Context Engineering: How We Work Around the Goldfish Problem

Context engineering determines what information an LLM uses, and when, impacting its usefulness. Summarization, the common approach, compresses context by discarding details, causing re-retrieval in debugging or refinement. This summarization struggles with iterative tasks, leading agents to re-read files, wasting tokens and affecting reliability. Other approaches like RAG, caching, agents, and fine-tuning exist, each with varied trade-offs in terms of cost, speed, and accuracy. Good context engineering requires understanding code structure, dynamic decision-making, and precision. Poor context engineering results in increased costs, slower speeds, and unreliable outputs, especially in AI-assisted coding. Users should watch for re-retrieval, match tools to tasks, ask about context strategies, keep sessions focused, and provide explicit context. The author is developing tools to improve context engineering by understanding code structure.