When To Use What — Decision Cheat-Sheet
The near-duplicate choices the exam loves, with a one-line rule for each.
| Choice | Pick the first when… | Pick the second when… |
|---|---|---|
| RAG vs long-context | The corpus is large, changing, or access-controlled | The context is small, stable, and fits the window |
| MCP vs direct function-calling | You want standardized, reusable, scoped tool access at scale | It's a one-off internal call with no reuse |
| Sync vs Message Batches | It's in a user's blocking path | It's bulk and can wait ≤24h (50% cheaper, supports tools) |
| Single-agent vs multi-agent | The work fits one context window | Truly independent subtasks justify the coordination cost |
| LLM-as-judge vs human review | High volume, calibrated rubric, cost-sensitive | High stakes, low volume, or judge can't be trusted yet |
| Canary vs blue-green | You want gradual exposure with fast rollback | You need instant cutover with a warm standby |
| Prompt vs deterministic control | It's guidance the model can reasonably follow | It's a must-hold rule — enforce it in code |
If you remember one thing
Match the mechanism to the constraint, and put must-hold rules where they can be guaranteed — not in a prompt that only asks.
All CCAR-P references
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.