Step-By-Step: Here's The AI System I Build To Help Business Owners Write Their Newsletters

I’ve been helping business owners run their newsletters for a couple years now, and up until a few months ago, it wasn’t really worth using AI seriously in that process.

AI tools struggled to absorb someone’s voice based on limited sample data, and even if they got close, they couldn’t really structure pieces in a compelling way. For a long time, it was faster for me to write client emails manually than it was to draft with AI, then fix and tune the piece afterward.

That’s changed.

The tools are getting better and better at producing useful text, and over the last six months, more of my client work has shifted away from writing for them, and toward helping them build the AI engine that allows them to write, edit, publish, and promote great work themselves.

The system looks something like this (below), and where my job used to involve owning every step of this chain for clients, increasingly, AI is taking over the right-side of this diagram and my time is shifting left – toward building the system, and helping develop ideas for stories.

In this piece, I’ll walk you through how I set something like this up so you can build and run it yourself.

(By the way, the piece you’re reading now was fully human-written, because that’s still worthwhile sometimes, but that’s a conversation for a different day).

We’ll spend most of our time on Step No. 2, because the principles there extend to the others.

This system is designed to make it easy for you to capture your raw ideas (either voice or text), and turn those into well-structured newsletters that actually sound like you, and capture the point you were trying to make.

To do that, here are the tools / resources I recommend:

Finally, I’ve got a Google Drive here with some sample skills and cheatsheets you can use if you want to speed run this.

This is where we’ll spend the bulk of the article, because the concepts here transfer elsewhere too.

We’re going to build a Claude Skill, that understands your business, and drafts newsletters in your voice, in the structure you like.

I’m using the term “draft” purposefully. You’ll still want to review/edit the output to fine-tune it for your voice.

But the system I’m about to show you is very good at getting close to your voice, and in my most recent client interactions, there tend to be only one or two small edits.

Simply put, Claude Skills are pre-packaged instruction books that tell Claude how to do a certain thing, so that it performs the same way and produces predictable results, over and over.

You only really need to know two things about them to get started:

We’re going to start by giving Claude some context, and using it to draft a few key documents that describe your voice, audience, and writing style. Then, with a single request, Claude will package all of that up for us into a custom skill that allows it to draft your emails reliably.

To build our skill, the first thing we’re going to do is upload a few documents that give Claude background on your business, and the way you either speak or write.

When I do this with clients, it starts with this Onboarding Q&A. The document has questions about their business, clients, and opinions on the industry, and the first thing they do is fill this out.

Then we record a conversation talking through their answers in more detail, and both of those assets (the written Q&A and our conversation transcript) get uploaded to a new Claude project where we’ll build their skills together.

If you’re doing this alone, make a copy of the doc and fill in your answers in writing first (bullets will do), then sit down and record yourself talking through your answers in more detail, and transcribe it (WisprFlow is great for this. iPhone also automatically transcribes voice memos).

The reason the transcript is important is that it captures how you speak naturally, which will help Claude better mimic your voice later.

I’ve been fascinated by this question recently, as I continue refining my process for teaching Claude to write like a given client.

What I’ve found so far is that your voice is definitely more than the substance of your opinions, and it’s more than the official industry terms you prefer to use.

It’s the filler words you use without realizing it, the transitions you use, or the ways you start sentences. It’s your use (or lack thereof) of emojis, or analogies, or movie references, or profanity. It’s your triumphs and heartbreaks. The things you learned to show the world, and the things you learned to conceal.

The more I explore it, the more I fall in love with the things that make people unique. Because at the base of it all, your voice is this sort of irreducible compilation of all your experiences – it’s everything that sticks to you as you move through life; A fingerprint of the tiny, unique set of events and experiences of the universe that we call, “you.”

It probably can’t ever be perfectly replicated by a computer, but it is fun to step back and analyze it every now and then. Kind of like a personality test.

And the best way is to have a few examples that were made when you weren’t thinking about it.

So, if I have 2-3 sample pieces of writing from the client, I’ll add those to the project files too. So that by the time the first stage is complete, Claude has access to…

The next thing we’ll do is have Claude summarize its own understanding of your audience, voice, and writing style by prompting it to create two files:

Here are two actual prompts that I’ve used recently to produce these documents. As you can see, they’re all similar, and very simple.

