The numbers (so the rest of this is grounded)
Between November 24 and April 22, 108 people paid $126,000 across Second Brain and Second Brain Factory to build their AI work system. The direct-purchase split: 40 DIY, 42 Kickstart, 22 Done-With-You, plus Factory creator-side revenue on top. 4 refunds across the whole book. No team. No VC.
And of the 80 customers who filled out the post-purchase questionnaire honestly:
- 55 self-identified as non-technical. 20 said "some CLI." 5 were technical. So 75 of 80 (94%) couldn't comfortably live in a terminal, even though that's exactly what V1 required them to do.
- Only 30 of 80 had ever used Claude Code before buying. 50 hadn't.
- The most popular build environment was Claude Web (35), then Codex (18), then Claude Code (16). Eight people picked "I don't know."
- 95% started personal-only or personal-first-team-later. Two people bought it for a team on day one.
That's the customer base. Now the lessons.
What worked
1. The 1:1 setup call paid for itself in pattern recognition.
The Kickstart and Done-With-You packages included a call. 89% of buyers booked it (48 of 54). I thought I was selling personalized setup. What I was actually buying was 90 minutes per customer to watch where they got stuck. By customer 30 I had a list of recurring failure modes that no questionnaire would have surfaced. By customer 60 I knew which features to cut.
The DIY tier didn't have this loop. DIY was the highest-friction package and produced 3 of the 4 refunds.
2. Personal-first scope kept the blast radius small.
95% of buyers started solo. That meant a setup gone wrong only broke one person's workflow, not a team's. It also meant feedback was fast, raw, and honest. Nobody had to escalate it through their manager.
3. Specific use cases beat "general productivity."
The seven use case buckets that customers self-selected, ranked by frequency, were:
- Strategic planning (51 of 80)
- Research and analysis (51)
- Operations (50)
- Content creation (50)
- Sales and outreach (48)
- Client work (47)
- Learning and knowledge management (43)
Every customer picked four or five. Nobody picked one. The Second Brain that worked was the one that covered most of those at a 70% level. The Second Brain that failed was the one that did one of them at 95% but ignored the rest.
4. Memory beat intelligence, every single time.
The customers who got the most value out of their setup weren't the ones who configured the smartest agents. They were the ones whose CLAUDE.md held real context: their rates, their clients, their methodology, their voice. One quote from the questionnaire said it perfectly:
"Having to re-explain my entire consulting approach every session. Claude does not know my rates, my methodology, or my quality standards."
That's the problem Second Brain solves. Not "smarter AI." Smarter AI is everywhere. The thing nobody is shipping is a system around the AI that remembers what you did yesterday.
What failed
1. The terminal was a wall.
V1 required Claude Code. Claude Code lives in a terminal. 75 of my 80 customers couldn't comfortably use a terminal. That's not a bug, that's a category mismatch. I sold a tool to people who couldn't open it without help.
The Kickstart calls revealed this within the first ten minutes every single time: "Where do I type this?" The fact that people still bought, still showed up, and still tried to make it work is its own data point. They wanted what I was selling. They just couldn't reach it.
2. The questionnaire was too long.
V1's onboarding questionnaire had 60 fields. People filled the first 15 honestly, defaulted the next 30, and skipped the last 15. The data I got back was structurally complete and substantively empty. I had to chase half of it on the call.
I kept telling myself the questionnaire was "rigorous." It was actually a self-imposed bottleneck.
3. Documentation-first onboarding is documentation-only onboarding.
I shipped V1 with a 14-page setup doc and assumed people would read it. They didn't. They scanned the headers, ran the first command, hit an error, opened a different tab, and never came back to the doc. The doc was correct. Correctness is not the same as usable.
4. Maintenance loops do not maintain themselves.
A handful of customers got the system running brilliantly in week one and then watched it decay over the next month. Their CLAUDE.md got stale. Their memory file accumulated junk. Their skills directory drifted from what they actually did. Nobody runs git status on their second brain.
This was the most painful failure because the customer experience was: "This was amazing. Then it stopped being amazing. I think I broke something." They didn't break anything. The system aged out from under them.
The three reasons people stalled
From the questionnaire write-ins, three failure modes show up over and over. None of them are exotic. All of them are fixable.
1. Setup friction stacks faster than excitement.
Real quotes:
"Still blocked on the setup. Went halfway through, set up a few skills, but never got the quality I wanted. I'd like to start from scratch."
"Buying the license. No clue how it works exactly."
"Separating business and private setup. Need to create two accounts."
"Just haven't started using it yet. If it requires a local install I'd need to set up admin privileges for that."
None of these are the Second Brain. They are the dependencies. Claude Pro license, IT admin privileges, separating personal and work accounts, terminal install, MCP config. By the time someone clears all of that, the dopamine of "I just bought something exciting" has bled out and the second brain is just another tab they're avoiding.
2. The quality gap between expected and built.
Customers who bought DIY especially hit this one. They imagined a polished agent that already knew their world. They got a starter kit that needed an hour of configuration to start producing what they imagined on day one. The gap between "I just paid for this" and "and now what" is where most quitting happens.
This is why the Kickstart call worked: I closed that gap in real time. DIY customers had to close it themselves, and most stopped trying after two attempts.
3. Mental load.
"The biggest blocker is mental load. I interact with a lot of people, and tracking context, timing, and loose threads in my head slows the whole workflow down. Nothing is technically hard, it's just cognitively expensive to remember it all."
Every customer who quit had this in some form. They came to Second Brain because their mental load was too high. The setup increased their mental load before reducing it. If the dip lasted more than a week, they were gone.
What I rebuilt for V2
Once those three failure modes were clear, V2 wrote itself.
1. Claude Desktop, not terminal. Same memory model, same agents, no terminal. Of the 80 questionnaires, 35 picked Claude Web as their primary build environment. V2 meets them there.
2. An onboarding agent that replaced me. The 1:1 calls were the highest-leverage thing I did, and also unscalable. So I rebuilt the call as an agent. It walks new customers through their setup the same way I did on Zoom, asks the same questions, picks the same patterns.
3. A health check that scores 45 dimensions. This is the answer to the post-purchase question that drove most of the anxiety: "did I set this up right?" Now there's a number. It's free. Anyone can run it, customer or not.
4. Memory that compounds AND self-cleans. The maintenance problem from V1. V2's memory doesn't just accumulate. It rotates, decays, and consolidates. The customer doesn't have to remember to clean it.
5. 30 agents, 80 skills, 20 MCP integrations out of the box. No more "blank canvas." Day one is already populated. You delete what you don't need. You don't start from zero.
6. Six-step onboarding instead of 60-field questionnaire. Cut the inputs by 90%. The remaining six are the only ones that actually changed what got built.
The bigger pattern
If I had to compress 108 setups into one sentence, it would be this:
You don't need better AI. You need a system around it that remembers what you did yesterday.
Every customer who came in asking "which model should I use" was asking the wrong question. Every customer who came in asking "how do I make Claude know my world" was asking the right one. The model gets smarter every six months whether you do anything or not. The context, the memory, the agents, the skills, the integrations: those are yours to build, and they compound.
That's what Second Brain 2.0 is. The same thesis as V1, finally usable by the 75 of 80 customers who couldn't open a terminal.
If any of this lands wrong, or you want me to expand a section, send me a message on LinkedIn.
— Iwo