One of the makers of Claude Code calls himself a context minimalist. His philosophy: tell the model only what it needs to know, and let it work out the rest. That ran straight against my instinct.
My first reaction was the wrong one. I thought: surely more context is better? The more I explain, the less room for misunderstanding. That sounds logical. It just is not true.
Too much explanation is micromanaging
The makers put it in an image that stuck with me. When you give the model too much context, you are really micromanaging it. You spell out the whole route, while the model sometimes knows a better way to the same outcome. By nailing everything down, you take that option away.
I recognise it in myself. My instruction file for this site grew to hold all sorts of things: how an article should go, which steps in which order, what to avoid above all. Well meant. But every rule I added was also a rule the model had to obey instead of think for itself. When I cut a lot of it, the work did not get worse. It got looser, and often better.
From prompt engineering to almost nothing
There is a line the makers sketch nicely. With Sonnet 3.5 you had to do prompt engineering: find exactly the right words. With Opus 4 it was context engineering: feed it the right information. But with today’s models you do less and less of both. You give the smallest possible system prompt, the smallest possible set of tools, and then you let the model figure it out.
That is not laziness. It is trust that the model knows enough to choose its own path, as long as you give it a way to pull in the context it needs. The craft shifts from telling it a lot to choosing well what you leave out.
Why this feels hard
Saying less feels like less grip, and that is exactly where it rubs. A long, detailed instruction gives the feeling that you are in control. But that feeling is partly an illusion. You control the words, not the outcome.
For me this is the same lesson as letting go of Plan mode. Grip is not in nailing everything down up front. Grip is in seeing what comes out and steering where it is needed. An agent given room surprises you in a good way more often than one kept on a short leash.
Since I prescribe less, the model more often takes a route I would not have thought of myself. With all those rules, I had made that impossible.
Part 2 of a series on how the makers of Claude Code work with it themselves. Earlier: I dropped Plan mode.