Claude Models, Prompting & Context Engineering

Selecting models on their trade-offs, designing system prompts and guardrails, applying few-shot and chain-of-thought techniques, and managing context and token usage with caching, compaction and Skills.

1. 2.1Selecting Claude ModelsPick the model tier that meets the quality bar inside the cost and latency budgets, route by task difficulty where it pays, and treat every model change as a release that needs an evaluation.
2. 2.2System Prompts, Templates and GuardrailsWrite system prompts with a clear role, explicit constraints and an output contract, build them as templates whose fixed rules cannot be overwritten, and enforce anything consequential outside the prompt.
3. 2.3Prompt Engineering TechniquesKnow when zero-shot is enough, when examples teach better than instructions, and when to ask for step-by-step reasoning, and add each technique only when an evaluation shows the need.
4. 2.4Context Windows and Token UsageDecide what goes into the context window and when, using monolithic, progressive, retrieval and compaction strategies, and prevent the two different ways a request can run out of room.
5. 2.5Prompt Reuse: Caching, Modular Prompts and SkillsReuse prompts three ways: cache stable prefixes to cut cost and latency, share prompt fragments as a modular library, and package governed procedures as Skills, and know which one each situation needs.