Start with the consuming task
A model receives an information environment: instructions, a request, selected sources, conversation history, tool definitions, and sometimes a record of earlier work. Context engineering concerns how that environment is assembled and maintained for the task at hand.
Begin by asking who or what will use the information next. A person investigating a disagreement may need original passages and their surrounding sections. A system extracting an invoice may need a small set of fields and validation rules. A broad answer is not always the useful next operation.
Make the first route inspectable
Choose a bounded source collection and a few realistic tasks. Build a route that makes it clear what was searched, what matched, which material was selected, and what entered the next step. A simple starting point makes it easier to see which additional machinery would solve an actual problem.
For an exact identifier, literal search and original source passages may be enough. A request expressed in different vocabulary may justify semantic retrieval. The useful baseline depends on the task. Preserve the intermediate results so a fluent final answer does not become the only indication that the information was found.
Preserve meaning and provenance
Information changes as documents are parsed, split, indexed, ranked, selected, and compressed. Each transformation can affect what the next participant is able to know. A short excerpt may omit an exception. A summary may merge a proposal with an accepted decision. An old policy may look current after its date disappears.
Keep useful links to original sources, versions, and surrounding context. Distinguish instructions, external evidence, generated interpretations, and unresolved hypotheses. Decide what belongs in temporary working context and what deserves durable storage. A retained statement should carry enough information to understand when it applies.
Use failures to locate the next experiment
When an answer fails, inspect the route before changing the prompt. Was the source absent from the collection? Did parsing lose a relevant table? Did retrieval miss a passage? Did selection exclude it? Did compression remove the qualification? Or did the model misuse information that was present?
These are different problems. Change one consequential choice and compare the result on cases that expose it. Keep source snapshots and surrounding conditions comparable when attributing an improvement. Examine what the change removes as well as what it adds, including review effort, latency, and cost.
Evaluate the assembled context
Finding useful material does not guarantee that the actual model input is adequate. Check what survives selection, ordering, formatting, and compression. Then evaluate whether the resulting workflow completes its intended job. Retrieval quality and task success inform each other, but they answer different questions.
Make interventions available where they resolve a meaningful uncertainty. A person may need to expand a passage, correct an interpretation, dispute a relevance judgment, or change source scope. Preserve the reason for that intervention. It can improve the current task and suggest a policy to test across future work.