A few months ago, I wrote about approaching the ceiling of my own brain, the context problem, and how the fix turned out to be a structured knowledge base rather than a more clever model. That piece ended with a promise, I’d document the system as I went. This newsletter is me keeping it, and the reason it’s a weekly log instead of another one-off essay is that somewhere in the past few months the project has moved from something aspirational and almost sci-fi to now showing real lift on my day to day.
A note I don’t want to skip past. There’s a serious and necessary conversation happening about where all of this productivity, the real and the merely perceived, actually leads, and whether it nets out as a good thing for people. I think about it a lot, and I don’t have a tidy (or consistent) answer. I’ve kept this newsletter pointed at the operational and the practical rather than turning it into a philosophical treatise on the future of knowledge work, mostly because that’s where I think I can actually contribute. But it’s a generationally important question, I do have (ever evolving) views on it, and I expect it will find its way in here more than once.
What this newsletter is: a weekly working log of the rapid changes in knowledge work, written from inside a real advisory practice. Some weeks I’ll take a system apart, the architecture in plain English and what broke along the way. Other weeks I’ll make an argument about where I think this is heading, what it does to the professions and to the people in them. The standing rule, and you can hold me to it, is that every opinion comes with something concrete behind it. I’m not interested in adding another hot take account to your inbox, there are plenty.
Some of what that has looked like in the last few weeks:
- I’ve pointed a few dozen agents at a business development question before I went to bed, and by morning there was a researched, sourced, fact-checked report waiting, work that would have taken a ton of analyst time.
- An internal tool and dashboard I made for our internal calls has morphed into something larger, and my firm is now folding it into its enterprise platform.
- The thing that captures the raw material, the data that feeds the whole system, is an open-source, locally processed, Mac transcriber I built called Detto, it’s live at detto.app and over a thousand people are now using it.
- A colleague on a same-day deadline picked up one of the skill packs I built and got the deliverable out that afternoon.
None of these were demos. They were real value adds to our practice.
The first essay argued that the model was no longer the constraint, the context was. The lesson since then, and I think it’s the bigger one: once the context layer exists, the constraint moves. Production stops being the scarce thing. The first few words on paper, the research, the monitoring, a lot of what used to fill my day now runs at machine speed. What stays scarce is judgment, knowing what’s worth doing in the first place, and whether the output is actually right. The judgment was always the real job. It just used to be buried under the production.
The proof I didn’t expect is by far the least serious one. A few weekends back I made alltimehockey.com, a version of a daily all-time NHL draft game, for no better reason than I wanted to see if I could do it and sounded like fun on a rainy Saturday morning, by Sunday night it was live and by Tuesday it had dozens of users. When production stops being the scarce thing, the distance between “that would be fun” and “that exists at a URL” is about a weekend. It can be both equal parts invigorating and overwhelming to realize that the gap between idea and reality is getting narrower by the day.
What broke this week: since I promised failures, here’s a recent one. That overnight research run didn’t finish its first night. Partway through it hit a usage cap and died quietly, no crash or alert, the machine equivalent of a contractor walking off the job site. The cap was self-inflicted: in another window the same account was busy building that hockey game, the research on one side, my fun project on the other, and together they drained the tank. Running two projects overnight, at the same time, sans AI is something usually impossible for me, short of the grey market supplements from a 2000’s-era GNC and no sleep, so I’m counting the failure itself as a small milestone in productivity.
What saved the work was a boring decision made weeks earlier: After various failed attempts at long, unsupervised tasks, and a ton of wasted tokens setting up too many brittle cron jobs, I’ve set up a scheduled system status check that runs periodically and logs what’s going on and tries to fix whatever is stalled or broken. So, when I opened up my computer the next morning, I was able to kick back off what had stalled and reviewed what had restarted on its own. The failures aren’t prompt problems anymore, they’re operations problems, and the fix is operational too: assume the run dies at some point and set things up so the work survives it.
The number: on a recent research run, the second set of agents, the ones whose only job is to try to prove the first set wrong, flagged more than half of the headline findings as stale, overstated, or just incorrect, before any of it reached me. That’s the one that stuck with me. Research that used to cost my practice days of analyst time now runs overnight, but the speed was never the hard part, the checking is. That verification layer is the difference between machine-speed work and machine-speed mistakes, and it’s next week’s issue, end to end.
Steal this: whatever you’re using AI for today, add one step before you rely on the output. Take its three most important claims and make it prove each one against a source. It costs ten minutes and it catches more than you’d expect. And if you want to overcome some of the remaining sycophancy these models still display, give your agent instructions to treat it as an adversarial, full red-team audit, you might get a clearer result.
Cool thing I read this week: Anthropic matched what 9,700 people said about AI at work against what they actually did with it, and the people who delegate the most turned out to be the least worried about their jobs.
If you’re still in the drowning-in-context stage, the original essay is here.
Next Week: the overnight research system, taken apart.
First published by email and on Substack, July 15, 2026.