Weight: 20%. About 12 of the 60 questions.
| Topic | What it tests |
|---|---|
| 4.1 Explicit criteria | Precision, false positives, defining what to report and what to skip |
| 4.2 Few-shot prompting | Targeted, diverse, balanced examples. Showing reasoning |
| 4.3 Structured output | Tool use plus JSON schema, tool_choice, nullable fields, enums, strict mode |
| 4.4 Validation & retry | Retry with specific errors, semantic checks, when retrying won't help |
| 4.5 Batch processing | Message Batches API: cost, latency, custom_id, limits, SLAs |
| 4.6 Multi-pass review | Independent instances, per-file plus integration passes, verification passes |
Mental model
- Precision comes from explicit, testable criteria with examples, not from "be careful" or "be conservative".
- Schemas guarantee structure, not truth. Validate meaning separately.
- A model reviewing its own work in the same context is a weak reviewer. Use an independent instance.
Updated 2026-10-28