Prompt Engineering — Glossary

Prompt Engineering & Structured Output · 20% of the exam. 10 terms.

The top-level instruction set establishing role, constraints, and behavior for the whole conversation.

Exam context: Distinct from a per-turn user message — the exam tests knowing which behaviors belong in which layer.

Wrapping distinct prompt components — instructions, context, examples — in XML-like tags so the model reliably tells them apart.

Exam context: Especially valuable for long prompts mixing several kinds of content in one request.

Including 2–5 worked examples in the prompt to show the exact input/output pattern wanted.

Exam context: More reliable than instructions alone when the desired output format is precise and easy to demonstrate but hard to describe.

Asking the model to reason step by step before answering, improving accuracy on multi-step problems at the cost of latency and tokens.

Exam context: Not free — the exam expects you to reserve it for genuinely multi-step reasoning, not simple lookups.

Seeding the start of the assistant's response to force a format — e.g. starting with { to guarantee JSON — or skip preamble.

Exam context: A fast, cheap way to eliminate an entire class of formatting failures without extra validation logic.

Constraining the model's response to a defined schema so downstream code can parse it reliably.

Exam context: The exam distinguishes 'ask nicely for JSON' from actually constraining and validating the output.

Programmatically validating a structured response against its schema and re-prompting with the specific error on failure.

Exam context: The retry prompt must include what specifically failed — a bare 'try again' rarely fixes the same mistake twice.

Using an asynchronous batch API to process many independent prompts at lower cost when latency isn't critical.

Exam context: Correct choice for large, non-interactive workloads; wrong choice whenever a user is waiting on the response.

Splitting generation and review into separate calls so the reviewing pass isn't anchored by the generation pass's own reasoning.

Exam context: A single call asked to 'write and then check your own work' tends to rubber-stamp itself — the exam flags this.

Common mistakes like negative-only instructions ('don't do X') with no positive alternative, or one prompt overloaded with unrelated tasks.

Exam context: When an option only tells the model what not to do, look for a sibling option that also states the desired behavior.