Voice and tone: a guide the model can actually follow

Tell a model your brand is “professional but approachable” and you will get hype back, every time. That phrase describes roughly every company on earth, so the model gives you the average of every company on earth. A voice guide that works is built the other way round: a hard list of things never to say, and real samples of things you already said. Here is the structure, and a method for pulling yours out of writing you have done rather than inventing it from scratch.

Why AI copy sounds like AI

This is not a mystery and it is not laziness. A language model picks the word that most plausibly comes next given everything it has read. Ask for marketing copy and it lands on the statistical center of all marketing copy ever published. That center is a specific, recognizable place: unlock, seamless, empower, at scale, game-changing.

So the generic voice is not a bug being patched out. It is the mechanism doing exactly what it does, pointed at a request with nothing in it to narrow the guess. If you want output that sounds like you, you have to narrow the guess yourself, and the narrowing has to be specific enough that a machine can act on it.

There is a business cost to skipping this. Copy that reads as machine-written tells a skeptical reader you did not care enough to write it, which is the opposite of the impression most of us are paying for.

Adjectives don't work. Prohibitions do.

Here is the test that settles it. Take two instructions:

A: “Write in a confident, friendly, professional voice.”

B: “Never use these words: unlock, seamless, empower, leverage as a verb, supercharge, cutting-edge, game-changing, at scale, world-class, future-proof. Do not promise anything is permanent or guaranteed.”

A changes almost nothing. B changes the first sentence you get back, because B is checkable. You can search the draft for a banned word and get a yes or a no. Nobody can search a draft for “confident.”

That is the underlying principle of the whole thing: a voice rule is only real if a stranger could check whether the copy obeyed it. Rules that pass that bar do work. Rules that don't are decoration, and they are why most brand voice documents change nothing when you hand them to a model.

The seven parts of a usable guide

Seven sections, in this order. The order matters because a model reading a long document weights the start and the end most heavily, so the audience frames everything and the checklist lands last.

1. The audience, in one paragraph. Not a segment. A person you could picture. Age, job, what they are already paying for, what they are skeptical about, how they found you, and what they are secretly worried about looking dumb over. Every later rule points back at this paragraph.
2. The rules. Eight to thirteen. Each one gets a title, a sentence of explanation, and a real quote from something you actually published. A rule with no example is an adjective wearing a numbered hat.
3. The forbidden list. Words, phrases, and promises. Two sections: the hype vocabulary, and the over-promises (anything asserting the future will not change).
4. Gold-standard samples. Three to five passages you wrote, liked, and shipped. Say underneath each one why it works. The why is what transfers.
5. Before and after rewrites. The generic version next to your version, with one line on what the difference is. This teaches faster than any rule, because it shows the move rather than describing it.
6. The channel split. Your voice on a landing page is not your voice in a client email is not your voice on camera. Say which register applies where, or a model will flatten all three into the landing-page one.
7. The pre-flight checklist. Six to ten boxes the writer ticks before handing anything back, ending with an escape hatch: if unsure, return a draft for review instead of publishing.

Extract, don't invent: the 40-minute build

The mistake is sitting down to decide your voice. You already have one. The job is archaeology, not authorship.

Step 1 · Gather what you actually wrote (10 minutes)

Five to ten samples, minimum. Emails you sent, not drafts you abandoned. A page you were happy with. A long message to a client. If a piece was written by someone else or heavily rewritten by a tool, leave it out. The corpus has to be genuinely yours or the guide will describe a stranger.

Step 2 · Make the model mine it (10 minutes)

Paste the samples and ask for patterns rather than praise:

Here are eight things I wrote. Don't compliment them. Tell me: what sentence patterns repeat? What kind of metaphor do I reach for? How long are my sentences and do they vary? What words appear often? What words would never appear in my writing, based on what you see? Give me ten candidate rules, each with a quote from these samples as evidence.

You will get eight decent rules and two wrong ones. That ratio is fine, and the two wrong ones are useful.

Step 3 · Correct it, and keep the corrections (10 minutes)

Where the model got you wrong, write the correction down as its own rule. This is the highest-value part of the whole exercise. Your objection to a mischaracterization is a sharper description of your voice than anything you would have volunteered.

Step 4 · Build the forbidden list from your own cringe (5 minutes)

Find one AI draft you rejected. Read it and write down every word that made you wince. That is the list, and it will be more accurate than any generic “words to avoid” article, because it is specific to what you personally cannot stand.

The strongest version of this move: take a draft a model wrote for you, edit it by hand until you would send it, then hand both versions back and ask what changed and why. Your edits are the rules. One section of our own guide was built entirely from the line edits on a single client email, and it captured things nobody would have thought to write down: cut the reassurance padding, sign with the short name, use the house term instead of a cute invented one.

