Domain 4

Prompt Engineering & Structured Output

6 build exercises to practice the concepts in this domain.

4.1Intermediate45 minutes

4.1 — Build an Explicit Criteria Code Review Prompt

What you will learn

Vague bars like "be conservative" or "high confidence only" do not hold in production.
Name the categories to flag and the categories to skip.
Calibrate severity with code examples, not adjectives.
Measure false positives, and turn off a category that keeps wasting reviewer trust.
Write the criteria first. Use confidence routing only after the criteria are explicit.

4.2Intermediate45 minutes

What you will learn

Reach for examples when formatting drifts, a judgement call is ambiguous, or a present field comes back empty.
Each example should include the reasoning, not only the input and the output.
Two to four examples, aimed at the failing cases, are enough.
Examples fix format and judgement. A schema change or a validation loop fixes a different failure.
Measure empty fields and format consistency before and after.

4.3Intermediate45 minutes

What you will learn

Optional or nullable fields stop the model from inventing a missing value.
tool_choice auto, any, and a forced tool name are three different guarantees.
A tool call removes syntax errors. It does not remove wrong values.
Unclear enums need an other value plus a detail string, and formats should be normalised in the schema.

4.4Advanced60 minutes

What you will learn

A retry sends the original document, the failed extraction, and the specific validation error.
Format, structure, and arithmetic can be fixed. A fact that was never in the document cannot.
A schema can carry calculated_total, stated_total, and conflict_detected.
Track the error pattern and which findings reviewers dismiss.
tool_use already removed syntax errors. The loop is for semantic checks.

4.5Intermediate45 minutes

What you will learn

A blocking user-facing call stays synchronous. A latency-tolerant job can use the batch API.
custom_id ties each result back to its request.
Resubmit only the failed items, with a change aimed at the failure.
The processing window is up to 24 hours, so submission time has to fit the SLA.
Test the prompt on a sample before you submit the full batch.

4.6Advanced60 minutes

What you will learn

A model that reviews its own output in the same session still holds the reasoning that produced it.
Review each file on its own, then run a separate pass over the findings.
Uneven depth, missed middle-file bugs, and contradictory flags are attention problems.
Route low-confidence findings to a person, using a threshold you checked on labelled examples.
A raw confidence number is not a calibrated threshold.