I dictated 250,000 words last year. When I split them by where the text was going, nearly three-quarters had gone into AI tools: instructions to Claude, Codex and ChatGPT. The rest went to people, as Slack messages, client emails, LinkedIn replies and notes to friends.

My voice file, the document every AI tool reads before writing as me, had been built from a mix of both. So it had learned a good chunk of how I talk to machines and was using that to write to humans.

The file learned from the wrong writing

The clearest example was one word. My old file listed “let’s” as a signature phrase, and it’s true that I say it constantly. But when I counted, it turned up 64 times per 10,000 words in my prompts to AI and about 4 times in my Slack and LinkedIn messages. “Let’s ship it” is how I direct an agent. It isn’t how I write to a colleague.

It went further than vocabulary. With machines I’m blunt: “this is broken”, “that’s not very good”. With people I’m steadily polite, I say “please” and “thank you so much” a lot, and I challenge a colleague with “I’m curious where these came from” rather than a verdict. My old file told every AI to be “direct over polite”, which is right for a code review and wrong for nearly everything else I send.

If you’ve been building a voice file from whatever writing is closest to hand, there’s a fair chance yours has the same problem. More and more of what founders write each day is prompts, and prompts are a register of their own.

Why investors are where it matters most

An email to an investor is a small sample of how you think. It’s also the email a founder is most tempted to hand to AI, because the stakes feel high and the blank page feels heavier than usual.

That’s exactly where generic AI writing costs the most. My read, from being in the room on raises as the CTO, is that investors now see a lot of AI-shaped writing and have become quick at spotting it. The smoothness itself isn’t the real risk, though. The real risk is what AI adds when nobody’s told it not to: a confident adjective where a number should be, a “we” that’s actually one person, a claim that sounds great and wouldn’t survive due diligence. In a sales email that’s a bit embarrassing. In an investor update it’s something you might have to explain later.

A good voice file fixes both at once. The output sounds like you, and it carries your rules about what you’d never claim.

Here’s the process I used, in the order I’d do it again.

1. Collect by who’s reading

Gather real writing and sort it by the reader, not the format: team messages, client or customer email, investor email, public posts, and personal messages if you want those covered too. The best source I found was my dictation history, because every entry records which app I was talking into. Sent mail and Slack exports work well too.

Then set aside everything you wrote to an AI tool. Don’t use it as example writing. You’ll come back to one part of it in step 4.

2. Measure each register

Run each pile through something that counts rather than guesses: average sentence length, how many sentences are short, contractions per 100 words, questions, and how often your softeners show up (“just”, “maybe”, “I wonder”, “hopefully”). My free voice.md generator does the core of this in your browser, and I explained why measuring beats adjectives in an earlier post.

The numbers matter most where they differ between registers. My LinkedIn replies average under 11 words a sentence and my journal averages 17. A single number for “my voice” hides the thing the model most needs to know.

Look for what’s missing too. Across a year of messages I never once wrote “no worries”, “cheers” or “circle back”. Absence is as useful to a model as presence, because a “never” list stops the most common tells.

3. Mine what you change before sending

This is the richest source most people ignore. Take drafts you edited before sending, whether AI wrote them or you dictated them, and compare the before and after.

In my 113 genuine edits, the pattern was consistent. I cut the soft closing line (“I’m here if you need me”, “There’s good evidence around this”). I turned “Please can you ensure…?” into “Please ensure…”. I swapped vague words for exact ones and barely changed the length. Each of those became a rule, and the best ones became before-and-after examples in the file, because models copy an example far more reliably than they follow a rule.

4. Mine your complaints about AI drafts

Here’s where the writing you set aside in step 1 earns its keep. Every time you’ve told an AI “no, that’s not right”, you’ve written down a piece of your taste. I found 622 of those, and they grouped more cleanly than I expected.

My biggest complaint wasn’t “this sounds like AI”. It was “there’s no story here”, which came up around 57 times. Next was “this makes too many assumptions and reads cold”, meaning the draft was written for someone who’d been in the meeting. Then “generic”, then hype, then “too harsh, make it more collaborative”. Each became a rule in the file: give a longer piece an arc, write for someone who wasn’t in the room, and never invent a fact.

For investor writing, that second rule matters enormously. An update that says “revenue picked up after the change we agreed” means nothing to someone who wasn’t there. Say which change, when, and what the number did.

5. Write an investor register

Once the general file exists, add a short section for investors. Mine would start something like this:

Investor register · add to your voice file
<investors>
- Lead with the ask or the number. One line of context before it, no more.
- Every figure must be one I could show in due diligence. If the source doesn't have it, ask me. Never estimate or round up.
- "I" when it's me. "We" only when there's a team behind it.
- No hype words, no dramatic closing line, no "we're uniquely positioned".
- Updates: what happened, what the number did, what's next, what I need.
- Follow-ups: a light nudge first, then name the gap, then ask directly. Warm every time, never pushy.
</investors>

Keep it short. Your general voice rules still apply underneath, so this section only holds what’s different when the reader is deciding whether to give you money.

6. Test it blind

Don’t trust a voice file because it reads well. Test it. I took six real situations from my history, had one AI draft each message with my old file and another with the new one, shuffled the pairs, and picked the one that sounded like me without knowing which was which. The new file won five of six.

The one it lost was useful. The new draft opened a request with pleasantries and a paragraph of context before getting to the ask, and it used my favourite phrase (“I’m curious if…”) in four of the six messages. That’s the other failure to watch for: a voice file that makes AI do an impression of you. I added a rule to lead with the ask, and a short list of signs a draft is overdoing it.

The file is never finished. Every time I correct a draft now, the correction becomes a new example, so the file gets better and the drafts need less fixing. It’s the first thing I set up when I install an AI Brain for a founder, because everything else the AI writes for you sits on top of it. If you’re about to start a raise and want the investor register built properly, book a call and we can do it together.

Share: