Ryan “RyMac” McKinney · USMC · 15 August 2026 · 6 minute read
Anthropic, the company that makes Claude, has published its own guidance on context engineering for agents. It is good and it is written for people building software. I am not one of those. Here is what it actually says, and what changes if your job is running a business that now has an AI in it.
Because most of what you will find on this subject is somebody restating a vendor blog with their own product bolted on the end.
The people who build the model have an unusual view of what breaks. Worth 20 minutes of your time, even if half of it is not aimed at you.
The consistent message across the serious writing on this, Anthropic's included, is that the limit is not model intelligence. It is what the model can see.
And more is not automatically better. Give a model a large amount of material and its ability to use any particular piece of it degrades. The signal gets buried under things that merely look relevant.
That single finding kills the most common instinct people have, which is to paste everything in and hope.
The technical version of the advice is: curate the context window deliberately, retrieve what is needed, keep the rest out.
The business-owner translation is shorter.
Decide what is on the desk. Put everything else in a drawer. Tape a card to the desk saying which drawer holds what.
That is the same instruction. One version needs a codebase. The other needs a folder.
~/business/ ├── CLAUDE.md <- the map. 11 lines. always read first ├── Context.md <- the desk. what we're doing today └── ops/ ├── clients/ ├── offers/ ├── sops/ └── archive/
Not on the principle. On what you are expected to do about it.
Almost every implementation guide assumes you will build the curation in code: a retrieval layer, an evaluation harness, a memory service.
If you are shipping a product to customers, build that. It is the right answer.
If you are trying to stop re-explaining your own business every morning, you do not need a retrieval layer. You need to write things down somewhere the agent can find them, and be consistent about where.
Folders do not scale infinitely and I am not going to pretend otherwise.
At some size you genuinely want search over your material rather than a path to it. The signal is when you cannot say which folder something is in without looking.
I have not hit that running several businesses. You might. When you do, the developer guides are waiting and they are correct.
Interpretable Context Methodology, ICM, from the paper by Jake Van Clief and David McDermott. Folder structure as agent architecture. Jake teaches it himself over at Clief Notes.
Plain English breakdown: Interpretable Context Methodology in plain English. The concept itself: context engineering.
CLAUDE.md, Context.md, the folder skeleton, a 1-page setup sheet, and a teardown checklist for whatever you are already paying for. 20 minutes, start to finish. No card, no call, no course to buy.
Send me the folder system