context-engineeringlisted

# Context Engineering for Agents Context engineering has displaced prompt engineering as the core skill. Coined by Tobi Lütke, amplified by Karpathy and Willison: it covers everything that enters the context window across a session — system prompt, retrieved content, tool descriptions, history, memory, examples. The framing comes from LangChain's Lance Martin: context engineering is **four operations applied to a finite resource**. ## The four operations ### Write — externalize state outside the context window When state must survive across steps but doesn't need to be visible every step, write it out: - **Scratchpads / working memory:** the agent saves intermediate work to a file or store, reads it back when needed - **Plans:** Anthropic's Research feature has the LeadResearcher save the plan to memory before spawning sub-agents, so the plan survives compaction - **Repo-resident knowledge:** `CLAUDE.md`, `AGENTS.md`, `docs/` directory — version-controlled, human-editable, agent-readable - **Memory store:** Mem0 / Zep / Letta (see `memory-architecture/`) for cross-session persistence **Rule:** if it's not actively driving the current decision, don't carry it in context. Externalize. ### Select — retrieve only what's needed for this step The default is to dump everything potentially relevant. The discipline is to select only what this specific step needs. Three places to apply selection: - **Documents:** RAG over your knowledge base. Standard. - **Tools:** when you