Finn'sights

My swipe file wasn't a museum. It was a storage unit. I learned about swipe files from Justin Welsh. The pitch is seductive. Great creators don’t start from blank pages. They collect. Every banger headline, every hook that stopped their scroll, every note that made them think I wish I’d written that — it all goes into one place. Your private museum of proof that great writing exists. Then, when you sit down to write, you don’t summon genius from nothing. You visit the museum. I was sold instantly.

The comedian finally has a room. It only took a year — and one subject line I almost didn't send. Twelve months ago, I published I Built a Note-Writing AI That Learns From Its Own Mistakes — a 32-billion-parameter model on my MacBook, writing Substack Notes, scoring its own drafts, iterating. I called it a learning loop. Idea, action, observation, evaluation, refinement. The honest answer at the time was: barely. It could critique a Note. It couldn’t hear whether anyone laughed.

I ran one question against my own list. 336 names came back. People who opened five of my emails — then nothing for thirty days. Not strangers. Not tire-kickers. Readers who were in the room with me. Warm. Engaged. Already halfway out the door. Substack’s dashboard never flagged them. My open rate looked fine. My subscriber count went up. I felt like I was doing the work. I wasn’t. I was watching a scoreboard while people quietly left. Honest question: how many of yours are doing the same thing right now?

Comment @stacksave before it scrolls away — then search your archive at stacksave.live. Substack’s save button only works on posts. The best stuff is usually somewhere else — a sharp Note, a reply buried in a thread, a one-liner you swore you’d remember. Then you scroll, and it’s gone. I ran a quick survey on June 10th. 54% of respondents said they lose Notes and threads regularly or constantly.

581 subscribers arrived through 49 peer recommendations. The top 12 sources account for 472 (81%). Thirty of them got nothing back. I have the query results sitting in a spreadsheet. In 2024, I published a Sankey diagram of follower overlap among Medium’s top 12 writers. It showed something most creators pretend isn’t true: your audience is never just yours. Nearly 40% of Derek Hughes’s followers also followed Eve Arnold . Tim Denning’s bar touched almost everyone.

The three-step loop I ran on Saturdays and Sundays while everyone else said “post daily.” When I stopped optimizing for likes and started running the weekend loop, revenue and subscribers finally moved on different schedules—and both went up. I used to think the problem was time. You hear it everywhere: quit your job, post daily, grind until something breaks your way.

Posting Notes is like throwing darts at midnight The Substack feed shows hits. It does not show attempts. That distortion makes “one viral Note” feel closer than it is. The truth is simpler: outcomes on social media are heavy-tailed. A few posts carry most of the visible signal. Everything else looks like failure when it is actually the sampling cost of finding who you serve—and who sits one ring out.

Pi + Ollama + Qwen: a newsletter stack that stops the subscription tax I was spending $60/month on AI tools for a newsletter with 47 subscribers, mostly to draft Notes. The tools worked; the monthly line item didn’t - at that scale, it never felt in proportion to where I was. I kept renewing in part because cancelling the AI tool felt like downgrading the project, not because the line item made sense for a list that small.

Same loop. One of these cost $1 billion to build. The other costs $49. Netflix runs thousands of A/B tests simultaneously across 325 million subscribers. Their personalization system — built on that constant stream of experiments — saves them an estimated $1 billion a year in churn. 80% of what you watch on Netflix was put there by an algorithm trained on billions of small decisions. Netflix spends nearly $1 billion per quarter on technology and development just to keep that machine running.

4 am. Dashboard open. Kill-product doc one tab over. This is where this post started. 22 months of building tools. $26,752 in net revenue — every dollar one-time, no subscriptions, no recurring billing tricks. 721 paying customers across 8 tools, from 69 different countries. And last month, I opened my dashboard and almost killed the product that pays my bills. This is the post I’ve been avoiding for two weeks.