Curriculum
CCAF logo

Domain 4 · 20% of the exam

Prompt Engineering & Structured Output

Craft prompts with explicit criteria, apply few-shot prompting, enforce structured output with tool use and JSON schemas, and build validation, retry, batch, and multi-pass review architectures.

  1. 4.1 System Prompts with Explicit Criteria

    Fix a noisy reviewer by replacing vague confidence language with explicit, categorical report-vs-skip criteria and concrete examples — not by telling the model to "be conservative."

  2. 4.2 Few-Shot Prompting

    When instructions alone give inconsistent format or judgment, add 2–4 targeted examples that show the reasoning — the most effective technique for consistency, generalization, and reduced hallucination.

  3. 4.3 Structured Output with Tool Use

    Guarantee schema-valid output with tool_use + a JSON schema — it eliminates syntax errors (not semantic ones) — and design the schema with nullable/optional fields and enum + "other"/"unclear" patterns to prevent fabrication.

  4. 4.4 Validation, Retry & Feedback Loops

    Catch semantic errors that tool_use can't, then retry with the exact error and the original document — but recognize that retry can't conjure information that is simply absent from the source.

  5. 4.5 Batch Processing Strategies

    Route latency-tolerant, non-blocking work to the Message Batches API for 50% savings — but never batch a blocking gate like a pre-merge check, where someone is waiting.

  6. 4.6 Multi-Instance & Multi-Pass Review

    Review generated code with an independent Claude instance that lacks the generator's reasoning context, and split large multi-file reviews into per-file passes plus a separate integration pass.