I Built a Substack Analytics System with Claude Cowork. The Data Changed How I Write.

290 notes, 42 articles, 323 subscribers analyzed: what Substack's dashboard won't show you (and why it changes everything).

TL;DR:

• Substack’s native dashboard hides more than it shows: Notes impressions only exist in the mobile app, and you can’t cross-reference growth with content. On a platform generating $510 million in annualized creator revenue, this is unacceptable.

• Using Claude Cowork, I reverse-engineered 6 private Substack API endpoints and built a complete analytics system: 290 notes, 42 articles, 587 comments, 323 subscribers analyzed. The result? Notes with attachments generate 78% fewer likes, and the top 20% of my articles produce 91% of total signups.

• The entire kit is open source on GitHub, fully parameterized, ready for any creator. No coding skills required.

You know the views. You know the open rate. But you know almost nothing.

You’ve been publishing for months. Maybe years. You open Substack’s dashboard, glance at the view count, check the open rate on your latest email, and feel like you’ve got a handle on things. But stop for a second: do you know which of your articles actually drive new subscribers? Do you know whether your Notes perform better with or without attachments? Do you know if your open rate is dropping because you’re losing audience or because you’re gaining new one?

No. You don’t. And it’s not your fault.

Substack reached 8.4 million paid subscribers in Q1 2026, up 68% year-over-year from 5 million in 2025. Annualized creator revenue surpassed $510 million. There are over 3 million active writers on the platform. Yet the native dashboard offers no subscriber segmentation, no behavioral tracking, no cross-analysis between content and growth, and no custom reporting. Notes impressions? Only visible in the mobile app, not on the web and not via any public API. The Growth Report launched in 2025 shows unique visitors, new subscribers, and new revenue, but it doesn’t tell you which article caused that spike.

Sure, SaaS tools try to fill the gap. WriteStack is a Chrome extension for scheduling and public Notes analytics, but it only accesses public data. Notes Studio Pro offers suggested timing and CSV export, but stays on the surface. NHagar’s Python package “substack_api” (version 1.0, March 2025) is a solid open-source project, but it too only touches public data. None of these tools access private data: subscribers, revenue, historical growth, cross-referenced Notes analytics.

The real problem is that you’re running an editorial business with hobbyist tools.

The number that made me stop everything

Here’s where the perspective shifts. When I started cross-referencing the data the dashboard wasn’t showing me, the first number that hit me was this: Notes with attachments (images, link previews, embeds) generate 78% fewer likes compared to short Notes without attachments. Not 10%. Not 20%. Seventy-eight percent.

This data point doesn’t exist in any SaaS tool. You won’t find it in WriteStack, Notes Studio Pro, or the native Growth Report. It only emerged from cross-referencing 290 notes downloaded and analyzed one by one. Which means every Substack creator adding attachments to their Notes is likely sabotaging their own reach without knowing it.

How I built the system (without knowing how to code)

As I shared in From Coder to Architect, my path isn’t that of a traditional developer. This project proves it.

Using Claude Cowork, I reverse-engineered 6 private Substack API endpoints, all undocumented, discovered by digging through the browser’s DevTools. I wrote Python scripts, handled authentication via session cookies extracted from the browser, and built the entire data pipeline in 30-60 minute Cowork sessions. The result is a dataset that includes: 290 notes analyzed, 42 articles with full statistics (views, open rate, likes, signups, engagement rate), 587 comments from 96 notes (with discussion depth, subscriber tier, author), 323 subscribers with join date, activity rating, and revenue, plus a daily time series of subscriber growth since November 2025.

One technical detail worth mentioning: the subscriber endpoint only works via authenticated POST from the logged-in browser. Direct terminal requests get blocked by Cloudflare. This explains why no external tool can access this data: Substack actively guards it. Everything runs on flat JSON files, no external database, zero infrastructure.

How to use the kit: the prompts to start with

The GitHub kit is already built. You don’t need to construct anything — you just need to know what to ask Cowork once you’ve set it up.

After cloning the repo and filling in config.json with your publication slug and session cookie (instructions are in the README), open Cowork and start here:

First: “Read the CLAUDE.md file and explain how this system works: what each script does, what order to run them in, and what I get at the end.”

Second: “Run fetch_notes.py. If you get an authentication error, explain step by step how to renew the cookie. When the fetch is done, tell me how many notes you downloaded.”

Third: “Analyze the downloaded data. Tell me which content types generate the most likes on average, and whether there’s a statistically meaningful difference between notes with attachments and notes without.”

Fourth: “Cross-reference the articles data with the subscriber data. Which articles generated the most new subscribers? What percentage of total signups do the top performers represent?”

These four prompts, in sequence, already surface most of the strategic insights I described above. Each one builds on the previous. Each requires zero technical knowledge: Cowork handles the errors, reads the data, and translates everything into operational language.

Playwright MCP: the next level (for those who want to go further)

Cowork, by default, can control Chrome through the Claude in Chrome extension. It works well for navigating pages, clicking, filling forms, and using your already-logged-in browser sessions. But it has a hard limit: it can’t access DevTools, can’t inspect network traffic, can’t see the API calls a page is making in the background. It’s like being able to drive a car without being able to open the hood.

