Task 3.2: Describe prompt engineering techniques and how they drive better results.

• AFew-shot prompting with varied example emails and answers

  Correct

  Few-shot prompts include examples of good responses; they are often used to regulate output format, phrasing and patterns, which is exactly the inconsistency here.

• BZero-shot prompting with a shorter, simpler instruction

  Incorrect

  Zero-shot means no examples; removing words from the instruction does not show the model the expected format.

• CRole prompting that makes the model a logistics expert

  Incorrect

  A persona does not demonstrate the seven categories or the exact output format.

• DAsking the model to explain its reasoning before answering

  Incorrect

  Explaining reasoning helps with complex reasoning tasks; here the problem is format and category consistency, and extra text makes the output longer.

Examples teach patterns that instructions alone leave open. Keep examples varied and consistently formatted, and keep the clear instructions too.

• AOne-shot prompting with a single final answer as the example

  Incorrect

  One example of a final number shows the format, not the steps, so the middle of the calculation stays unguided.

• BChain-of-thought prompting that asks it to show each step

  Correct

  Chain-of-thought prompting has the model work through the intermediate steps of a problem; explaining its reasoning can improve accuracy on challenging, multi-step queries.

• CRole prompting that tells the model it is an accountant

  Incorrect

  A persona does not make the model show and check each step of the arithmetic.

• DPrompt chaining through a series of unrelated questions

  Incorrect

  Unrelated questions do not build on each other, so they cannot guide the middle steps of one calculation.

For multi-step reasoning, ask for the steps. Chain-of-thought makes the model reason before it concludes, and the steps can be checked.

• ARole prompting, giving it a pediatric nurse persona

  Correct

  Role prompting tells the model to take on a person or role, which helps it get into the right "mindset" and steers its style, tone and focus.

• BFew-shot prompting with examples of the hospital's medical codes

  Incorrect

  Examples of medical codes would teach a code format, not a warm voice for parents.

• CRaising the maximum number of output tokens

  Incorrect

  Max output tokens sets how long the answer may be; it does not change its tone.

• DZero-shot prompting with only the question

  Incorrect

  A bare question leaves the tone to the model's default style, which is the problem to solve.

Assigning a persona is the quickest way to set tone and audience; combine it with clear instructions about what to cover.

• AReAct prompting, which reasons, acts and observes

  Correct

  ReAct (reason and act) lets the model reason about a query and take actions such as calling a tool, observing the result and reasoning again until it can answer, the core loop of AI agents.

• BChain-of-thought prompting on its own, with no tools

  Incorrect

  Chain-of-thought produces reasoning steps, but on its own it takes no actions such as running a search.

• CFew-shot prompting with several sample answers

  Incorrect

  Sample answers show the expected output; they do not let the model call a search tool and react to what it finds.

• DRole prompting that casts it as a librarian

  Incorrect

  A persona does not let the model act and observe.

ReAct adds acting and observing to reasoning, which is what turns a model into an agent that uses tools.

• AOne-shot prompting with one sample translation first

  Incorrect

  A single example helps when the output must follow an unusual pattern; a plain translation needs no pattern shown.

• BChain-of-thought prompting that explains each step

  Incorrect

  Step-by-step reasoning helps complex, multi-step problems, not a one-sentence translation.

• CPrompt chaining that drafts, reviews and polishes

  Incorrect

  Chaining splits a complex task into a sequence of prompts; translating one sentence is a single, simple step.

• DZero-shot prompting with a clear instruction

  Correct

  Zero-shot prompts give the task with no examples; for simple, common tasks such as translating a sentence, a clear instruction is usually enough.

Start with the simplest prompt that works; add examples, reasoning or extra steps only when the task needs them.

• AOne-shot prompting

  Correct

  One-shot prompting gives the model a single example of the desired response, which is enough when one example shows the pattern.

• BZero-shot prompting

  Incorrect

  Zero-shot gives no example at all, leaving the unusual pattern to description alone.

• CReAct prompting

  Incorrect

  ReAct interleaves reasoning with tool actions; formatting tags does not need tools.

• DRole prompting

  Incorrect

  A persona shapes tone and focus; it does not show the tag pattern.

Zero, one or few examples: use as many as it takes to make the pattern unambiguous, and no more.

• AFew-shot prompting with one example of the whole result

  Incorrect

  One example of the full output does not make a three-stage task more reliable in a single prompt.

• BRole prompting that casts the model as a senior copywriter

  Incorrect

  A persona can improve the email's tone but leaves the three stages tangled in one prompt.

• CPrompt chaining, passing each step's output to the next

  Correct

  Breaking a complex task into subtasks and chaining prompts, where one prompt's output becomes the next one's input, makes each step simpler and easier to check.

• DA larger maximum output token limit

  Incorrect

  More output room does not fix a prompt that asks for too much at once.

Complex jobs work better as a chain of simple prompts than as one long prompt.

• AIn system instructions that apply to the whole conversation

  Correct

  System instructions guide the model's behavior across a session, such as language, persona, tone and topics to avoid, separately from each user prompt.

• BIn a few-shot example added to the first user message only

  Incorrect

  Examples in one message can fade in long conversations; standing rules belong in system instructions.

• CIn the stop sequences set on each request to the model

  Incorrect

  Stop sequences are a generation parameter, not a place for instructions; they cannot set language, topics or tone.

• DIn a lower top-P value set on each request to the model

  Incorrect

  Top-P controls sampling randomness, not language, topics or tone.

Rules that should hold for every turn are system instructions; task details belong in each prompt.

• ASupervised fine-tuning of the model

  Incorrect

  Tuning changes the model's weights rather than improving the prompt, and needs a tuning job and deployment.

• BThe data-driven prompt optimizer

  Correct

  Agent Platform's prompt optimizer can improve prompts automatically; its data-driven mode uses labeled examples to optimize the instructions for a target model.

• CModel Monitoring

  Incorrect

  Model Monitoring tracks drift in deployed models; it does not rewrite prompts.

• DGrounding with Agent Search

  Incorrect

  Grounding connects answers to a data store; it does not optimize the prompt's instructions.

Using AI to improve prompts, sometimes called metaprompting, is built into Agent Platform as the prompt optimizer.