AI Prompt Engineering — Proficiency Exam

15 questions · pass with 50%

1. In this course, a prompt is best described as:
A secret keyword that trains a new model on your laptopThe brief you give the model — task, context, and constraints — so it can draft a replyA guarantee that every sentence is a verified factA replacement for your company's document store
2. You asked for a short email, then later wanted a legal memo, in the same chat. The model keeps writing emails. A reliable fix is to:
Type only the word 'better' until it switchesDelete Excel from your computerNever use constraints in promptsState the new goal explicitly or start a new chat so old instructions stop dominating
3. "Make this better" is a weak prompt mainly because:
It does not define audience, length, tone, or what better means, so the model must guessModels cannot rewrite text at allIt uses too many constraintsIt includes too much source material
4. Which instruction best reduces invented numbers in a summary of a pasted policy?
You are a world-class geniusBe creative with figuresIf the excerpt does not contain the fee, say you cannot find it — do not guessIgnore the excerpt and use typical industry fees
5. Few-shot examples in a prompt are:
Short samples of the input-to-output shape you wantA method for training a new model from scratchRequired to be at least ten pagesA replacement for providing any context
6. You need a customer email and a separate internal checklist. The clearest approach is to:
Hide both requests in a jokeAsk for two labelled outputs or use two prompts so you can judge each jobAsk only for 'make it better' oncePaste unrelated policies for extra context
7. You change tone, format, sources, and length all at once and the next draft is still wrong. The process problem is:
You cannot tell which change helped or hurt, so you should alter one lever per roundModels forbid constraintsYou must never save a promptIteration is only for programmers
8. A draft has the right outline but the wrong audience. An efficient follow-up is:
Start every new chat with a random poemDelete the outline and ask for 'something totally new'Keep the same outline; rewrite for this audience and give a two-line sample of the toneAsk the model to train itself on your inbox
9. You should usually stop iterating when:
The model uses a long wordThe first token appearsYou have generated exactly two replies, alwaysRemaining fixes are faster by hand, or new versions are not adding information
10. You ask a model for academic citations on a topic and paste no sources. The citations it lists are:
Always real and correctly quotedOften plausible-looking and still wrong — you must verify they exist and support the claimImpossible because models cannot output URLsAutomatically filed with your library
11. Which prompt pattern most directly discourages guessing a fee that is not in a pasted policy?
Use only this text; if the fee is not present, say so and do not invent oneBe maximally helpful and completeAdd extra unrelated examples of fees from other industriesAsk for a confident number even when unsure
12. The model summarises your policy and disagrees with a sentence you can see in the excerpt. You should:
Average the two versions into a new ruleTrust the model because it sounds certainFollow the source (or a human expert), and fix the prompt if you need a faithful summaryDelete the policy
13. You want help rewriting a customer email that contains an account number and a medical detail. You should:
Paste the whole thread into a public chatbot because rewriting is low riskPost the email in a public forum for extra contextUse the account number as the prompt title for trackingRemove sensitive details or use an approved internal tool — do not paste secrets or PII into a public model
14. Which job is the best fit for a language model as used in this course?
Automatically firing employees based on a generated scoreDrafting three subject-line options for an email whose facts you already approvedFiling taxes with numbers the model inventedDiagnosing a patient and sending the result without a clinician
15. A coworker says a public chatbot confirmed your unpublished revenue figure. You should treat that as:
Not a source — unpublished figures should not have been pasted, and the model is not a system of recordAudit-ready evidenceA reason to paste the rest of the finance folderProof the number is public now

Answer all 15 questions to submit.