Context engineering in 2026: treating the window as a budget, with six techniques, two metrics and a one-week rollout plan.

Context Engineering 2026: Six Things That Matter More Than Prompts

Prompt engineering optimises the sentence you write. Context engineering optimises everything the model sees: system instructions, retrieved documents, tool schemas, history and memory, plus how they compete for a fixed window.

Treat the window as a budget, not a container. A workable split: system instructions 5-10 percent, tool definitions 10-15, retrieved content 30-40, history 20-30, and 15-20 reserved for output. The reserve is the most commonly ignored line; a full window produces truncated or suddenly terse answers.

Six techniques by payoff. Structure the system prompt into role, task, constraints, output format and examples with explicit delimiters. Retrieve dynamically: filter by similarity, rerank, inject only the top 3-5 chunks, and if nothing clears the threshold tell the model there is no material rather than filling with noise. Layer memory into working, session and long-term, with only working memory inline. Roll old turns into summaries that preserve constraints and decisions. Trim tool schemas to the current step, since twenty definitions cost tokens and raise selection errors. Put critical constraints at the end, where attention is strongest, not buried in the middle.

Two metrics weekly: context utilisation, held at 60-75 percent, and retrieval precision, trending up. Healthy engineering means the same quality with less context.

One-week rollout: rolling summary, then tool trimming, then constraint reordering, then instrument both metrics, then A/B on an eval set. Change one thing at a time or you cannot attribute a quality drop.

Three misconceptions: bigger windows do not mean the model uses the middle of them; context engineering is not just better wording; and this is system design, not a trick, because RAG, agent memory and routing all face the same question of what enters the window.