How to Write a Week of Substack Notes in 20 Minutes (With an AI Agent You Build Yourself)
You'll walk away knowing how to build a four-step AI agent that generates 21 Substack Notes in your voice, using your own performance data — so Sunday content prep goes from 90 minutes to 20.
What You'll Build
A four-step AI workflow that replaces your weekly Notes production entirely. The agent pulls your Substack performance data, proposes a topic mix based on what actually performed, writes 21 Notes in your voice using a structured prompt, and exports them to a CSV you can batch-schedule. One command, 20 minutes, a full week of content.
The core outcome: stop spending Sunday writing content. Start spending it on the work that actually compounds — coaching, client calls, building offers.
Why Your Manual Notes Process Keeps Failing You
Most creators write Notes from memory and gut feel. You scroll your dashboard, estimate what did well, pick topics based on what feels right, paste a prompt into Claude, and then spend 30 more minutes editing. That's still 90 minutes of production work — every week.
The prompt is doing maybe 30% of the job. You're still doing the other 70%: gathering data, choosing topics, enforcing quality standards, and scheduling one by one. The problem isn't effort. It's that you're running each step manually when they could be chained together.
One data point that changes how you think about this: a Note with 34 likes can convert more subscribers than one with 266. Virality and conversion are not the same metric. You'll never catch that pattern by scrolling your dashboard. An agent running on real numbers will.
How It Works — The 4-Step Agent
Step 1: Pull performance data. The agent reads your last 7 days of Substack Notes — likes, comments, restacks, subscriber conversions for each one — and summarizes patterns. Which formats pulled the most engagement. Which topics converted subscribers. Which Notes flopped. This is the step most creators skip. The agent catches it in 8 seconds.
Step 2: Generate your topic mix. Based on the performance summary, the agent proposes 21 topics — 3 Notes per day for 7 days. It pulls from your content pillars and weights toward what performed best last week. If you published an article mid-week, it pulls the most surprising data point, the most actionable tactic, and the core insight — and turns each into a Note. Repurposing happens automatically.
Step 3: Write 21 Notes in your voice. The agent reads your voice profile (sentence rhythm, banned phrases, format preferences) and writes all 21 Notes in one pass. The output mix is intentional: at least 2 micro-Notes under 100 characters, at least 3 in the 100–250 range, and longer pieces up to 600 characters. No two consecutive Notes use the same format. The agent enforces rules you used to check manually.
Step 4: Export and schedule. The agent writes all 21 Notes to a CSV with columns for content, date, and time. You upload that CSV to a batch scheduler — Substack only lets you schedule one Note at a time manually, so the CSV route gets around this entirely. Notes post automatically at the times your audience is most active.
📌 Key Takeaways
📌 Performance data beats intuition. Feed last week's engagement numbers into the prompt before generating. Three specific examples beat a page of instructions every time.
📌 Voice rules are non-negotiable. Read 10 of your published posts first. Pull phrases you actually use and phrases you'd never write. That contrast is what prevents generic AI output.
📌 The feedback loop is what compounds. Performance data from this week feeds back into next week's generation. The agent gets better every cycle because it's learning from real numbers, not static formulas.
📌 Format variety is a growth lever. Accounts that alternate between ultra-short and medium-length Notes get 40% more profile visits than those posting the same format repeatedly.
📌 Start without the full automation. No Claude Code? Paste the prompt into any Claude conversation. You lose the automation layer but keep the voice matching and format variety. That's your v1.
Files and Tools You Need
Claude (claude.ai, any paid plan) — for the Notes generation prompt
Claude Code (optional, for full automation) — handles data pulling via Puppeteer and CSV export via scripts
A Substack batch scheduler — Claudia built one at substack-note-scheduler.up.railway.app
Your last week of Notes data — likes, restacks, and subscriber conversions per Note (pull from your Substack dashboard before you start)
15 minutes to customize the prompt — fill in your newsletter name, voice rules, banned phrases, and content pillars
Ready to Build?
Claudia Faith published the full walkthrough — including the complete copy-paste prompt, the system diagram, and exact setup instructions for the Claude Code version — over at Level Up with AI. Read the full tutorial here: levelupwithai.substack.com
Start with the prompt in any Claude conversation. Once you see the output quality, you'll understand why automating the rest of the steps is worth it.
Want more tutorials like this? Every week I find the most practical AI tutorials on the internet and break them down for women in tech and consulting who want to build faster, work smarter, and stop leaving hours on the table. Subscribe to TechwithM.
Original tutorial by Claudia Faith, published May 20, 2026 on Level Up with AI (levelupwithai.substack.com). This post summarizes and adapts the key steps for the TechwithM audience.
