Prompt Engineering

Question 1

A model given a strict JSON-format constraint returns perfectly valid JSON that is factually wrong. What does this demonstrate?

Accepted Answer

Format constraints govern structure, not content truth — Output-format constraints (a JSON schema, XML tags, etc.) control the shape of the response, not its truthfulness — a model can be perfectly well-formatted and still hallucinate the content.

Question 2

Prompt engineering steers a model by editing its ___ rather than its weights.

Accepted Answer

context — Prompting changes the model's behavior via the context it reads at inference time — the cheapest, fastest lever before retrieval or fine-tuning.

Question 3

What distinguishes few-shot prompting from zero-shot prompting?

Accepted Answer

Few-shot includes labeled demonstrations in the prompt; zero-shot does not — In-context examples teach the model the desired format and decision boundary without any weight updates (in-context learning).

Question 4

Chain-of-Thought (CoT) prompting improves performance on…

Accepted Answer

Multi-step reasoning tasks — Asking the model to 'think step by step' surfaces intermediate reasoning, raising accuracy on math/logic/multi-hop problems.

Question 5

Adding an OUTPUT-FORMAT constraint (e.g. 'respond with JSON') mainly…

Accepted Answer

Makes output machine-parseable — Constraining the format makes responses predictable and easy to parse downstream — critical for production pipelines.

Question 6

The system/role message is used to…

Accepted Answer

Set the model's persona and task rules — The system prompt establishes who the model is and how it should behave, framing all subsequent interaction.

Question 7

Which prompting approach sends only the task instruction with no demonstrations?

Accepted Answer

Zero-shot prompting — Zero-shot gives only the instruction; capable models often succeed, but few-shot examples help when format or edge cases matter.

Question 8

When would you prefer few-shot prompting over FINE-TUNING?

Accepted Answer

Few examples or a changing task — Prompting fits low-data, rapidly-changing tasks; fine-tuning suits large stable datasets or behaviors hard to achieve via prompts.

Question 9

When would you skip chain-of-thought and rely on output constraints instead?

Accepted Answer

When the output must be a terse, parseable single token — CoT costs tokens and latency — the wrong trade when the answer is a discrete token a schema can enforce. Constrain the output and skip the reasoning steps for terse, parseable results.

Question 10

Using DELIMITERS (e.g. triple backticks, XML tags) in prompts helps by…

Accepted Answer

Separating instructions/data — Delimiters mark boundaries so the model distinguishes the instruction from the text to process, improving reliability and safety.

Key insight

Assemble a production prompt

Test your knowledge

A model given a strict JSON-format constraint returns perfectly valid JSON that is factually wrong. What does this demonstrate?

The short version

In more depth

Prompt Engineering

Building blocks

Interview Questions

References & Further Reading

Related topics

Love this visualization?