Task 3.2: Choose effective prompt engineering techniques.

• AZero-shot prompting

  Incorrect

  A zero-shot prompt gives the task without any example input-output pairs. This prompt includes three.

• BChain-of-thought prompting

  Incorrect

  Chain-of-thought prompting guides the model to reason step by step before answering. Nothing here asks for intermediate reasoning.

• COne-shot prompting

  Incorrect

  A shot is one paired example of an input and the desired output. This prompt contains three examples, not one.

• DFew-shot prompting

  Correct

  Providing a few paired examples of input and desired output in the prompt is few-shot prompting, also called in-context learning.

Count the examples: none is zero-shot, one is one-shot, a few is few-shot. Examples steer the format and labels without changing the model.

• ARaise the temperature so the model has more freedom to find a format

  Incorrect

  A higher temperature makes output more random, which would make the format even less consistent.

• BState the task, the audience and the relevant context in clear, specific words

  Correct

  Models work best with clear, well-structured, specific instructions, and providing contextual information helps align the output with the target scenario.

• CSpecify the output format and length, such as three bullet points

  Correct

  Output indicators and explicit format constraints tell the model what to produce; without them the model chooses its own length and format.

• DRemove all background information so the model is not distracted

  Incorrect

  Context helps the model; the guidance is to start by providing contextual information in the query.

• EPack several unrelated requests into the same sentence to save tokens

  Incorrect

  Models on Amazon Bedrock work best with simple, clear and complete instructions; mixing unrelated requests makes the prompt harder to interpret.

Specific instructions with context, plus an explicit output format, are the first fixes for vague and inconsistent responses.

• APrompt injection

  Correct

  Prompt injection is a user prompt designed to ignore and override the instructions specified by the developer, which is exactly this request.

• BJailbreak

  Incorrect

  A jailbreak tries to bypass the model's native safety and moderation capabilities to generate harmful content. This user wants the topic restriction removed, not harmful output.

• CPrompt leakage attack

  Incorrect

  Prompt leakage tries to extract or reveal the system prompt or developer instructions. The user is not asking to see them.

• DHallucination

  Incorrect

  A hallucination is incorrect or misleading output from the model, not an attack by a user.

Override the instructions: injection. Reveal the instructions: leakage. Bypass safety to get harmful content: jailbreak. Bedrock Guardrails can detect all three as prompt attacks.

• AAmazon SageMaker Model Registry

  Incorrect

  Model Registry catalogs and versions trained models and their approval status. It does not manage prompts.

• BAmazon Bedrock Prompt Management

  Correct

  Prompt Management saves reusable prompts with variables, a chosen model and inference parameters, lets you compare variants, and creates versions, snapshots you deploy while you keep iterating on the draft.

• CAWS Secrets Manager

  Incorrect

  Secrets Manager stores and rotates secrets such as database credentials and API keys, not prompt templates.

• DAmazon Bedrock Knowledge Bases

  Incorrect

  Knowledge Bases retrieve information from your data for Retrieval Augmented Generation. They do not version prompt templates.

Treat prompts like code: template them, test variants, and promote versions. Amazon Bedrock Prompt Management is built for that.

• AFew-shot prompting

  Incorrect

  Few-shot prompting includes example input-output pairs. This prompt has none.

• BChain-of-thought prompting

  Incorrect

  Chain-of-thought prompting asks the model to reason step by step before it answers.

• CPrompt caching

  Incorrect

  Prompt caching reuses repeated prompt prefixes to cut latency and cost; it is not a prompting technique.

• DZero-shot prompting

  Correct

  A zero-shot prompt states the task without any example input-output pairs, as in this sentiment classification prompt.

Zero-shot relies on the model's general ability; add examples (few-shot) when the output format or labels need calibrating.

• AChain-of-thought prompting

  Correct

  Guiding the model to think step by step and reason before reaching an answer is chain-of-thought prompting; it improves accuracy on problems such as math word problems.

• BZero-shot classification

  Incorrect

  Zero-shot means no examples are given; the change here is asking for intermediate reasoning steps.

• CPrompt injection

  Incorrect

  Prompt injection is an attack that tries to override the developer's instructions, not a technique the developer uses.

• DRetrieval Augmented Generation

  Incorrect

  RAG adds retrieved information to the prompt; nothing is retrieved here.

Asking for the reasoning before the answer gives the model room to work through intermediate steps, which helps on multi-step problems.

• APrompt leakage

  Incorrect

  Prompt leakage tries to extract the system prompt or developer instructions. The user wants harmful content instead.

• BJailbreak

  Correct

  A jailbreak is a prompt designed to bypass the model's native safety and moderation capabilities to generate harmful or undesirable content.

• CFew-shot prompting

  Incorrect

  Few-shot prompting is a legitimate technique that adds example input-output pairs to a prompt; it is not an attack.

• DHallucination

  Incorrect

  A hallucination is the model's own incorrect output, not a user's attempt to manipulate it.

Attacks aimed at the model's safety protections are jailbreaks. A Guardrails prompt attack filter can detect them, along with prompt injection and leakage.

• AValidate and sanitize all user input before it reaches the model

  Correct

  Input validation, including removing or escaping special characters and enforcing expected formats, is a listed best practice against prompt injection.

• BConcatenate raw user input directly into the instruction text

  Incorrect

  AWS recommends avoiding string concatenation for input as a secure coding practice.

• CGive the application broad permissions so that blocked requests can be retried

  Incorrect

  The guidance is the opposite: apply the principle of least privilege when granting access to resources.

• DRaise the temperature so that injected text is less likely to be followed

  Incorrect

  Temperature controls randomness in the output. It is not a security control.

• EUse an Amazon Bedrock guardrail that detects prompt attacks

  Correct

  A guardrail can help protect against prompt injection; its prompt attack filter detects attempts to override instructions.

Prompt injection is defended in layers: validate inputs, keep least privilege, test regularly, and add a guardrail that detects attacks.

• AA prompt template

  Correct

  A prompt template specifies the formatting of the prompt with exchangeable content in it, such as values filled in for each request.

• BA few-shot prompt

  Incorrect

  A few-shot prompt contains example input-output pairs; placeholders for data are what make a template.

• CA guardrail policy

  Incorrect

  A guardrail filters undesirable content in inputs and outputs. It is not a reusable prompt format.

• DAn inference profile

  Incorrect

  An inference profile defines a model and the Regions requests can be routed to; it does not hold prompt text.

Templates keep prompts consistent and testable; Amazon Bedrock Prompt Management can store them with variables and versions.