The prompt above may look complex, but all I’m doing is giving it several samples from the client, and telling it which are our favorites. You don’t have to go this deep. But I find that the results get better when you give Claude a little more guidance about what to focus on.

This exercise will produce two key documents: Your voice guide and audience guide.

The first thing you want to do is read through, and make sure you agree with what’s currently in there.

We’ll be adding more later, so it doesn’t have to be comprehensive yet. Just make sure it hasn’t created any rules you disagree with.

If it has, you can edit those by simply talking in the chat (example below). It’s tough to tell, because I was being pretty loose with the structure of my request, but the bottom part (in red) is a line from the voice guide that I’m referencing, and the top part (in green) is my feedback/edit.

This was enough for Claude to correctly update the guide the way I wanted.

After editing, add the voice guide and audience guide files to your project so Claude can reference them on any future tasks.

At this point, we’re about 80% there.

In the next few steps, we’ll draft a sample newsletter, then revise it, track what changed, then feed that info back to Claude, to update and improve these core documents.

This refining process is how we fine-tune its ability to write in our unique voice.

Now that Claude has a few samples of how you like to write, and clear voice- and audience-guides, it’s in pretty good shape to start generating drafts for you.

So let’s talk for a second about how that’s done.

Currently, my favorite way to use Claude to draft an email for a client is to…

The live interview allows them to explore an idea out loud, which is a lot easier for most people than writing. Often these conversations unearth much better insights too, because I’m there, as a journalist, and can help pull threads I think their readers would want to know more about.

So we talk for a while, I mentally bookmark the best parts of the conversation, then upload the transcript to our Claude project with a prompt that looks something like this:

The key here is in that tiny bit of direction I gave – telling it which part of the transcript to focus on.

A recorded conversation can be pretty wide-ranging. But Claude responds quite well to even a small amount of direction. This, plus the voice & audience guides, will often produce a good first draft.

But it won’t be perfect… And that’s the point.

We’re going to have to revise it. And every single edit is a signal of something important that Claude doesn’t know about our writing style yet. So after we revise, we’ll track the changes, and feed them back to Claude so it can update its understanding of our style.

Here’s how…

1. Drop the draft in a Google Doc: Edit until you’re completely happy with how it sounds and would be okay sending it to your most important client (that’s a good litmus test for whether you’re hitting the tone you want).

2. Use Version History To Review All Edits: Look for changes that are significant, and represent an important rule that you want Claude to know about how you talk to your audience. Examples might be…

Looking back, some of your edits will not be substantial. Feel free to ignore those. But take note of any meaningful change that you want to instruct Claude on.

3. Document The Before/After/Reason: Start a fresh tab in your Google Doc, and for each important edit, record what the initial text was, what you changed it to, and a brief explanation of why. Continuing with the example above, your note would look something like this:

4. Feed Your Notes Back To Claude: Give it the final, edited version of your draft, along with the full page of notes on any important changes. Then tell it to use those notes to update your voice and audience guides (here again, I don’t think my phrasing is particularly important - as long as you communicate the gist of what you’re trying to do)

You should have:

You can go through this process of giving feedback to Claude as many times as you’d like. But I find that revising and annotating 1-2 newsletter drafts is good.

Don’t get stuck here too long.

Remember, we’re building context that we can ultimately package into a Claude Skill. While your project will be pretty good at drafting emails, we want to push through and finish the skill so we get consistent execution, over and over.

To do that, there are two more steps…

Similar to the way you created the Voice- and Audience-Guides, you’re now going to ask Claude to review all of your context up to this point (your revised drafts, notes, and updated voice and audience docs) and draft a Newsletter Production Guide.

Tell it that the goal is to to intake rough notes or voice transcripts and turn them into final newsletters like the samples you’ve uploaded.

The main difference between this and the others is that it will be more procedural – walking through how to turn raw input into a finished draft. That means it’ll spend more time focused on the structure of your emails and how to replicate them:

Here’s a sample of the kind of thing you want to see inside:

Once again, review the guide to make sure you agree with all the rules in there. Have Claude update anything you don’t like

The last thing you’ll do is ask Claude to turn all of this context into a skill for newsletter production.

