Domain 2: Claude Models, Prompting & Context Engineering

Models & prompting terms: tier trade-offs, system prompts, PE techniques, context budgets, caching, and Skills.

• Model tier trade-offs

  Selecting among relative capability tiers (fast/cheap, balanced, highest-capability) by matching task difficulty and risk to an accuracy floor plus latency and unit-cost SLAs at expected volume.

  Exam Always-pick-largest is usually wrong — choose the smallest clearing tier.

  See also: 2.1 Model selection

• System prompt criteria

  Explicit acceptance criteria, required output formats, and boundary conditions in the system prompt — stronger than vague style advice alone.

  Exam Vague encouragement without criteria/formats fails exam-style prompts.

  See also: 2.2 System prompts & templates

• Deterministic vs prompt guardrails

  Probabilistic prompt guidance steers behavior most of the time; deterministic controls (tool-path blocks, allowlists, policy checks) enforce must-always / must-never outcomes.

  Exam Hard safety rules need enforcement beyond the system prompt.

  See also: 2.2 System prompts & templates

• Zero-shot prompting

  Instructing the model with a clear task and rubric without demonstrations — appropriate when the specification alone is enough.

  Exam Prefer zero-shot for clear, well-specified tasks before heavier techniques.

  See also: 2.3 Prompt techniques

• Few-shot prompting

  Providing a small set of high-quality exemplars so the model learns desired format or edge-case behavior.

  Exam Few-shot targets format drift and edge failures — not every problem.

  See also: 2.3 Prompt techniques

• Chain-of-thought (CoT)

  Asking for explicit multi-step reasoning (or a scratchpad) when those extra tokens improve measured accuracy enough to justify latency and cost.

  Exam Do not default CoT on simple latency-sensitive tasks.

  See also: 2.3 Prompt techniques

• Context budget

  An explicit allocation of tokens across system instructions, retrieval, tool results, and history — with headroom for the response — instead of unbounded dumps.

  Exam Dumping entire corpora into context is a common wrong answer.

  See also: 2.4 Context & tokens

• Lost-in-the-middle

  A position effect where critical facts buried in the middle of a large context are easier to miss — mitigated by short live-state blocks and trimmed history.

  Exam Place must-follow constraints and IDs carefully; avoid giant undifferentiated pastes.

  See also: 2.4 Context & tokens

• Prompt caching

  Reusing a stable prompt prefix across requests so unchanging tokens are not fully re-billed/recomputed each turn when the platform supports it.

  Exam Cache stable prefixes; do not treat volatile ticket text as the cache key content.

  See also: 2.5 Prompt reuse

• Modular prompts & Skills

  Versioned shared instruction modules plus Skills (isolated reusable workflows) that keep live turns lean — avoiding full inline playbooks every message.

  Exam Prefer modules/Skills/caching over pasting verbose skill text every turn.

  See also: 2.5 Prompt reuse