I ran this session on my own business. The job: turn my monthly subscribers into yearly ones, and make sure my yearly subscribers are getting everything they pay for. Two parallel builds, one machine.
The whole thing turned on a single instruction I gave the model halfway through.
Don’t segment.
Kill the segment
When the plan came back, it proposed the obvious thing — bucket the subscribers into four segments by role and company type, then build one gift for each bucket. I stopped it. It should not be segmented.
Here’s why. The entire reason to do this with agents is that you no longer have to lump people together. Segmentation was always a compromise — the best you could manage when a human had to design a finite number of messages by hand. A model doesn’t have that limit. So take each subscriber’s full LinkedIn profile, their company, everything I’ve ever written in the wiki, my live Substack feed, and all ninety-to-a-hundred skills I’ve shipped — and have it return the top three things that are genuinely useful to that one person, with the reasoning behind each. If there’s no fit, it says so, and offers to build something instead. Nobody gets a generic experience.
I’m betting this lifts conversion and cuts churn at the same time, because the whole promise of these models is a pathway built for the individual. That’s the feature. Almost nobody implements it, and the few who do win twice.
“The whole purpose of LLMs is to give people pathways that are relevant to them.” — Jordan Crawford
Your best signal is in your own won deals
One member on the call had started recording every sales call, and it paid off immediately. A wine company he’d closed mentioned, in passing, a courier that kept smashing their bottles. That’s not a testimonial — it’s a targeting signal. So he pointed Claude at it, scraped who else uses that same courier, built a list, and had a campaign live the same day.
That’s the move nobody makes. The pain that closes your next deal is usually sitting in the transcript of your last one. Record every call, then mine it. The best signal you’ll ever get isn’t for sale — it’s in your own won-and-lost conversations, and the moment you write it down, you can go find everyone else who has the same problem.
Run it in parallel, then trust it by running it twice
A build like the subscriber engine dies if one agent tries to hold the whole thing. So chunk it: one wave of agents enriches contacts, the next validates them — are the titles right, is the LinkedIn live as of today? — the next figures out who works at a real company, and only then does the recommendation step run. Parallel, not sequential.
Then the discipline that keeps it honest, because a machine blending this much data will confidently hand you garbage. I duplicate a slice of the work — submit the same person through more than one window, with maybe twenty percent of the context overlapping — and have a judge compare the answers. If the same subscriber comes back with three different recommendations, the data isn’t trustworthy, and I’ve caught it before it ships instead of after. When the agents disagree, that disagreement is the signal.
Who Gets This
The worked examples below are the actual builds from the room — what we made, what I told each person to do with it, and the exact moves.
Free: the shift — kill the segment, mine your own deals, verify by running it twice.
$50/mo (most readers start here): everything below the line in this post, and every paid article.
$2,499/yr: Every tool I ship. Edge Copilot is how you talk to all of it through Claude Code. Current tools: Edge Copilot, AutoClaygent, Agent 7, Who to Target and What to Say, Blueprint Cloud, Technology Finder, Video List Extractor, Competitor Monitor, LinkedIn Engagement, Domain & LinkedIn Finder, Dossier Builder, PDF Contact Finder, TAM Contact Harvester, Find a Rep, Blueprint Playbook, Crawford, JoJo. Whatever ships next is included. Plus all 3 courses + weekly Applied Office Hours. (Go annual — $2,499/yr.)
Below the line in this post:
The per-subscriber engine: enrich → validate → recommend top three, judged, one person at a time
How to mine a closed deal into a live campaign — and why nine-person campaigns aren’t worth running
The economics: sub-agents over the API, and the Batch API trick that’s 50% off
Every question from the room, with the answer
Every week I run Applied Office Hours on Zoom — bring what you're building and we'll work it live.
