Context Engineering in Claude Code and Cursor: What to Copy

Coding agents are among the most sophisticated context engineering systems in production. You can learn the discipline by observing how they manage context, and apply the same patterns to your own systems.

If you want to learn context engineering from the best, study the coding agents. Claude Code and Cursor are among the most refined context systems running in production, used by millions of developers on real codebases that far exceed any context window. The techniques they use to make that work are exactly the four strategies, applied with care, and you can observe and copy them directly.

Procedural Memory: CLAUDE.md and AGENTS.md

Both tools use persistent instruction files, CLAUDE.md for Claude Code, AGENTS.md and similar for Cursor, that hold conventions, patterns and project knowledge. These are the write strategy applied to procedural memory: instructions that survive across sessions and get loaded as context automatically. When you add a CLAUDE.md describing your project conventions, you are doing context engineering, persisting knowledge so the agent applies it consistently without re-explanation.

Auto-Compaction: The Compress Strategy

Claude Code uses auto-compact: when a conversation grows long, it automatically summarises earlier portions to free up space while preserving the thread. This is the compress strategy in production, and it is what lets the tool maintain coherent long sessions without degrading into rot. Observing when and how it compacts teaches you how to implement threshold-triggered compression in your own agents.

Just-in-Time Loading: The Select Strategy

Rather than loading entire files or codebases into context, these tools sample on demand. Claude Code uses commands to read specific portions of large files when needed. Cursor uses codebase indexing to enable intelligent selection of relevant files, and the @ context system lets you explicitly select what enters the window. Both embody just-in-time loading: keep the baseline lean, fetch detail only when the step requires it.

Codebase Indexing: Smart Selection

Cursor codebase indexing deserves special attention. By building a semantic map of the entire project, it can select the most relevant files for any task, the select strategy operating over a codebase too large to fit in any window. This is exactly the retrieval-and-relevance pattern that powers good RAG, applied to code. The lesson: invest in good indexing so selection can be precise.

What to Take Away

The coding agents prove the four strategies work at scale on hard problems. To apply their lessons: add procedural memory files to give your agents persistent conventions, implement compression for long-running interactions, use just-in-time loading instead of pre-loading everything, and build good indexing so selection can be precise. You do not need to invent these patterns; you can copy them from the most battle-tested context systems in production and adapt them to your own.