I'm self-taught. Online courses and a lot of trial and error, not a computer science background. What's below is the automation stack I built and still maintain myself, running unattended every night across every retainer client at Three Seven. Not a tool I subscribed to. Not one company's setup. Mine, for all of them.
None of that happened on the first try.
June 2026: the version that kept breaking
The first version worked just often enough to convince me it was close, and failed just often enough to be useless. Four separate root causes, and I found each one the hard way.
A stray API key export was silently routing every run to the wrong billing org, which meant it was getting rate-limited without ever telling me why. Overnight wake-ups had no wifi nearby half the time, so the system burned two hours retrying a connection that wasn't coming back. There were no failure alerts at all — if a run died at 2am, I found out the next afternoon, if I thought to check. And it had no memory of what it had already finished, so a daytime rerun would redo the entire batch and blow through a token budget that was supposed to last the week.
I traced all four, one at a time, over a few weeks of watching it fail in new ways.
July 14: rebuilt from the root causes, not the symptoms
The rebuild didn't add more automation on top. It fixed what was actually broken.
Every client run is now state-tracked, so if one client fails partway through, it just retries that client the next night instead of redoing the whole batch. Pre-flight checks confirm network, auth, and Drive access before a single token gets spent — the system checks its own footing before it starts walking. It sends me a Slack message after every run, success or failure, and if it failed, it says why. And nightly batches are capped per client, so one bad run can't blow past a single token window and take the rest of the night down with it.
Proof it holds, not just a claim that it does
A few weeks ago I was on a camping trip with no wifi for several nights in a row. The overnight batch had no connection to work with, so it didn't work — it logged exactly why, sat quietly, and picked back up the moment I was back in range. Nothing lost, nothing duplicated, no 2am panic-checking from a tent. That's the design doing what it's supposed to, not a lucky week.
Why this isn't something you could just buy
Before it touches a client's account, it reads that client's own history first — goals, past meetings, the things they've explicitly said not to do — so the work it produces matches what that specific client wants instead of running everyone through the same template.
It routes that work to a team of named specialists I built and scoped myself, coordinated by an orchestrator I also built:
Coordinated, not generic — eight roles instead of one bot claiming to know all of it.
Nothing gets marked done or sent to a client without me reviewing it first. Every time. The automation drafts. I approve. When something fails, it tells me in a Slack message, with a reason, instead of quietly retrying forever or pretending the run succeeded.
The volume, for scale
The part people don't believe: this runs on one $20/month Claude subscription, not a metered API bill that scales with how many clients I take on. There's no engineering team behind it and no vendor I'm licensing it from. I learned to build this the same way anyone can now — by using Claude Code itself to write, break, and rewrite it, one real failure at a time, until it held.
source: this system's own run logs, categorized by automation and summed from raw token usage.