Domain 2: Claude Models, Prompting & Context Engineering

Model selection trade-offs (2.1)

Match relative tiers (fast/cheap, balanced, highest-capability) to difficulty and risk. Require accuracy floor + latency/cost SLAs at expected volume. Revisit when evidence moves.

• Smallest tier that clears floor + SLAs
• Escalate hard/high-risk — do not over-provision bulk
• Never always-pick-largest by default

System prompts, templates, guardrails (2.2)

• Explicit criteria, formats, boundaries > vibes
• Templates for multi-step consistency
• Prompts = soft; must-always = deterministic tool-path controls
• Priority when instructions conflict; eval soft vs hard separately

Prompt engineering techniques (2.3)

• Zero-shot: clear, well-specified tasks
• Few-shot: format drift / edge cases
• CoT: multi-step reasoning when it earns tokens
• Lightest technique that fixes the named failure — no default CoT on hot simple paths

Context windows & tokens (2.4)

Budget by layer. Verbatim vs summarize. Persist IDs outside narrative summaries. Trim tool results with retention rules. Watch lost-in-the-middle.

• No dump-all corpora
• Live-state / facts block for transactional truth
• Structured tool adapters beat raw 50k log pastes

Prompt reuse (2.5)

• Cache stable prefixes
• Modularize shared instructions (versioned)
• Skills isolate verbose reusable workflows
• Keep live context lean — do not inline full skill text every turn