Prompt Engineering · CyberQuiz

Q1 Easy 10 pts

What is 'few-shot' prompting?

Limiting the model to a few tokens Giving the model a few worked examples of the task in the prompt Asking only a few questions Using a small model

Q2 Easy 10 pts

'Zero-shot' prompting means:

Giving no task description Asking the model to do a task with no examples, just instructions Turning off the model Using zero tokens

Q3 Medium 15 pts

Adding 'Let's think step by step' elicits which reasoning technique (three words / acronym)?

Q4 Easy 10 pts

The role of a 'system prompt' is to:

Store the user's password Set the model's persona, rules and context before the conversation Compile the model Pick the GPU

Q5 Medium 15 pts

To reliably get JSON output you should:

Ask nicely once Specify the exact schema, give an example, and use structured-output/JSON mode Raise temperature to 2.0 Use ALL CAPS

Q6 Medium 15 pts

Raising the 'temperature' parameter:

Makes output more random/creative Makes output more deterministic Speeds up the GPU Shrinks the context window

Q7 Hard 20 pts

Prompting a model to critique and improve its own answer is broadly called self-____ .

Q8 Medium 15 pts

Setting temperature to 0 tends to make output:

Random Nearly deterministic / repeatable Longer Encrypted

Q9 Medium 20 pts

'top-p' (nucleus) sampling controls:

The max tokens The cumulative probability mass of tokens the model samples from The GPU count The system prompt

Q10 Medium 15 pts

A 'stop sequence' is used to:

Start generation Tell the model where to end its output Encrypt the prompt Pick a persona

Q11 Easy 10 pts

'One-shot' prompting provides:

No examples Exactly one example of the task Ten examples Only the answer

Q12 Easy 10 pts

'Role prompting' means:

Assigning the model a persona/expertise (e.g. 'You are a security expert') Rolling dice for tokens Rotating API keys Randomizing output

Q13 Medium 15 pts

Using delimiters (```triple backticks``` or XML tags) helps by:

Encrypting the prompt Clearly separating instructions from data/content Reducing accuracy Disabling safety

Q14 Hard 20 pts

'Self-consistency' improves reasoning by:

Using one greedy answer Sampling multiple reasoning paths and taking the majority answer Lowering max tokens Disabling CoT

Q15 Hard 20 pts

'ReAct' prompting interleaves:

Two models Reasoning steps with actions/tool calls Images and audio Two languages

Q16 Medium 15 pts

Breaking a complex task into smaller sub-prompts is called:

Compression Task decomposition / prompt chaining Tokenizing Caching

Q17 Medium 20 pts

Which most reduces hallucination?

Higher temperature Grounding the model on provided sources and asking it to cite them Longer persona ALL CAPS

Q18 Medium 15 pts

RAG stands for:

Random Answer Generation Retrieval-Augmented Generation Rapid API Gateway Recursive Agent Graph

Q19 Easy 10 pts

The maximum amount of text (tokens) a model can consider at once is its context ____ .

Q20 Medium 15 pts

Providing an example of the desired OUTPUT FORMAT primarily improves:

Model speed Consistency/structure of the response GPU temperature Token price

Q21 Medium 15 pts

'In-context learning' refers to the model:

Being retrained on your data Learning the task from examples in the prompt at inference time Updating its weights live Caching answers

Q22 Easy 10 pts

Best way to enforce a specific persona/tone across a chat?

Repeat it every user message Set it once in the system prompt Use a bigger model Raise top-p

Q23 Medium 15 pts

Negative instructions ('do NOT include X') are:

Always perfectly followed Helpful but less reliable than stating what TO do Ignored entirely Only for images

Q24 Medium 15 pts

'max_tokens' limits:

Input length only The maximum number of tokens the model will generate in the response The temperature The persona

Q25 Medium 20 pts

Why include the reasoning AND the final answer separately (e.g. tags)?

