Weight: 13%. About 8 of the 63 exam questions.
This domain covers choosing Claude models and steering them with prompts and context.
| Objective | What it tests |
|---|---|
| 2.1 Model selection | Capability vs latency vs cost; tier mixing; platforms and regions; migrations |
| 2.2 System prompts, templates, guardrails | Structure, scope, untrusted data, templates with variables |
| 2.3 Prompting techniques | Zero-shot, few-shot, chain-of-thought, extended thinking |
| 2.4 Context windows and tokens | Long documents, compaction, tool output size, token budgets |
| 2.5 Prompt reuse | Prompt caching, modular prompts, Skills |
Mental model
- Pick the smallest model that meets the quality bar on your evaluation set.
- Give Claude what a human expert would need: audience, purpose, constraints, examples.
- Context is a budget: keep it relevant, stable where possible (for caching), and loaded on demand where large.
Exam heuristics
- High-volume simple tasks → Haiku-class; complex, low-volume, high-stakes → Opus-class; Sonnet-class is the usual balanced default.
- A cache that never hits almost always means the prefix changes (timestamp first, unstable tool order, per-user data before static content).