Context engineering: give AI the right working information

More than writing a prompt

Prompt wording describes the task. Context engineering concerns the whole working package available to the model: instructions, the user’s question, retrieved evidence, relevant conversation history and tool results. The objective is sufficient, relevant information at the moment it is needed. Adding every available document can introduce noise, contradiction and unnecessary cost.

Choose and separate the inputs

Keep business instructions distinct from source documents. A retrieved email is evidence to interpret, not authority to rewrite the application’s rules. Identify current material, preserve useful source references and reserve room for the model’s response. Context has a finite size, so selection and summarisation are design decisions rather than housekeeping.

Use the supplier scenario to see the difference

In our existing supplier-delay example, a useful working package could contain the supplier’s message, the authorised order record and the team’s escalation policy. A full customer archive would add little. A summary saying only “delivery delayed” would omit the order and date that make the task actionable. The design question is what the reviewer needs to reach a justified conclusion.

Review the package, not just the answer

When a response fails, inspect what was actually supplied. Was a relevant passage missing? Did a stale record appear authoritative? Was a tool result truncated? Keep a repeatable evaluation set and compare changes to retrieval, instructions and context selection. This makes improvement more systematic than repeatedly rewriting the same prompt.

Microsoft sources and further reading

• Microsoft Architecture Center: context management in RAG prompts