This domain is 15% of the exam. All five weightings:
The smallest domain by weight, but its ideas show up inside almost every scenario question. It asks how you keep a long-running system accurate as context grows.
Lost in the middle
Claude attends more reliably to the start and end of a long context than to the middle. A critical constraint buried deep in the middle of a huge window is processed less reliably than the same text near the top. A bigger window does not fix this — restructuring where information lives does.
- Keep a persistent "case facts" block (the customer, the order, the constraints) at a stable, prominent position.
- Put key findings at the top of aggregated inputs, with detail after.
- For codebase exploration, write findings to a scratchpad file instead of relying on the conversation to carry them.
Compaction and summarisation have a cost
Compressing a long conversation saves space but can drop precise details — numbers, identifiers, dates. When exact values matter, preserve them explicitly (in a facts block or a file) rather than trusting a summary.
Escalation and reliability
- Escalate to a human on defined triggers — policy exceptions, an explicit request for a person, repeated failure — not on Claude's own confidence score.
- Ambiguous or multi-issue requests should be broken down so each issue is handled and none silently dropped.
Provenance and uncertainty
When results from several sources are merged, keep track of where each claim came from and say where coverage is incomplete. Conflicting sources should be reported as conflicts with attribution, not blended into one confident statement. Dates matter: two numbers from different periods are not a contradiction.
Prompt caching and cost, at exam depth
You only need the idea: when the same large context is reused across real-time calls, prompt caching cuts cost (up to 90% on cached input) without adding the delay of batch processing. You do not need the internals.
Build to learn it
Take any long-running exercise from the study plan, deliberately place a key constraint mid-context, and see it get missed — then restructure with a facts block and a scratchpad and watch the behaviour improve.