A few people asked for my voice doc after I posted about it on LinkedIn, so I shared it. The voice doc is one page, though, and I think the reason it works is the rest of how my Notion is set up. So this post is about that.
Where the voice doc came from
I got really annoyed one day and spent like an hour yelling at Claude about its writing. Every draft had the same stuff in it: em dashes, the neat quotable line at the end of a paragraph, words I’d never say out loud. I’d fix it in the chat, and the next day I’d be fixing the same thing again in a brand new chat.
So we built a document instead. It pulls from a dozen of my real client and prospect call transcripts, because that’s the closest thing to how I actually talk, plus some older writing of mine. It has the traits that make something sound like me, the words and phrases I actually use, a few before and after rewrites, and a long list of things I never want to see again.
The list is the part that does the most work. Every time a draft comes back with something that bugs me, it goes on the list and I stop having to fix it in the chat. Some of the entries are really specific. Write “no one” and “everyone” instead of “nobody” and “everybody.” Never use “honestly.” Don’t tell the reader something is quick or cheap, just say what it is and let them decide. And no made-up details bolted onto the end of a sentence to make it sound vivid, which I had to add because the AI kept inventing little scenes that never happened.
There’s also a checklist at the bottom that the AI runs through before it hands anything back. My favorite one is to read the last clause of every sentence on its own, and if deleting it loses no information, delete it.
The Context Store
The voice doc is one entry in something bigger that I call my Context Store. It’s a Notion database where every row is one fact about me, my clients, or how I want something done. Each entry has a category, how important it is, how long it’s likely to stay true, and a usage guidance field that tells the AI when to pull it in.
My instructions page for Notion AI tells it to search the Context Store before it responds to anything that isn’t trivial. Claude gets the same instruction when it’s working through the Notion connector (an MCP, which is just a standard way for an AI tool to read and write inside another app). When something gets marked Critical, its core rule also gets pulled up into the instructions page itself, so it applies to every conversation without a search.
The other half is that the agents write to it. I’ve told all of them that if they figure out something new that might be useful later, they should log it. ChatGPT, Claude, Notion AI and Gumloop can all add entries, and each entry gets stamped with which tool wrote it. So when two entries conflict, I can see where each one came from, and so can they.
A few examples of what’s in there:
- A LinkedIn post I rewrote myself, word for word, with notes on what the AI version got wrong. It added a joke I didn’t need, explained the pattern instead of just giving my opinion, and ended on a tidy piece of advice. I cut all of it.
- How my HubSpot is set up: what each deal stage means, which properties always get filled in, and that notes have to use HTML formatting because HubSpot shows markdown as raw characters.
- Rules for building on Val Town that every agent has to read before it touches anything there, even a read-only call.
- Where my meeting transcripts live, and that “look at the transcript” means the full transcript, not the AI summary.
Why Notion instead of text files
A lot of the advice right now is to build up skills and memory files, which are text files the AI reads before it does anything. I think that’s useful, but no one’s opening those files. They’re for the AI and only for the AI.
In Notion it’s the same information, but it’s a page or a database I can open and work out of. I can build a board on it, sort it, or fix an entry when it’s wrong. And because I’m in Notion all day anyway, the context builds up while I’m doing normal work. If you’re keeping text files for the AI, you have to go out of your way to make them.
I also don’t think a pile of markdown files holds up once a company gets big. At some point there’s too much for the AI to read before every action, and you need structured data in there, tags and categories it can filter on, alongside the looser notes. That’s a big part of why the Context Store is a database and not one long page.
It doesn’t have to be Notion. You can do this with Obsidian, a Google Doc, or a plain text file. I just think Notion’s the easiest to use for both the people and the AI.
Instructions that get a little better every time
The thing I’m most interested in right now is making whatever I build grow a little every day. If you’ve read Atomic Habits, it’s the 1% idea, and it’s really easy to do with AI.
I start with a pretty rough one-off build. If I can tell I’m going to use it again, I make it repeatable. Then every time I use it, I add one small thing, or I build an agent that does that part for me. A lot of the time the thing getting updated is a Notion page or database with a growing set of instructions, and the agent just points back to it and reads it before it runs.
My client work runs this way now. Every recurring task I have, one for each client plus marketing and admin, reads the same Work run rules page in Notion at the start of every run. That page says what to load first, what it’s never allowed to do (it drafts, it never sends anything), and how to log what it did. If I want all of them to behave differently, I edit that one page and leave the task prompts alone.
The voice doc works the same way. It’s a lot longer now than when I started it, and almost every addition came from a draft that annoyed me.
Save the chat as a page
This is the habit that got the whole thing going for me. After a useful conversation with AI, I tell it to save the conversation as a page in the right place. “Add a page to Marketing about this chat.” Now the next conversation about marketing starts with that context already there.
Same idea when I start collecting information on something new. My first move is to make a database for it, even if I don’t plan to read it myself. Tagged, structured data is a lot easier for AI to pull from than a wall of text.
I don’t design those databases up front either. My companies database started as a random list of notes about companies I’d talked to. When it got messy, I told Notion AI to turn it into a database and tag everything, and the first pass was pretty close to what I wanted.
If you’re starting from nothing, I’d use PARA: Projects, Areas, Resources and Archive, plus an Inbox for anything you don’t know where to put yet. It comes from the book Building a Second Brain. Dump everything in and organize it later. I used to care a lot about putting things in exactly the right spot. Now I mostly just ask the AI to go find it.
Meeting notes do a lot of the work
All of my call transcripts land in Notion, recorded by Notion’s note taker, which runs on my machine instead of sending a bot into the call. Every client and prospect call ends up in one searchable place, linked to the company and the person.
That’s where the voice doc came from in the first place, since it’s built off real calls. The bigger everyday use is what I do after a call. I’ll chat with the meeting note and tell it to add the follow-ups to my to-dos with the context in the page body, so each to-do carries the conversation with it. I’ve built proposals almost entirely out of a transcript plus an old proposal, and then edited from there.
A nicer screen on top of it
Once I had 25 or 30 Notion databases I was using every day, jumping between them got annoying. So I had Claude build a dashboard over those databases and my HubSpot, where every card links back to the Notion page. I call it my Client Command Center, and it’s also where I kick off those client work runs.
I’ve also started building a lot of throwaway dashboards, ones I don’t plan to come back to. My agents use Notion as a scratchpad, and when I need to see something, I have Claude build a quick view of it for that day. The next time it gets built a little differently, because the instructions it reads have changed since.
Doing this for a team
The most common problem I see right now is a company that buys Claude for everyone, and then the power users keep using their personal accounts anyway. I don’t blame them. Their personal account has years of their context in it and the company one has nothing.
You can’t just mandate it. You have to make the company version more useful than the personal one. One of the first things I build for a company is an org-wide instruction that watches for this. When something comes up in a chat that might be useful to the rest of the team, the AI drafts a Notion entry and asks the person to approve it before it goes into the shared space. That alone slowly builds things up, and the people who know AI best end up feeding the context everyone else uses.
Where I’d start
If you want to try this, start with my voice doc. I took out the client quotes and stories so it works as a template, and there’s a short section at the top on how to swap in your own lines. Yours will probably look a lot different in a month. Put it somewhere your AI tools read before they write anything, and every time a draft bugs you, add the fix to the doc instead of the chat.
Get the voice doc. It’s the same file my AI tools read before they write anything for me, as a markdown file you can drop into Notion, a Google Doc, or a Claude project.
Thanks. Here it is.
Download kevin-stout-voice-and-style-guide.md
Once that’s working, do the same thing for whatever you explain to AI the most, whether that’s your CRM setup or how you like proposals laid out.