To waste tokens So you can parse a clean final answer while still benefiting from reasoning To confuse the model To encrypt output

Q26 Medium 15 pts

'Prompt template' with variables is useful because it:

Prevents any output Standardizes prompts and lets you inject dynamic values safely Slows the model Removes the system prompt

Q27 Medium 15 pts

'Meta-prompting' means:

Prompting about the weather Using the model to help write or improve prompts Doubling the temperature Disabling tools

Q28 Medium 15 pts

For a classification task, a good prompt:

Leaves labels undefined Enumerates the allowed labels and asks for exactly one Requests an essay Sets temperature high

Q29 Medium 20 pts

'Frequency penalty' is raised to:

Increase repetition Discourage the model from repeating the same tokens Speed up output Change the persona

Q30 Easy 10 pts

The three standard chat roles are:

admin, guest, root system, user, assistant read, write, execute get, post, put

Q31 Hard 20 pts

'Tree of Thoughts' extends chain-of-thought by:

Using one path Exploring and evaluating multiple branching reasoning paths Removing reasoning Lowering max tokens

Q32 Medium 20 pts

To make outputs reproducible across runs you can:

Only change the prompt Fix temperature low/0 and set a seed where supported Increase top-p Add more examples

Q33 Medium 15 pts

A prompt that says 'If unsure, say you don't know' helps to:

Increase hallucination Reduce confidently-wrong answers Slow the model Change the role

Q34 Medium 15 pts

Which improves extraction of fields from messy text?

Vague instructions A clear schema listing each field and its type, with an example High temperature No examples

Q35 Hard 20 pts

Placing key instructions at the START and END of a long prompt helps because:

Models ignore the middle sometimes ('lost in the middle') It looks nicer It saves tokens It encrypts them

Q36 Medium 15 pts

'Guardrail' prompts add:

More randomness Explicit boundaries on what the model may/may not do A bigger context A new tokenizer

Q37 Medium 15 pts

Why version and test (A/B) your prompts?

To waste time Prompts drift/regress; testing tracks quality and catches regressions To use more GPUs To hide them

Q38 Easy 10 pts

For summarization, specifying length/format (e.g. '3 bullet points') mainly improves:

Latency Control and consistency of the output Temperature The persona

Q39 Medium 15 pts

Feeding the OUTPUT of one prompt as the INPUT to the next is called prompt ____ .

Q40 Medium 15 pts

Asking the model to 'show your work, then give a final answer on the last line' is a form of:

Encryption Structured chain-of-thought with a parseable answer Fine-tuning Rate limiting

Q41 Medium 15 pts

Embeddings are used in prompt engineering mainly to:

Encrypt prompts Find semantically similar documents for retrieval (RAG) Increase temperature Set the persona

Q42 Easy 10 pts

A concise, specific instruction usually beats a vague one because:

Models prefer long text Ambiguity leads to inconsistent or off-target outputs It costs more It disables safety

Q43 Medium 20 pts

'Presence penalty' encourages the model to:

Repeat topics Introduce new topics/tokens not yet used Stop immediately Lower temperature

Q44 Medium 15 pts

When a prompt exceeds the context window, a common fix is to:

Ignore it Summarize/chunk the input or use retrieval to include only relevant parts Raise temperature Add a persona

Q45 Medium 15 pts

Providing counter-examples (what NOT to output) alongside good examples:

Always hurts Can sharpen the boundary of the desired behavior Is impossible Disables the model

Q46 Medium 20 pts

A robust production prompt should also:

Trust all user input as instructions Separate untrusted input from instructions to resist prompt injection Skip validation Expose secrets

Q47 Easy 10 pts

Which yields more consistent tone across many calls?

Rewriting instructions each time A reusable system prompt + template Random personas Higher temperature

Q48 Hard 20 pts

'Least-to-most' prompting solves hard problems by:

Answering instantly First solving simpler sub-problems, then composing them Skipping steps Lowering max tokens

Q49 Medium 15 pts

The best single habit for reliable prompts is:

Guessing Iterating with evaluation on real examples Maxing temperature Avoiding examples

Q50 Medium 15 pts

Giving the model the current date/context when needed prevents:

Faster output Stale or wrong time-sensitive answers Lower cost Shorter prompts