Domain 2 — Claude Models, Prompting and Context Engineering (13%)

13% of the exam, roughly 8 questions. The Foundations material, asked at organisational scale.

2.1 Model selection as a portfolio decision

At this level you are rarely choosing one model. You are choosing an allocation across a system, and justifying it.

The executor–advisor or orchestrator–worker arrangement is the one to know: a cheaper model carries the bulk of the tokens and escalates only the hard decisions to a frontier model. It is frequently the correct answer to "reduce cost without losing quality", and it is defensible in a way that a flat downgrade is not.

Whatever you choose, the examinable justification is evaluation on your own task. Benchmark claims are not evidence about your workload.

2.2 Context engineering at scale

Context is a budget spent on every request, and at scale it is also a cost centre and a latency contributor. The levers, in the order a professional applies them:

• Order for caching. Stable content first — system prompt, tool definitions, reference material — volatile content last. This is simultaneously the right structure for grounding and the only structure that lets prompt caching work.
• Cache deliberately. Four cache_control breakpoints, ordered tools → system → messages, on the last block that stays identical. Reads cost 0.1×; writes 1.25× at five minutes or 2× at an hour. Cache reads also do not count toward your input-tokens-per-minute limit, so caching buys throughput as well as cost.
• Clear what has stopped informing anything. Context editing does this server-side: clear_tool_uses_20250919 with a trigger threshold, keep for recency, exclude_tools for what must survive, and clear_at_least so each activation frees enough to justify the cache invalidation it causes.
• Isolate. Subagents keep exploration out of the parent's window entirely.
• Offload. Files or a memory tool hold state that does not need to be resident.

Still wrong at Professional level — answering context pressure with a larger context window. Attention dilutes; a bigger window lets the problem grow and raises cost per request while doing it.

2.3 System prompts as governed artefacts

In an enterprise system the prompt is not a string in someone's editor. It is a versioned artefact with an owner, a change process and an eval suite that gates changes to it. The professional-level points: prompts belong in source control; a prompt change is a deployment and gets the same regression run as a code change; and the model version should be pinned so that behaviour changes only when you decide it does.

← Previous: Domain 1 — Solution Design and Architecture  |  Next: Domain 3 — Integration →