Context Engineering: The Discipline Formerly Known as Prompting

"Prompt engineering" always undersold the job. The instruction paragraph is maybe 5% of what the model sees; the other 95% — retrieved documents, tool results, history, schemas — is assembled by your code, and that assembly is where quality lives or dies. The field has started calling this context engineering, and it deserves the engineering treatment: budgets, structure, and measurement.

Budget tokens like you budget performance

Every context has a spending plan. Ours, for the support copilot:

system + tools:        1,200 tokens  (cached, stable-first)
user profile + state:    400
retrieved knowledge:   3,000  (top-5 chunks, headers included)
conversation window:   2,400  (recent turns + rolling summary)
current message:         ~200
response reserve:      1,800
────────────────────────────
budget:                9,000 of 128K available

Yes, the window fits 128K. No, you shouldn't use it: cost scales linearly, latency grows, and — the finding that keeps replicating — models attend worse in bloated contexts. Middle-of-context information gets lost; irrelevant chunks actively degrade answers by offering plausible wrong material. A disciplined 9K context beats a lazy 60K one on our evals, every time we've tested it.

Structure for salience

Order and framing are levers, not aesthetics: stable content first (cache-friendly), the most relevant retrieved chunk closest to the question, explicit section delimiters the model can navigate (<knowledge>, <history>), and a one-line task restatement at the very end — the position models attend to most. When an answer misuses context, the fix is usually placement, not more instructions.

Measure usage, not just outcomes

The context-engineering equivalent of a profiler: for each eval case, check which context sections the answer actually drew from (citation tracking makes this cheap). We consistently find dead weight — a "helpful" metadata block that no answer ever referenced, costing 600 tokens per call. Delete, re-run evals, confirm no regression, keep the savings. Context, like code, accumulates cruft that only measurement reveals.

The framing that stuck with our team: the model is a brilliant contractor who reads your brief once, at speed. Everything depends on what you put in the brief — and briefs are written, structured, budgeted, and revised. That's not prompt whispering. That's engineering.