My second brain works the night shift.
Capture is where every guide stops. Mine got interesting the day it quit waiting to be asked: it files overnight, hands me a brief at breakfast, and gets graded on both.
Everybody is building a second brain right now, and good. The guides are right. Point a model at a pile of your own notes, capture everything, link it, let it read across all of it, and never re-explain yourself again. If you have not built one, build one. Mine is the highest-leverage thing I run.
But the version the guides get you to is a brain that answers questions. Mine got fun the day it stopped waiting for me to ask.
The night shift
Here is what mine did last night. It pulled in what landed during the day, the emails, the meeting notes, the files, and put each piece where it belongs, linked to the ventures and people it touches. Then it wrote the morning brief: what moved, what needs a decision from me, and what an eval flagged as off. I did not run a query to get any of that. It ran while I slept.
That is the leap. Not ask my notes a question. My brain handled what it could and surfaced the few things only I can decide. Capture is table stakes. Acting on the capture is the product.
Answers are the floor, not the ceiling
Most setups store what you give them and answer questions about it. That is genuinely useful, and it is also passive. You still show up, ask, read the answer, and go do the work yourself. The brain got smarter. Your week did not get lighter.
Karpathy’s original idea, the one these guides are built on, was never store more. It was compile. Raw material goes in and gets refined into pages that hold a real understanding, so the next question starts from something better than the raw notes. Storage does not compound. Compiled knowledge does, and a brain that acts keeps compiling whether or not you showed up that day.
And evals grade all of it, or it drifts
A brain that acts will sometimes act wrong. That is why mine is covered in evals. Not one grade, all sorts: the overnight filing has one, proposals have one, decks have one, even the video work gets scored. Each eval is a standard I set, what good looks like to me, written down so the system can check its work against it on every run.
Without them, an acting brain just makes confident mistakes faster. With them, every miss becomes a lesson, and my taste compounds into the machine instead of living only in my head. The order that matters: capture, then act, then grade. The grades are what turn a pile of notes into a system that is better in month three than it was in month one.
One rule, because this part bites people
When your brain can act, give it keys, not instructions. Telling an agent do not delete that is a suggestion. If it technically can delete the file or send the email, assume one day it will. Control what it can do at the permission level, scoped and read-only where you can, not with a polite line in a prompt. A brain that acts is powerful exactly because it can reach things, so be deliberate about what it can reach.
If you want the full machine under this, the substrate, the cockpit I run the fleet from, and how the grading works, that is The Loop Is the Product. Build the second brain. Just do not stop at the part that answers. The fun starts when it acts, and the evals keep score.