Karpathy's line — "filling the context window with just the right information, at just the right time" — has quietly become the organizing principle of the field. Five GitHub repos with roughly 170,000 combined stars are now using the context-engineering framing in their READMEs, including context-engineering-intro and Context-Engineering. Two years ago the same folder structure would have been called "prompt engineering." Today the phrase sounds dated.

The shift is methodological, not semantic

Prompt engineering was a craft: you wrote a clever instruction, tuned it, and hoped the model did the thing. Context engineering is a system: you decide what goes into the window, in what order, with what citations, with what constraints, and you treat the whole bundle as the unit of work. The prompt is a small part of that bundle. The big parts are the files, the retrieved snippets, the tool results, the memory, the constraints.

This matters because the failure modes are different. A bad prompt produces a bad answer. A bad context produces a confidently wrong answer that looks right, which is far more expensive. The discipline of context engineering is the discipline of not letting that happen.

CLAUDE.md is the canonical implementation

The pattern that won is file-system-native. Claude Code reads a CLAUDE.md from your repo, loads it into every session, and now your "prompt" is a version-controlled document that lives with the code. Cursor Rules, Windsurf memories, and Aider conventions are the same idea in different dialects. Obsidian vaults with an LLM on top use the same shape applied to knowledge instead of code. The convergence is hard to miss.

The implication for builders: stop iterating on prompts in a chat window. Move them into files. Treat the context bundle as the product. Version it, review it, test it.

What to watch

Two tells. First, whether the framing gets absorbed into training and evaluation — benchmarks like "long-context faithfulness" and "context packing" are already drifting this way. Second, tooling. A standard format for context bundles (something between Dockerfiles and prompt files) is the obvious next step. The repo that nails it will be next year's 50k-star project.