The Newsletter Production Guide will serve as the core of the skill – the instruction set that Claude executes on over and over to produce predictable outcomes – and your voice guide, audience guide, transcripts, and finished emails can be rolled into the reference folder as useful context.

With Claude, this is super easy.

It’s got a “Skill Creator Skill” that allows it to make new skills for you based on basic conversations.

When it’s done, it’ll output a file like this. You can activate it in your account by clicking “Save Skill” or you can download it, and share it with anyone else on your team who may need to draft newsletters in your voice.

The skill can be updated any time, since all it is is a collection of markdown files. The only thing you need to keep in mind is if you have employees using it, they’ll need to install the latest version any time you update.

Now, you’ve got a skill, trained on your voice, audience, and newsletter style, and designed to take in a voice transcript and turn it into a structured newsletter draft for you to review.

Let me give you three ways to think about this…

1. Messy Voice Note: Claude can actually work quite well with relatively unstructured voice notes. So if you have an idea while you’re out on a walk, grab your phone and start recording. I like Wispr Flow. But Voice Memos on iPhone automatically transcribes too. Talk for a few minutes and don’t worry about structure or sounding refined. Just get the idea out.

2. Conversation with Claude: Claude’s voice mode is also quite good, and automatically transcribes everything into text. If you want to develop an idea, have it interview you and ask you questions about the concept. You can go back and forth as long as you like, then have it feed the entire conversation to your new Newsletter Skill for a draft.

3. Human Conversation: This is one of my favorite, and ends up being a fair amount of the work I do with longer-term clients. We meet on a regular basis, I interview them like a journalist to find out what’s on their mind – ideas, insights, latest developments in the business that readers should know about – then the transcript for that becomes the input for the newsletter. The benefit here is mainly two things: Accountability (for busy founders, it helps to have a meeting on the calendar to think about content) and thought partner. People aren’t always aware of which parts of their expertise are interesting to other people. I stand in as the reader, and have an ear for which parts of an idea they should dive deeper on, or unpack more fully.

The beauty of al three of these is that they allow you to get our of your head about writing and focus on the idea you’re exploring instead.

The truth? You could probably publish the un-edited transcript and it’d be better and more interesting than most thoughtless AI pieces. But that transcript, plus this system, offers the best of both worlds – your best and most interesting, unfiltered thoughts on an idea, plus the structure to make them clear.

Used well, this tool will do a good job replicating your voice (I’d say A- to A on any first pass). I feel confident saying that because I’ve watched over and over with clients as the number of suggested edits on drafts falls to near-zero as we go through the process I outlined above.

The skill works.

But there’s a psychological component to this process that I think is important to factor in.

Because, at the end of the day, we’re not just trying to match how someone sounds, but really, their perception of how they sound, and that’s a very interesting game to play.

I had one client who spoke English fluently, but whose first language was Italian. Importantly, he spends a lot of time thinking in Italian.

We went through the first part of the process above, and when Claude output an early draft in English, the client said it was excellent. I was surprised because, having spent a few hours talking with him live, I could see changes I wanted to make to more closely match his tone. But he insisted it was a near-perfect match.

His clients are Italian though, and he was thinking of publishing in Italian. And so, we had Claude translate the piece. Immediately, he saw changes he would make.

Obviously, the translation plays a role here. But it was interesting to me to see his sense of his own voice was much more acute in Italian.

This has become one of the things that fascinates me about this work – understanding how people perceive their own voice, and the role that understanding plays in our work.

Technically, you could stop here, and you’d have a very useful little AI assistant to take your rough ideas and transform them into newsletter drafts. On its own, that’s useful.

But we can do more, and with my clients, I do.

Because for them, the newsletter is part of a broader strategy to grow their influence, and their company. These are people who play at the top of their field. The stakes are high, and their calendar is busy, so we add a few additional tools to make this useful.

Specifically, we…

I’ll run you through each, but these will be faster since they are basically slight tweaks on the process we just went through.

The idea here is simple, and powerful: AI is good at taking on the persona of a given subject matter expert. So we create a single “Review Panel” skill with 3-7 “personas” on it, each one representing a different type of reader, and designed to battle-test the concepts in our draft, and reflect ideas and outside perspectives back at us.

