context-engineering

# Context Engineering ## Purpose Decide what an LLM sees, and what it does not. Context is a finite budget with a nonlinear cost curve: more context is not more capability, and past a point it is actively less. ## When to Use - Building an agent or assistant that runs over many turns. - A model that performs worse as the conversation gets longer. - Managing tool results, retrieved documents, or long files. - Reducing token cost in a long-running system. ## Capabilities - Context budgeting and allocation. - Retrieval: putting the right thing in, not everything. - Compaction and summarization of history. - Memory: what persists across sessions, and in what form. - Tool-result pruning and progressive disclosure. ## Inputs - The task, and what information it genuinely requires. - The context window and its cost. - The failure symptom, if the system is degrading over long runs. ## Outputs - A context budget with an allocation per component. - A compaction strategy that preserves what matters. - Measurably better performance on long tasks. ## Workflow 1. **Budget the window explicitly** — System prompt, tools, retrieved context, history, and the response all compete for the same space. Decide the allocation rather than letting history consume everything. 2. **Retrieve narrowly** — Ten highly relevant chunks outperform a hundred marginally relevant ones. Irrelevant context does not sit inertly; it distracts. 3. **Compact, do not truncate** — Dropping the oldest turns los...