends to be specific and slightly uncomfortable. One fintech client of mine ended up with “we assume the reader has been lied to by a bank before.” That’s a voice instruction a model can act on. It changes verb choice, it changes how claims get hedged, it kills the phrase “seamless experience” on contact.

The five components that actually change output

After rebuilding voice guides for roughly a dozen teams, the format that works is short: about 600 to 900 words total, structured as five blocks. Longer than that and models start ignoring the middle, a well-documented attention pattern in long prompts.

1. Reader state, not reader persona. Skip the demographic profile. Write two sentences on what the reader believes, fears or has already tried before they land on the page. “They’ve bought a tool like ours before and it failed during onboarding” is more useful than “Head of Ops, 35-50, SME.”

2. Banned list, with reasons. Not a style guide, a kill list. Twenty to forty items. Include the AI tells (delve, leverage, in today's fast-paced world, it's not just X, it's Y, navigate the complexities of), the industry clichés, and your own overused words. Add the reason in brackets: models follow constraints better when the rationale is present than when it’s a bare prohibition.

3. Sentence mechanics as numbers. This is the component most guides skip and it does the most work. Average sentence length target. Maximum sentence length before you break it. Whether you use fragments. Whether you use second person. Paragraph ceiling in sentences. Contractions yes or no.

4. Three annotated examples. Not clean examples: annotated. A paragraph of real published copy, followed by two or three bracketed notes explaining why it works. [Opens on the reader's problem, not our product. Verb in the first six words.] The annotation is what makes the example transferable rather than something to copy stylistically.

5. One negative example. A paragraph that is grammatically fine, on-topic, and wrong for you. Label it clearly. This is surprisingly powerful, because it gives the model a boundary rather than just a target.

A worked before-and-after

Here’s the voice section of a real guide (details changed), first as it arrived:

Our tone is friendly and approachable while maintaining professional credibility. We speak as a trusted advisor. We avoid jargon and speak plainly. We are optimistic about the future of work.

Now the rebuilt version of the same brand:

Reader state: They run HR for 80 to 400 people and have been handed a compliance deadline by someone senior who doesn’t understand the work involved. They are not excited about software. They want to know if this removes a specific task from their week.

Mechanics: Average 14 words a sentence. Hard stop at 28. Fragments allowed, roughly one per four paragraphs. Second person throughout, first person plural only when describing our own decisions. Maximum three sentences per paragraph. Contractions always.

Never write: empower, journey, solution (use the actual noun), seamless, robust, unlock, “in an era of”, “it’s worth noting”, “at the end of the day”, rhetorical question as an opener, any sentence beginning “Whether you’re”. [These all signal copy written to fill space rather than answer something.]

Always: name the specific regulation or form number rather than “compliance requirements”. Give a number when one exists. If we don’t know the number, say we don’t.

Good example: “The deadline is 6 April. If your payroll data lives in more than one system, start now, because reconciling two sources takes most teams a fortnight.” [Date in the first four words. Assumes competence. The warning is concrete and costed in time, not urgency-flavoured.]

Wrong for us: “In today’s evolving regulatory landscape, HR leaders are navigating unprecedented complexity. The good news? You don’t have to do it alone.” [Three sentences, zero information. Ends on a sales move before earning it.]

The second version is 230 words. Dropped into the same prompt, the difference in first-draft usability is not subtle: the team went from rewriting roughly 70% of each draft to closer to 25%, measured over eight pieces.

Where to put it so it actually gets used

A voice guide nobody attaches is worth nothing, so the delivery mechanism matters as much as the content. Four options, roughly in order of how much setup they need:

Claude Projects or ChatGPT Custom GPTs. Paste the guide into project instructions or custom instructions. Every conversation in that project inherits it. This is the default for a one-to-five person team and takes about ten minutes. One project per brand if you’re an agency.

A pinned snippet. If your team works across several tools, keep the guide in a single Notion page or a text file and paste it at the top of every session. Crude, but auditable, and people actually do it if it’s one copy-paste.

Claude Code or the API with a system prompt. If someone on the team is comfortable with a terminal, putting the guide in a CLAUDE.md file or a system prompt means it’s version-controlled and you can diff changes to your voice over time. Overkill for most, excellent if you publish at volume.

Writer or Jasper’s brand voice features. These will train on your existing content and generate a voice profile automatically. Worth trying, with one caveat: they tend to average your archive, which means if 60% of your back catalogue is mediocre, the profile learns mediocre. Use the generated output as raw material for a human-edited guide, not as the guide.

The broader question of where voice fits alongside briefing, drafting and editing is covered in our pillar on drafting, brand voice and editing workflows, which is worth reading if you’re building the whole production chain rather than fixing this one piece.

Testing it, which takes 40 minutes

Don’t ship a voice guide you haven’t tested. The protocol:

Pick three content types you actually produce (say, a blog intro, a LinkedIn post, a customer email). Generate each one twice: once with no voice guide, once with it. Same model, same prompt otherwise. You now have six outputs.

Give them to a colleague unlabelled and ask them to sort into “could be us” and “could be anyone.” If the guided versions don’t sort cleanly into the first pile, your guide isn’t specific enough, and the usual culprit is that the mechanics section is missing or the examples aren’t annotated.

Then count your edits. Take one guided draft, edit it to publishable standard, and count how many sentences you changed. Under 30% and the guide is earning its keep. Over 50% and something in it is being ignored, usually because the guide is too long or contains contradictory instructions (the classic being “be conversational” alongside a ban on contractions).

What to revisit, and when

Voice guides rot in a specific way: the banned list ages fastest, because model defaults shift with each release and new tells appear. Delve was the giveaway in 2024. By now it’s more often the three-item list with a colon, or the habit of restating the question before answering it. Add a line to your quarterly content review to read three recent AI-assisted drafts and note any phrase that appears in all three. That phrase goes on the list.

The mechanics section, by contrast, should barely change. If you find yourself adjusting sentence length targets every quarter, you’re not refining your voice, you’re chasing a preference. Pick numbers, keep them for a year, and see what your best-performing pieces have in common at the end of it.

One last thing worth knowing: the guide will improve your human writers too. Teams that go through this exercise usually find that the act of writing down “maximum 28 words” and “name the form number” settles arguments that had been running unresolved in Slack for months.