Context is the new working storage

When I started out, memory was expensive. You thought hard about every field you declared in working storage, because there wasn't room for anything you didn't need.

Large language models brought that feeling back. A context window looks huge, but it fills up fast, and what you put in it shapes what comes out. People call this "context engineering" now. I call it a very familiar discipline.

Three habits that carry over

Load only what the step needs. Pasting the whole codebase or the whole manual "just in case" doesn't make answers better. It dilutes them. Give the model the two files that matter and say why they matter.

Name things clearly. A well-named field saved the next programmer an hour. A clear section heading in a prompt ("Constraints", "Example input", "What not to do") does the same for the model.

Free what you're done with. In long sessions, summarize and restart instead of letting stale detail pile up. A fresh context with a tight summary often beats a crowded one with everything in it.

The hardware got bigger. The principle didn't change: the quality of the output depends on how carefully you manage what's in memory.