Context engineering: provide the evidence a task needs

Context engineering is the work of selecting and maintaining the information available to an AI system during a task. Better context can help, but does not remove model limitations or the need to verify the output.

Inspect the cause before adding more context

A weak answer may reflect missing facts, poor retrieval, conflicting instructions, model limits or an unclear task. Adding more text can also introduce distraction or contradictions. Compare alternatives on representative tasks and retain the evidence from those evaluations.

Five inputs I review

My working checklist covers the task, the audience, approved source facts, constraints and examples of an acceptable result. It is an editorial aid, not a measured prevalence model. Each source should have an owner and review date; examples must not introduce unsupported product claims.

An illustrative brief and response

Suppose a fictional product brief states that setup takes 90 seconds. A valid draft may repeat that supplied claim while marking it for verification. It must not turn it into 90 minutes, invent a customer case study, add a G2 badge or assert an unprovided data-refresh period. Fluent output is not evidence that the brief was followed.

Evaluate the final artifact

Check every factual statement against the supplied sources, distinguish assumptions, and review failure cases. A context library needs maintenance as policies, products and evidence change. Approve output only after factual and task-specific checks; do not treat a polished first draft as a verified result.

Sources and scope

Practitioner analysis by Fernando Angulo. Proposed examples are not measured outcomes.

Anthropic: Effective context engineering for AI agents