Context engineering: choose what AI needs

What belongs in context? #

Context engineering addresses the information available to a model as it works: instructions, selected evidence, conversation history and tool outputs. Anthropic treats this as a broader, ongoing curation problem.

Build a focused context pack #

For a practical task, assemble a small set of clearly labelled material. More text is not a substitute for deciding what matters.

• Task: what is being done now?
• Evidence: which current, permitted sources support it?
• State: which earlier decisions are still relevant?
• Constraints: what is out of scope or unknown?

Current task: compare two workshop proposals.
Approved sources: proposal A, proposal B and the current budget.
Earlier decision: online delivery only.
Unknown: final participant count.
Ignore: superseded draft budgets.

Review what remains #

Remove superseded instructions and duplicate material. Preserve the original source when a summary could lose a qualification. Treat retrieved documents as evidence to inspect, not authority to issue new instructions.

Sources & further reading

• Anthropic — Effective context engineering for AI agents ↗