Context engineering

Context engineering chooses and organizes the tokens available to a model for one inference: instructions, tool definitions, examples, retrieved evidence, message history, state, and runtime results.¹

Prompting versus context engineering

Prompt engineering	Context engineering
Writes and organizes instructions	Curates the model’s complete working set
Often changes a prompt template	May retrieve, filter, compact, or remove information on every turn
Asks “How should I say the task?”	Asks “What should the model know right now?”

Tiny example

For a refund question, the useful context may be the system rule, the signed-in user’s region, the current regional policy passage, one order lookup result, and enough message history to resolve “that order.” The entire policy library and complete chat history would add noise.

For a Python authentication change, useful context may be the current route, service interface, user model, token utility, dependency versions, and relevant tests. “Write production-ready authentication code” is an instruction, not a substitute for those repository facts.

FDE note

Treat context as a limited budget. Keep the smallest high-signal set that supports the next decision, and keep authorization outside the model.

Read the full prompt-versus-context lesson for a runnable context-selection example.

Footnotes

1. Anthropic, “Effective context engineering for AI agents”. ↩