Context engineering is the practice of selecting, structuring, retrieving, and delivering the information, memory, tools, constraints, and examples an AI system needs to perform a task well.
Imagine asking an AI agent to redesign a billing flow.
The model might need the current screens, design system, account permissions, pricing rules, recent research, analytics, technical constraints, and the user’s requested change.
Context engineering is the system that decides which of those pieces are relevant, how they are represented, and when they are supplied.
Prompt engineering focuses on how the instruction is expressed.
Context engineering includes the prompt but also retrieval, memory, source selection, tools, state, and permission-aware filtering.
More context is not always better. Irrelevant data can distract the model.
Outdated facts can create confident mistakes.
The system should respect product, organization, role, and permission boundaries.
Important claims should remain connected to their source where possible.
When sources disagree, the system should surface or resolve the conflict rather than silently combining them.
Existing-product design depends on details that cannot be inferred from a generic model: component usage, product rules, previous decisions, user behavior, and implementation constraints.
Strong context turns generation from “make a plausible screen” into “make a candidate that fits this product.”
Relevance and prioritization matter.
Stale decisions should not silently override current product truth.
Context systems need clear access boundaries.
Context engineering asks: what does the AI need to know, from which sources, at this exact moment, to make a useful and trustworthy decision?
AI agent · AI prompt · Product context · Multimodal AI
Read Context Is the New Canvas for the product-design implications of persistent context.