Playwright MCP is the open hood.

It’s a protocol developed by Microsoft that gives Cowork full access to the browser’s developer tools: Network tab, request headers, response bodies, authentication cookies. It’s what I used to find the 6 private Substack endpoints that no SaaS tool knows about — by navigating the site with Cowork controlling the browser while it read, in real time, every API call the dashboard was making in the background.

How to configure it: In Claude’s MCP configuration file (claude_desktop_config.json), add:

{

“mcpServers”: {

“playwright”: {

“command”: “npx”,

“args”: [”@playwright/mcp@latest”]

}

}

}

One detail that makes all the difference: once installed, Cowork doesn’t use Playwright automatically. It defaults back to the Chrome extension unless you tell it explicitly. The simplest way is to add it to your prompt — or specify it in the project’s CLAUDE.md file, which is already configured this way in the GitHub kit.

With Playwright MCP active, you can do things that used to require a developer. For example: “Use Playwright MCP to open the browser, navigate to [site], interact with the page while watching the network traffic. Find all the API endpoints being called, with URL, HTTP method, and authentication headers. Then write a Python script that replicates them automatically, test it, and tell me where to find the final file.”

Cowork runs the entire chain in a single session. This isn’t assistance. It’s complete engineering on data a platform never intended to give you.

The numbers that rewrote the strategy

Beyond the attachment finding, what emerged from the analysis redefined how I think about content.

The top 20% of my articles generate 91% of signups. This isn’t a normal distribution. It’s an outlier-driven distribution. The channel doesn’t thrive on averages; it thrives on a few breakout pieces. This fundamentally changes the strategy: it’s not about publishing more, it’s about understanding what makes that top 20% different and replicating it. As I highlighted in my Substack Growth Manual, growth isn’t about volume, it’s about precision. Now I have the data to prove it.

Then there’s topic clustering. I write about artificial intelligence. The analysis revealed that “AI How-To” notes, the practical, operational ones with concrete instructions, deliver the best production-to-return ratio: 5.2 average likes. The interesting part? They represent only 8% of the content I produce. I’m underproducing the exact format that works best. The specific numbers are mine, but the method works for anyone: with this Cowork setup, any creator can build their own analysis for their own topics, whatever the niche.

Finally, an insight I would have misread without data. The correlation between views and likes is strong and positive: r=+0.80. More people see a piece, more people like it. Straightforward. But the correlation between views and open rate is negative: r=-0.63. The more external reach you generate, the more your percentage open rate drops. Sounds like a problem, right? It isn’t. When content performs well and reaches new audiences, that audience is “cold”: they don’t know you yet, they don’t habitually open your emails. Your open rate drops not because existing readers lose interest, but because you’re growing. Without data, as I previously discussed when analyzing how LinkedIn hides your data, you risk making bad decisions based on misread metrics.

You can do this too. Right now.

The entire system is open source, free, and ready to use. You’ll find it on GitHub. The repository contains all parameterized Python scripts (none of my data hardcoded), a generic CLAUDE.md file with Cowork instructions, a setup prompt, and a config.json example.

Download the repo and put them in a Cowork folder. Claude Cowork is the great equalizer: it removes the technical barrier. You don’t need to know how to code. You need to want to understand your data. As I demonstrated with the Moltbot experiment, today’s AI tools let anyone do things that a year ago required a technical team. This project is the concrete proof.

Remember that 30% of subscriber growth on Substack happens without external promotion, through the platform’s internal network. Understanding exactly which content activates that network is the difference between growing by accident and growing by strategy.

Reality Check: Let’s be honest. Substack is a platform generating half a billion dollars a year for creators and it doesn’t have a proper analytics panel. Notes impressions accessible only from mobile? In 2026? That’s a design choice that says a lot about how the platform views its creators: as users, not entrepreneurs. The third-party SaaS tools that promise to bridge the gap only do so superficially, because Substack actively protects data behind Cloudflare. The result is an ecosystem where the people producing content and generating revenue for the platform don’t have access to their own strategic data. It’s not a bug. It’s a business model. And until creators notice, nothing will change.

---

Want to turn your data into strategic decisions, not just dashboards? From building analytics systems to automating decision-making processes, I help businesses and professionals use AI for measurable outcomes. Explore my AI Consulting services or reach out directly for a discovery call.

---

Where we go from here

This project isn’t about Substack. It’s about a bigger principle: your data is your asset. On any platform. LinkedIn, YouTube, Instagram, Substack, it doesn’t matter. Every time you accept seeing only what the platform decides to show you, you’re delegating your strategy to someone else.

The tools to take back control exist. Claude Cowork makes them accessible to anyone with the curiosity to look beyond the dashboard. The question isn’t “do I have the technical skills to do this?” The question is: “do I actually want to know what my numbers are saying?”

Those who build seriously don’t wait for the platform to give them permission to understand.