When I was first starting with AI, about two years ago and it seems longer than that, the main buzzword was prompt engineering. Now, that's commodity. Anyone can go in and build a proper prompt in seconds. The new buzzword, and it's not so new, but something I keep wanting to improve. Context engineering.
The file that holds everything my AI knows about this business had quietly grown to about thirty-five thousand tokens. Last week I watched it load on boot and stop partway through, cut off before it reached the bottom. The single most important memory file I own had become too big to read in one sitting. There is a name for that problem and for the fix: context engineering.
I wrote back on Day 89 that you should give every project its own brain, one plain-text file the agent reads first and writes back to last. I stand by it. It is still the highest-return move a non-developer can make with AI. What I did not say, because I had not hit it yet, is that a brain you only ever add to eventually gets fat, and a fat brain is a slow, unreliable one.
Nobody tells you this part when they say "give your AI memory." The model does not read your whole vault. It reads a window, a fixed amount of text, and then it is full. Everything you want it to know has to fit in that window alongside the actual question. Context engineering is the unglamorous work of deciding what goes into that window and what stays out. Not a cleverer prompt. A cleaner desk. You are choosing what the agent sees before it starts thinking.
I've spent two days on proper context engineering, using the thin harness fat skills build ideas.
I ran an audit of my own setup this week and rebuilt the whole thing on four dull rules that came out of it. None of them are clever.
One home per fact. If the price of a service, or the date a thing shipped, lives in two files, the two will drift, and one day the agent reads you the wrong one with total confidence. Every fact gets exactly one place to live.
Supersede, do not append. This is the one that had fattened my file. Every session I stacked a new block on top and left the old one sitting there, true last month, wrong today, still loading. The rule now: when something changes, I edit it in place and date it. The file stays the same size instead of swelling by a block a week.
A thin index on top. Instead of one giant file the agent swallows whole, there is now a short map, a page of pointers, and one small file per project underneath it. Boot reads the map and the single page for whatever we are working on. It reads a table of contents and one chapter, not the entire book, every morning.
The cold-boot test. I open a brand-new session, ask it where a project stands, and it has to answer from the map and one small file, no hand-searching. Then I ask it something the files do not contain, and it has to say "that is not written down" instead of inventing a number. Doesn't forget, doesn't invent. If it fails either half, the memory is not built right yet. It is the same discipline I first wrote about on Day 77, a scheduled audit so the agent never quietly rots.
I did the same for every project I run, including the small one I keep for the fund in my son's memory. Each brain is small, current, and quick to read again.
Not a build-it-and-leave-it deal, but a constant tweak to make it better. These were some of mine.
Monthly Revenues $11,000 | Clients 2 | Prospects 1 outbound live, Meta and WhatsApp still down
Day 113 of 365.