How to set up your writing voice so every draft sounds like you

The first draft comes back and it is fine. Correct, polite, well-structured, on topic. And you will not send it, because it sounds like a company rather than like you, and you cannot immediately point at the sentence that is wrong.

That reaction is worth trusting and worth diagnosing properly, because the instinctive response — go looking for a friendlier tone setting — does not work. The draft was not written in the wrong tone. It was written for the wrong business: a plausible generic version of your industry, assembled because nothing more specific was available.

What is the goal?

A configuration you fill in once, after which every draft — comment, DM, planner post, rewrite — carries your vocabulary, your rhythm and your habits, without you restating them each time.

This matters more than any other setting because of where it sits. Your voice instructions are layered last into the generation, after the structural scaffold, the length default and the style rule. Last means they win where they contradict anything else. It is the highest-leverage field in the product.

What is broken today?

Three things, and they compound.

Generic is the correct answer to a generic brief. Tell a model you run a marketing agency and it writes for the average of every marketing agency it has seen. That average is real, coherent, and exactly what you are reacting to.

Fluency is the default. Absent instruction, models produce tidy, evenly weighted, slightly formal prose with a fondness for three-part lists and a habit of restating the question before answering. Most people do not write that way.

People describe their business the way their website does. A document written to sound impressive rather than to be specific. "We help ambitious businesses grow" is a mood, and a draft written from a mood is about growth in general.

If your business description would also be true of four competitors, you have written a category, not a business — and the draft will sound like the category.

How do you configure your voice, step by step?

Six fields. Two of them do most of the work.

1. What you actually sell. The services, named, including the ones you do not offer.

2. Who buys it. The specific customer, including the ones you turn away.

3. How you talk. Formality, sentence length, contractions, first person singular or plural.

4. Words you never use. Ten to fifteen. This is the second high-leverage field.

5. Three real examples. Replies you actually sent. Not marketing copy.

6. What you will not claim. Covered separately in the guardrails guide.

If you have not installed anything yet, do the fifteen-minute setup first — these fields are step two of it.

Written out, fields one and two look like this — note that every sentence is a fact rather than an adjective:

Domestic plumbing and heating: boiler repair, servicing and replacement, plus
bathroom installs. No commercial work, no new-build. Covering [area] within
about 20 miles.

Most customers are homeowners aged 35+, usually calling because something has
already broken rather than as planned maintenance. Price-aware but not the
cheapest-quote type — they want it done once, properly.

We turn down: emergency callouts outside hours, anything gas-related on
commercial premises, and jobs where the customer has already had two other
people out and wants a third opinion for free.

That last paragraph is the one people leave out and the one that changes the most. What you turn down defines the business more sharply than what you accept, and it is what stops a draft being accommodating about work you do not want.

Then field three, in the same concrete register:

Short sentences. Contractions. First person singular — I, not we, because it is
mostly me. Direct: say what the problem is before saying what it will take.
No exclamation marks, ever. Never say "reach out", "touch base", or "at your
earliest convenience".

Why does a banned-words list beat a tone setting?

Because a prohibition is checkable and an adjective is not.

"Be conversational" has to be interpreted, and interpretation drifts — over a long thread, under a complicated question, when the previous message was formal. "Never use the word leverage" does not drift. It either appears or it does not.

It is also a more accurate model of how voice works in practice. Most people cannot describe their own writing style, but everyone can immediately name five words they would never type. That list is a sharper fingerprint than any self-description, because it is made of real reactions.

A starting list to react to: delve, leverage, seamless, robust, unlock, elevate, game-changing, revolutionise, synergy, reach out, touch base, at your earliest convenience, absolutely, exciting news. Cut the ones you actually use, add the two from your own industry that make you wince.

Everyone struggles to describe how they write. Everyone can name five words they would never type. Use the second list.

How does Learn my writing voice work?

If describing your own style feels like guesswork, do not describe it — derive it.

Learn my writing voice takes a sample of writing you have already published and produces a compact, reusable description of how you write: tone and register, sentence length and rhythm, vocabulary level, punctuation habits, emoji and hashtag use, how you open and close, and any recurring quirks.

The workflow:

1. Collect a sample of your own posts or replies — your real ones, not your best ones.
2. Run Learn my writing voice over the sample.
3. Read the description it produces. It will name habits you did not know you had.
4. Edit it — cut anything that was an accident of the sample rather than a habit.
5. Save it as your voice instructions.

Two things worth knowing. It describes how you write, not what you wrote about — if it comes back summarising your topics, the sample was too narrow. And it describes only what the sample actually evidences, so a thin sample produces a thin description. Twenty real replies beat three polished posts.

The best results come from combining the two approaches: derive the description from a sample, then add your banned-words list by hand. The sample captures rhythm; the list captures taste.

How do you handle several brands?

One persona per voice, and be strict about it.

A persona is a brand. It carries its own business facts, its own voice, its own platform set and its own style default, and you switch between them from the panel without signing out.

The tempting shortcut — one persona with a "depending on which client" note in the context — produces drafts that sound like the average of all of them. That is worse than either brand's generic default, because it is inconsistent as well as bland.

The cost is real: three brands means writing the fields three times, about half an hour. That is the difference between drafts you send and drafts you rewrite.

One thing that trips people up: a persona is not a login. Adding a persona does not add a teammate. Colleagues who need their own access are workspace members, which is a separate thing.

How do you test that it worked?

Do not test on an easy message. Test on the four that are actually hard.

The saying-no test is the most diagnostic of the four. If your "no" comes back as a soft maybe, the problem is not the voice fields at all — it is that you have not written down what you turn down.

Change one field at a time between runs. Changing three at once means you cannot tell which one fixed it.

What are the best practices and common mistakes?

Do: write like a briefing, not a brochure · name what you turn down · keep the banned list short and real · use real replies as examples · re-run the four tests after any significant edit.

Do not:

• Stack adjectives. "Professional yet friendly, authoritative but approachable" cancels out.
• Paste your About page. It was written for a different job and it teaches brochure English.
• Use polished marketing copy as your examples. Use the messy real ones.
• Write thirty banned words. The list stops being something you can hold in your head.
• Configure once and never revisit. The first version is a guess; the version after you have edited ten drafts is the real one.

What are the limits?

It does not learn by watching you. There is no silent adaptation and nothing you write trains any model. The configuration is what you typed — which makes it inspectable and editable, and also means a thin configuration stays thin until you improve it.

Voice is not accuracy. A draft can be perfectly in your voice and still claim something you do not offer. That is a different setting, covered in the guardrails guide.

It cannot supply the specific detail. It has your context, not your experience. The sentence that could only have come from having done the work is still yours to add, and it is usually what makes a reply worth reading — see the in-page reply workflow for where that edit fits.

Structural tells survive. Even a well-configured draft runs slightly tidier than human writing. Editing one sentence before sending fixes it, and you were going to read it anyway.

Your next step

Open a blank note and write the two fields that matter, in the concrete register above: what you actually sell including what you turn down, and ten words you would never type. Six minutes.

Then run the saying-no test. Take the last message you had to decline and generate a reply. If it says no clearly and still sounds like you, your setup is working. If it comes back as a soft maybe, you have found the field to fix.

See the documentation for the exact, step-by-step setup instructions.