Prompt Engineering & Structured Output
Prompt engineering is where architectural intent becomes reliable model behaviour. This domain — 20% of the CCAR-F blueprint — covers the techniques that turn Claude from a capable chat model into a dependable component in a production pipeline: writing **system prompts with explicit criteria**, teaching format and judgement with **few-shot examples**, guaranteeing schema-compliant **structured output** via tool use, and wrapping it all in **validation-retry loops**. Beyond single calls, you will learn to scale and harden extraction workflows: when to trade latency for cost with the **Message Batches API**, and how **multi-pass, multi-instance review** overcomes attention dilution and self-confirmation bias. The exam tests judgement — knowing which technique fixes which failure mode — so every lesson pairs the concept with the traps that distinguish a correct answer from a plausible distractor.
6 lessons in this domain
- 4.1
System Prompts
The single biggest lever in a production system prompt is replacing ambiguous instructions with explicit, categorical decision boundaries that tell Claude exactly what to flag and what to skip.
6 min read - 4.2
Few-Shot Prompting
Few-shot examples are the most reliable way to get consistent formatting and judgement from Claude — often more effective than adding another paragraph of instructions.
6 min read - 4.3
Structured Output
Tool use with a JSON schema is the reliable way to guarantee syntactically valid, schema-shaped output from Claude — but it prevents syntax errors, not semantic ones.
6 min read - 4.4
Validation & Retry Loops
A validation-retry loop turns extraction failures into self-correcting workflows by feeding Claude the original document, its failed output, and the specific validation error.
7 min read - 4.5
Batch Processing
The Message Batches API trades latency for roughly 50% cost savings — ideal for latency-tolerant work, wrong for anything a person is actively waiting on.
5 min read - 4.6
Multi-Pass Review
Independent Claude instances review output without generation bias, and splitting large reviews into per-file then cross-file passes overcomes attention dilution.
6 min read