Step 5 · Write the checklist (5 minutes)

Turn the rules into yes-or-no boxes. “Did I use any word from the forbidden list?” “Does every claim have a specific a skeptic would accept?” “Did I vary sentence length?” Put it at the end of the file, where it is the last thing read before writing starts.

Inside a real one: the Wasser AI guide

Ours is a single Markdown file that every agent reads before touching any published sentence. A few of its actual rules, so you can see the shape:

• Lead with the risk, not the benefit. The opposite of standard marketing, and the most distinctive thing in the file. “It feels like magic, and that's the problem. When you don't understand how it works, you can't tell when it's about to burn you.”
• Metaphors come from work, not from technology. The overconfident intern. Tools in a toolbox. The PhD student you finally gave a lab. Never “neural networks” or anything that smells like a data science course.
• Admit the limits of the thing you are selling. Vendors don't say their product is unreliable. We do, and that is why the recommendation after it gets believed.
• Use em-dashes rarely. A dead giveaway of machine-drafted text is an em-dash every third line. Prefer a colon, a period, or a rewrite.
• Under-promise and over-deliver. No “free for life,” no “forever,” no “guaranteed updates.” State what is true right now. “It's free while we're in the founding phase” is honest today and still honest after the price changes.

The forbidden list runs about two dozen entries: unlock, seamless, supercharge, empower, next-level, revolutionary, game-changing, cutting-edge, leverage as a verb, synergy, disrupt, transformative, at scale, enterprise-grade, paradigm shift, future-proof, best-in-class, world-class, learnings as a noun. Plus a separate over-promise list.

One entry on that list is worth explaining, because it is counterintuitive. The word calm is banned, as a description of our own brand. Not because the quality is wrong, but because announcing it is. If the writing has to tell you it is measured and hype-free, it isn't. Demonstrate the quality with specifics and let the reader name it. Show, don't state, applies to your own adjectives first.

And the rewrite that explains the whole file in two lines:

Generic: “Unlock the power of AI to transform your business.”

Ours: “Stop theorizing. Start building.”

Keeping it alive

One file, one location, referenced by path in your agent's standing instructions so it is loaded before any copy gets written rather than pasted per task. Date it at the top so you can tell when it last reflected reality.

Then update it on a trigger, not a schedule. Three triggers are worth acting on: a new drift pattern appears in generated output that no rule covers, you rewrite a whole page and your voice moved with it, or you catch yourself hand-fixing the same thing for the second time. That last one is the loudest signal. A repeated correction is a missing rule, which is the same principle behind the reflexion loop.

The guide is also what makes delegation safe. When a cheaper model or a junior writer does the drafting, the guide is the contract, and the checklist is how you check the work without reading their mind. There is more on that in delegation patterns.

What breaks

The guide is too long to be read

Past a few thousand words the middle gets skimmed by humans and weighted less by models. Keep the rules tight and push the history, the rationale essays, and the retired sections to the bottom or into an archive.

The rules have no examples

“Be direct” is not a rule. “Be direct, as in this sentence I actually shipped” is. Every rule needs a quote from your own writing carrying it, or it degrades back into an adjective.

Tone gets edited after generation instead of before

If you write the prompt without the guide and fix the voice afterward, you are rewriting every piece by hand, which is doing the work twice and getting paid for neither pass. The guide goes in before the first word, not after the draft.

It is written once and never touched

A guide that has not changed in a year is describing who you were a year ago. Ours got rebuilt when it turned out it had never captured how Robert sounds when speaking, only how he writes. That rebuild changed the em-dash rule, added a spoken register, and banned a word the old version had recommended.

Only you can tell whether copy passed

If applying the guide requires your taste, it is not a guide, it is you. Push every subjective rule toward something checkable until a stranger could audit the draft with it.

Your turn: build yours today

1. Download the worksheet below. It is the seven sections with the prompts already written, and blanks where your material goes.
2. Spend ten minutes gathering samples. Sent emails, a page you liked, a long client message.
3. Run the mining prompt from Step 2 and let the model propose rules with evidence.
4. Argue with it. Keep every correction as a rule.
5. Save the finished file where your agent reads it, and name it in your standing instructions: “read this before writing anything user-facing.”
6. Test it once. Ask for the same paragraph with and without the guide loaded. If the two are hard to tell apart, your rules are still adjectives. Go back and make them checkable.

Forty minutes, and every piece of copy after it starts closer to done. The measure of success is not that the drafts are perfect. It is that you stop rewriting the same sentence for the same reason.

Back to

All workflows

→