Skip to content
Nerdy beta

Context Engineering

AI Practice

A brilliant prompt aimed at a clueless model produces nothing useful. Context engineering is the skill of deciding what the model needs to know and, just as critically, what it doesn't.

Supplying the right background information to an AI system, in the right format, so it can do useful work. A perfect prompt with no context produces generic output. Context engineering is the discipline of deciding what the model needs to know: relevant documents, examples of good output, constraints, domain-specific terminology, and the standards you’re working against. It’s as much about what you leave out as what you include; flooding the context window with irrelevant information degrades quality just as surely as providing too little.

The difference is concrete. A team asked a model why their delivery had slowed and fed it everything: the full quarterly board export, every standup Slack thread from the quarter, and the org chart. Buried in that pile, the model did the only safe thing and generalized — “communication could be improved, and more automation may help.” True of every team that has ever shipped software, useful to none.

The curated set answered the actual question. They cut the standup chatter and the org chart, which carried no signal about flow. They kept the cycle-time trend and the current work-in-progress count. They added two things the dump never had: the team’s working agreement, and two past decision memos the team was proud of, so the model knew what “a decision we’d act on” looks like here. This time it named the bottleneck — work stacking up at the handoff to the platform team — and proposed a WIP limit to make it visible. Same model, same question; the only variable was what went into the context.

Context engineering is where prompt engineering meets workflow design. A prompt framework like CO-STAR or STOKE will remind you that a request needs context; context engineering is the harder work of deciding which context earns its place in the window. In our Effective AI workshop, it’s the bridge between writing good prompts and building compounding setups. A Claude Project with well-curated context files, custom instructions, and representative examples is context engineering made durable: the context persists across conversations, improves over time, and compounds in value.

Resources