3 prompts to give your Substack MCP before it analyzes your post data
Help your AI exclude unfinished data, make valid comparisons and check distribution before it gives you editorial advice.
A few weeks ago I wrote about setting up your Substack MCP like a publishing advisor — what kind of judgment you want it supporting and why reactive questions are too small to carry publication-level strategy.
That first prompt taught Claude or ChatGPT how to behave like a publishing advisor.
Yesterday, I found the next piece: teaching it how to decide whether the post-level data in front of it is actually ready to influence a publishing decision.
I asked Claude for a 90-day audit of The Publishing Spectrum.
It noticed that one post had brought in 41 new subscribers while a nearly identical post brought in three, and Claude had a theory: the difference lived in my editorial approach.
Study those two June posts, it said. Then build August around what the more successful one was doing. Its guidance was confident and built from real numbers. Every figure was accurate, live and mine.
It was also wrong.
And it was wrong in a specific way that I think a lot of us are about to run into.
The MCP has the data. It does not automatically have your method.
The Substack MCP has access to a remarkable amount of information about your publication.
It can see post performance, early unsubscribes, read-to-bottom engagement, traffic sources, subscriptions generated and much more.
Truly, I think the post-level data bank is one of the most useful things Substack has released this year (I wrote about it in April; see footnote1).
But put all of that information inside an AI conversation, and the post-level data can turn sideways real fast, my friends.
The MCP gives us access to the data. It does not automatically bring along the analytical rules you use to decide when a number is mature enough — or comparable enough — to influence an editorial decision.
So the AI reaches for what is nearest and builds something reasonable-sounding on top of it.
I do not see that as a flaw in the Substack MCP.
I think it is wise for Substack to hand the data, as unshaped as it can reasonably be shared, to its community of independent publishers. It is a very democratic way of allowing people to nurture the future of their own publications because we can and should be free to build our own guardrails.
But it does mean that anyone querying their data through the Substack MCP is missing a layer of discernment.
I think of that missing layer as publication intelligence: the discipline of deciding which data is mature, comparable and context-rich enough to influence a publishing decision. (Further reading: I was interviewed by Mediabistro earlier this year on publication intelligence. See footnote here.2)
Today’s post aims to give you a few starting points for building in that intelligence directly into your Substack MCP. (If you’re curious how I came up with these, feel free to read this footnote here.3)
Give these prompts to whichever AI tool you are using to query your Substack MCP (or your CSV files; those work fine, too) — before you ask it for publishing- or editorial-level guidance.
The three reliable-pattern rules
Before we get into the complete prompts, here is the framework:
Do not let a post under seven days old carry a trend.
Compare like posts against like posts—and state the size of the comparison set.
Rule out distribution differences before drawing conclusions about the writing.
The reliable-pattern test is simple:
Is the data mature? Is the comparison valid? And did distribution — not the writing — create the difference?
Prompt 1: Exclude posts under 7 days from trend analysis
This is the rule that broke Claude’s original analysis yesterday.
One of the two posts holding up its theory was only five days old, and I like posts to have more travel time before analyzing them.
Inside the SubSight framework, a post that young is considered inadmissible for trend analysis.
A five-day-old post can be interesting, sure. But it cannot carry a trend — especially one that might influence your editorial direction or how you nurture your audience next month.
Try this prompt
Before identifying a pattern or trend, exclude every post published fewer than seven full days ago.
Its opens, views, restacks and distribution may still be changing, so do not use it to support a broader editorial conclusion.
List each excluded post, its publication date and why it was excluded.
Prompt one is yours to work with, my friends.
And prompts two and three are for paid subscribers — including the rule that actually explained the difference between 41 new subscribers and three.
It turned out to have nothing to do with my writing at all.
Upgrade today to access the rest of this post and the complete copy-and-paste prompts.
Earlier this year I launched SubSight as a data-grounded publishing tool for my paid subscribers. I used publishing, marketing and production best practices along with my years inside Substack data audits to build out 150+ publishing rules for querying Substack data. If you’re interested in what SubSight can tell you about your writing and your next steps inside your publication, you can learn more here.