The goal here is to surface weaknesses in our argument so we can address them before we publish.

Originally, this idea came from a client of mine who writes in the medical/longevity space. His goal was to develop a panel of specialists who could gut-check his writing to make sure he wasn’t over-stepping the bounds of what can be said with confidence. (Also, there was an “attorney” on the board, to flag anything that might create legal liability)

The process for creating it involved some back-and-forth, but when we were done, I had Claude summarize it into a single, general prompt you can use:

So far, most of my experiments with this step have involved fictitious experts that are created specifically for the task. But you could imagine a version where you hand-pick voices from your industry that you respect, and train Claude on a bunch of their material in an attempt to assemble a kind of Dream Team of editors.

Tim Ferriss recently said that he’s experimenting with this as a primary use of AI in his own writing practice, hand-picking the editors of his favorite authors to offer feedback on his work.

There are a lot of reasons I recommend Beehiiv over every other newsletter tool right now. Great features, killer team, shipping fast, etc…

Case in point: They recently released an MCP connection, so you can interact with your account from Claude.

As of this writing, it only has read-access.

But V2 is coming soon, and will have write-access, which means you could envision a world where you sit down, have a conversation with Claude (in writing or voice), explore an idea, then…

In time, a lot of the handoff from one step to the next can be automated, so that more and more of your work is focused on the front-end of the chain – coming up with interesting ideas, and exploring them in voice or writing.

The last thing we set up is a LinkedIn skill designed to take finished newsletters and generate posts you can use to either hype the newsletter before it goes out (the “pre-newsletter C2A”) or summarize the email after it goes out, and encourage others to sign up for more (the “post-newsletter C2A”).

There’s a generic version of the skill that you can snag here.

It’s programmed with the four basic emotions that tend to make content go viral on LinkedIn, and is designed to look at your finished draft, find angles that align with those emotions, and draft one potential post for each.

You can then pick your favorite, or pre-schedule all of them to run several weeks apart, thus re-prommoting some of your evergreen content over and over in unique ways.

This skill is curbed directly from Ryan Carr’s piece on this with one important alteration: There’s an instruction at the bottom of the skill.md file to check our references folder for voice samples, and when I work with clients, I update this by adding our voice-guide and sample content to that file.

The reason it’s packaged as a skill is twofold (and helps show why skills are so useful):

Generally, this skill is designed to re-hash and promote a newsletter after it already sent (e.g., “Here’s a great takeaway from my last newsletter, if you want more stuff like this, sign up here”).

That’s the main thing it’s good at.

But with a slight change to the instructions, it' does a decent job with pre-send C2As as well (though I will probably go in and update this more specifically soon).

From there, you can schedule/post as you see fit.

At this point you have a few foundational skills now active in your Claude account, designed to make drafting, publishing, and promoting easier:

You can move through this chain manually, or automate it using chron jobs, and Claude’s onboard agentic features.

The ability to move content through this pipeline is getting better, so that more and more of your time is spent at the front-end: Coming up with the interesting ideas, and exploring them in whatever format feels natural for you.

The concept of having AI write for you is still a controversial one, especially in writing circles. Like all technology, there’s a good and bad way to use it.

To be honest, it’s something I’m still exploring my own thinking on. But so far, the thing I like about this setup in particular is that it’s empowering more and more people to get their ideas out into the world.

Most people are not writers. But they still have interesting experiences and insights that are worth sharing and learning from. This system makes that possible, in a way that keeps the idea pure.

That said, I have a few rules for how best to use this…

1. Don’t use AI for any skill you want to maintain: I don’t use AI in any of my long-form writing, because I find the writing itself useful and think it’s worth doing for its own sake.

2. Don’t let AI add context that’s not there: One of the lines I include in our newsletter skill is that the AI cannot add ideas or concepts that aren’t in the original voice note it’s given. We’re not looking for help building an argument, we’re just looking for help formatting it once we’ve already made it.

The reason I feel fine advocating this form of AI assistance is because the important part – thinking through the ideas – is still on you.

At this pointI’m focused on how to tie these steps together and effectively automate them so that more and more time is spent exclusively on that first step.

More on that to come…