The Voice File

You have deleted “I hope this finds you well” forty times this year

You asked for a client email about a delayed delivery. It opened with “I hope this finds you well” and closed with “Please don’t hesitate to reach out.” You deleted both lines, then rewrote the middle, because it described your work the way a competitor’s website would. Eleven minutes later you had an email you could have written yourself in eight.

And it’s the same with the LinkedIn post, the proposal intro, the bio a podcast host needed by Friday. The tool is fast. You’re the slow part, because you’re the one turning generic writing into your writing, by hand, every single time.

I don’t think a better prompt fixes this. What fixes it is a file.

Generic writing is the correct answer to a vague question

If you ask a model for “an email to a client about a delay” and give it nothing else, it hands back the middle of everything it’s ever seen on that subject. The safest structure. The average of thousands of delay emails, most of them written by someone in a hurry who didn’t much care.

You asked a general question and got the general answer. What you actually wanted was the version one person writes to one client, and that gap is exactly what you keep closing by hand.

A style guide is written for humans. A voice file is written for machines.

If you have a style guide, it probably says things like “warm but professional”, “confident, never arrogant”, “approachable”. A person reads that, nods, and carries on writing the way they always have, because the adjectives were never really doing the work. Their judgement was.

A machine doesn’t have that judgement to fall back on. “Warm” describes a million sentences, so it picks the most common ones. “Write in a friendly, professional tone” changes almost nothing, because that’s where it was heading anyway.

What does work is examples and rules. Your exact words in, so your kind of words come out. Real sentences of yours next to the flat version of the same point. Numbers instead of adjectives. That’s what a voice file is: a page or two of plain text, saved into the instructions of whichever AI tool you use.

A voice file has seven parts

1. Who you’re talking to, specifically. Not “small business owners”. Something more like: a physio who owns two clinics, does the roster on Sunday night, and tried ChatGPT for captions earlier this year before quietly giving up. One person you could picture ringing.

2. What you sound like, described in ways a machine can follow. How long your sentences tend to be, and how much they vary. Whether you use contractions. First person or second. “Mixes long explanatory sentences with the occasional short one to land a point, uses contractions, writes in first person” is something a model can act on. “Punchy” isn’t.

3. Your own vocabulary. Twenty words and phrases you use that a generic writer wouldn’t. Then twenty common business words that don’t appear anywhere in your published work. That second list does more than you’d expect.

4. The banned list. Keep it separate from vocabulary, because it’s absolute rather than a preference. Mine includes unlock, elevate and game changer. Yours should have at least three words you’re tired of seeing in your own industry.

5. Five real before-and-after pairs. If you only have time for one section, make it this one. Take a sentence you’ve published, and beside it write the flat version of the same point. One good pair teaches more than a paragraph of description ever will.

6. Structural habits. How you open and how you close. Whether your headings are labels or statements. Where your main point usually sits, in the first sentence or the last.

7. The things only you say. Named ideas, the analogy you always reach for, the phrase clients quote back to you. Anything you invented goes in here with its definition, or the model will paraphrase it into something worse.

Pull the file out of writing you’ve already published

The good news is you’ve already produced the evidence. Gather eight to twelve pieces of your own writing: newsletters, a page of your website, that long comment you wrote when something annoyed you. Then run this:

You are building a voice file for me: a reference document another AI will use
to write in my voice.

Below are samples I published. All mine, none edited by anyone else.

Read them, then produce a voice file with the seven sections below. Quote my
own words wherever you can. Do not describe my writing with an adjective
unless you immediately follow it with a sentence from the samples that proves it.

1. READER. Who am I writing to? Describe one person: their job, what annoyed
   them this week, what they already know, what they would never say out loud.
2. MECHANICS. Average sentence length. Shortest and longest sentence found.
   Contractions, yes or no. First or second person. Fragments, yes or no.
   Paragraph length in sentences. Give numbers, not impressions.
3. VOCABULARY. Twenty words or phrases I use repeatedly that a generic business
   writer would not. Then twenty common business words absent from these samples.
4. BANNED. Words that would be obviously wrong in my voice, inferred from what
   is absent.
5. EXAMPLES. Five pairs. Take a real sentence from my samples (label it AFTER)
   and write the flat, generic version of the same point (label it BEFORE).
6. STRUCTURE. How do I open a piece? How do I close one? Headings, lists, bold:
   which, and how often? Does my main point land in the first sentence or last?
7. SIGNATURE. Anything only I say. Named concepts, recurring analogies, phrases
   that repeat across samples.

Then list three things you are unsure about and would need more samples to
settle.

Output plain markdown I can paste into another tool. No preamble.

SAMPLES:
[paste everything here]

Then read what comes back with a pen in your hand. It will get two or three things wrong, often a habit you dropped a while ago. Fix those yourself. The file is yours, not the model’s.

Same model, same question, different writer

The request: the opening of an email telling a client their project is two weeks late.

Without a voice file:

I hope this finds you well. I wanted to reach out with a brief update regarding the status of your project. Due to unforeseen circumstances, we are anticipating a slight adjustment to the timeline.

With the voice file attached:

Your project is running two weeks late. The new date is 14 October. Here’s what happened and here’s what I’m doing about it.

Nothing changed except the file. The second version was written for someone the model had, in a sense, already met.

The file should run without you pasting it

Most of the big assistants now let you save standing instructions somewhere, so they apply to every chat without you having to remember. What each tool calls that changes often, so look for the capability rather than a feature name.

If your tool has	Put the voice file
Projects, or custom instructions that apply to every chat	In those instructions. If you use projects, one project per kind of work
Saved skills, gems or agents	In that assistant’s instructions, or as a skill it loads without being asked
An instructions file it reads before it starts (common with coding agents)	In that file, so it applies unprompted
An API you’re building on	In the system prompt
No saved instructions at all	Pasted at the top of the chat, from one saved snippet

Check where yours keeps standing instructions before you assume it can’t. Whichever row you land on, keep one master copy and update every tool from it. Six slightly different versions across six tools is honestly worse than none, because you lose track of which one wrote the draft you didn’t like.

When not to bother

If you write less than a few hundred words a month with AI, I’d skip it. Building the file properly takes about ninety minutes, and it won’t pay that back.

Skip it where house style matters more than voice: a compliance notice, a safety procedure, instructions for a piece of equipment. Flat is the right choice there.

Skip it if you haven’t published enough to work from. If you can’t find eight pieces, the model starts inventing a voice for you and hands it back very confidently.

And don’t force one file onto three different writers. You’ll end up with a house average, which is the thing you were trying to get away from. Give each person their own, and share only the banned list.

What you’ve got now

The seven parts of a voice file, and a prompt that pulls them out of your own writing. Once the file is saved somewhere it runs without you, the next thing you ask for should come back needing a couple of edits instead of twenty.

Tonight, copy eight pieces of your own writing into one document. Run the prompt over them before you open your inbox tomorrow.