I Run My Substack Through Claude Now. Here's What That Looks Like.

Let me be upfront. When people say "use AI for your newsletter," I picture the same thing you do. Generic sentences. A tone that belongs to nobody. Notes that read like a press release from a company that does not exist.

I was not excited to try this.

But I did, and some of it surprised me. So here is what I actually do now, without making it sound better than it is.

What the setup actually is

StackBuddy has an MCP server. That means Claude can reach my real StackBuddy account inside a conversation.

I do not open a dashboard. I do not copy numbers into a prompt. I ask, and it pulls from my actual data.

That is the whole trick, and it is less magical than it sounds. It is a connection, not an intelligence.

What I ask it

The first thing I did was ask which of my notes performed best. It had analysed 933 notes. My average sat at 0.9 reactions per note, with 83 restacks in total.

Those are not impressive numbers. But they are mine, and seeing them laid out honestly told me something real about where I was starting from.

Then I asked what time my notes tend to do well. That question sounds simple and is miserable to answer by hand: you would be scrolling back through months of posts trying to spot a pattern. I got a straight answer, and I schedule differently now because of it.

Whether that change pays off long term, I genuinely do not know yet.

A starting point, not a draft

Here is where I want to be careful about how I describe this.

I asked for three notes in my voice on a topic I was working on. It had my past notes to work from. The drafts were not bad. They were also not good enough to post.

The rhythm was close but not right. The opinion was vague where mine would be specific. One of them said something I do not actually believe.

I edited all three heavily and rewrote one almost completely. The one that worked best kept maybe thirty percent of what came back.

A rough draft you can argue with beats a blank page. But if you are hoping to hand this off entirely, that is not what this is.

What it cannot know

This is the limit that matters most.

It does not know what happened to you this week. Not the thing that broke on Friday, not the call that went sideways, not the idea that arrived at 11pm and felt obvious in hindsight after three hours of confusion.

That material is usually your best. The model can help you shape it once you bring it in. It cannot invent it.

And a note that sounds like nobody is a note nobody restacks.

The analysis has a ceiling too

933 notes is a decent sample. For most writers it will be much smaller.

If you have published 40 notes, the patterns that surface are interesting but not reliable. This is not magic, it is arithmetic on whatever data exists. Small sample, noisy conclusions.

I try to keep that in mind when reading my own insights, including the flattering ones.

What a week looks like now

I open a conversation and ask what my best notes had in common. I ask for three sketches on a topic I am already thinking about. I react to the sketches, pick one, rewrite it, and add the part only I know. Then I schedule it for the time the data suggests.

The loop takes less time than it used to. But I am still in every sentence that goes out.

What I do not hand over

Anything that depends on my actual life or my actual opinion.

That is not a philosophical stance about AI, it is practical. What gets restacked is almost always the specific thing, not the general thing. "Here is what I learned building this" beats "here are some things people often notice." The second one is easy to generate. Only I can write the first.

I also want to push back on the "in your voice" framing, because it oversells. Past notes inform the drafts, and that helps: the output is closer to my style than a cold prompt would give. But voice is not only word choice. It is what you notice, what you care about, what you find funny, what you think is overrated. The model picks up surface patterns. You still supply the rest.

Is it worth it? For me, narrowly, yes. The scheduling and analytics save enough attention to pay for themselves. The drafting is useful the way a bad sketch on a whiteboard is useful: it gives you something to argue with.

If you want the same setup, the MCP server is on the Pro plan and takes a few minutes to connect. Just go in with realistic expectations. It is a useful tool inside a workflow you still run. It is not the writer.