Domain 4 — Prompt Engineering & Structured Output

CCA Foundations course · Page 5 of 8 · ← Back to all courses · Weight: 20%. Mark complete at the bottom to advance.

4.1 — Explicit Criteria Over Vague Instructions · Core

• Use categorical inclusion + exclusion lists ("flag X; never flag Y") — not "be conservative".
• False-positive trust cascade: one noisy category undermines trust in accurate ones → disable the bad category first, fix its prompt separately.
• Severity calibration with concrete code examples per level.

4.2 — Few-Shot Prompting · Core

• Show the reasoning, not just the answer.
• 2–4 examples, on ambiguous cases only.
• Five useful types: ambiguous · format · positive-negative · varied-structure · informal.

4.3 — JSON Schemas & Structured Output · Core

• Nullable fields + remove them from required → prevents the model fabricating values to satisfy the schema.
• Add "unclear" / "other" enum values for ambiguous cases.
• Self-correction fields: extract stated_total and calculated_total, plus a conflict_detected boolean.

4.5 — Message Batches API · Core

• 50% cost savings, 24-hour window, NO latency SLA, NO multi-turn tools.
• custom_id correlates requests and enables failure resubmission.
• Use the synchronous API if anyone is waiting on the result, or if tool loops are needed.

4.6 — Prompt Chaining & Multi-Instance Review · Core

• Attention dilution beyond ~14 files → split into per-file passes + a cross-file integration pass.
• Pydantic retry only for format/structural errors — NOT when data is simply absent from the source.

⚠ Often-missed — Multi-Instance Verification Architecture (VIP routing) · Gap

• Pass 1 generates; Pass 2 is an independent instance that verifies and outputs a confidence score.
• Route on that score: send / human_review / regenerate.
• Use tiered thresholds per customer segment.

Exam reflexes for Domain 4

• "Absent data causing retry loop" → nullable fields + remove from required.
• "Conflicting / can't-sum totals" → self-correction fields (calculated_total, conflict_detected).
• "High volume, can wait" → Batches API (50% / 24h / no SLA / no tools).
• "Self-review bias" → independent instance, zero prior context.
• "Misrouting" → fix descriptions, not few-shot (Domain 2).