How to Hire a Prompt Engineer

• aMove everything to the user message each turn
• bCut or consolidate rules, put the most critical ones at the start and end, and verify with an eval whether each surviving rule still changes behaviorCorrect
• cTranslate the rules into bullet-point emojis for salience
• dDuplicate every rule three times so none is missed

Explanation: Attention and instruction adherence decay over long prompts, especially mid-prompt; pruning to high-impact rules at the edges plus eval-verifying each rule is the disciplined fix. Duplication worsens dilution, and moving rules to user messages weakens their priority.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the mid-level experience tier.

• aAsking the model to double-check its own single answer; free and always better
• bRunning the same prompt with a different system prompt once; cheap and reliable
• cTraining a second model to grade the first
• dSampling several independent reasoning paths at temperature above 0 and taking the majority answer; worth it when accuracy gains on hard cases justify multiplying inference costCorrect

Explanation: Self-consistency marginalizes over reasoning paths: errors decorrelate across samples, so majority voting lifts accuracy on verifiable-answer tasks. It multiplies cost and latency by the sample count, so it is reserved for hard, high-value queries, unlike cheap self-checks which mostly re-anchor on the same reasoning.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the mid-level experience tier.

• aA few examples per class including borderline cases (especially mixed), covering real input diversity, ordered consistently and checked for contradictionsCorrect
• bTen examples of the easy positive class
• cPerfect synthetic examples written by the prompt author from imagination
• dThe longest examples available, to teach thoroughness

Explanation: Few-shot sets teach the decision boundary, so class balance, borderline coverage, and real-data diversity matter more than volume; contradictory or idealized examples teach the wrong boundary. Easy-only or single-class sets bias the classifier, and length teaches verbosity, not accuracy.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the mid-level experience tier.

• aChain-of-thought is encrypted and unreadable
• bChain-of-thought is always hidden by all APIs
• cThe written reasoning is a generated rationalization that may not reflect the model's actual causal computation (unfaithfulness), so it cannot serve as proof of why the answer was producedCorrect
• dReasoning traces always contain PII

Explanation: Research on CoT faithfulness shows stated reasoning can diverge from the features actually driving the answer (e.g., models stay silent about using biasing hints). Traces are useful for accuracy and debugging, but they are not verifiable audit evidence of causation.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the mid-level experience tier.

• aMake the prose description 600 words
• bProvide the schema as an explicit JSON Schema or tool definition, show one concrete example, and validate responses programmatically with a retry on violationCorrect
• cAsk the model to promise it will follow the schema
• dSwitch to YAML because it has fewer brackets

Explanation: Formal schemas plus a worked example plus programmatic validation with error-feedback retries give layered guarantees prose cannot. Longer prose dilutes, promises are not enforcement, and changing serialization format swaps one failure surface for another.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the mid-level experience tier.

• aThe rule was not capitalized
• bThe rule is itself just text in the same channel the model reads, so an injected instruction competes with it as another plausible continuation; durable defense is structural: delimit and label retrieved data, gate tool capabilities, and filter outputsCorrect
• cRetrieved documents bypass the context window
• dThe model cannot read system prompts

Explanation: Prompt-based rules against injection are self-referential text that a crafted injection can outweigh; the model has no hardware-level channel separation. Structural controls (data delimiting, least-privilege tools, output filtering) bound what a successful injection can achieve.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the senior experience tier.

• aThe new model is strictly worse; refuse the upgrade
• bVendors intentionally break old prompts
• cYour eval set was too easy
• dThe prompt had overfit to the old model's quirks instead of stating intent clearly; re-baseline by rewriting prompts to plain intent plus examples, then re-validate and keep a prompt-compat eval per model versionCorrect

Explanation: Incantation-style fragments discovered by trial and error encode the old model's idiosyncrasies and break when the model changes. Durable prompts state intent plainly with good examples, and a per-model-version regression suite catches compat breaks at upgrade time.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the senior experience tier.

• aAsk for a short rationale as a dedicated JSON field before the extracted fields (reasoning happens, structure survives), or use the provider's hidden-reasoning mode, keeping the final answer strictly structuredCorrect
• bDrop reasoning entirely and accept the 6-point loss
• cParse the free-text reasoning with regex
• dRun the model twice and concatenate the outputs

Explanation: A schema-respecting rationale field or hidden-reasoning mode keeps the accuracy mechanism (compute before answer) while the final output stays machine-readable and validation-friendly. Accepting the loss is premature, regex over reasoning is brittle, and double-running doubles cost without the mechanism.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the senior experience tier.

• aPosition is irrelevant; the model reads everything equally
• bOnly the very first token matters
• cAttention and instruction salience are strongest at the beginning and end of context and weakest in the middle, so critical constraints and the output contract belong at the edges, with bulky reference material in the middleCorrect
• dThe model reads prompts backwards

Explanation: Empirically, models weight early and late content most (primacy/recency, lost-in-the-middle), so layout is a real variable: rules and output contracts at the edges, reference mass in the middle. Equal-weight and single-token theories contradict observed behavior.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the senior experience tier.

• aAdd the word 'base64' to the keyword list
• bScreen content after decoding and normalization (canonicalize input before matching, run the safety classifier on decoded text), and bound what the model can do regardless of what any input saysCorrect
• cForbid all encoded content, including legitimate attachments
• dSwitch to a model that cannot decode base64

Explanation: Obfuscation defeats surface matching, so screening must happen on canonicalized/decoded content, and capability limits make a successful jailbreak low-value. Blacklisting encodings whack-a-moles endlessly, blanket forbids break legitimate use, and models that cannot decode still follow translated or paraphrased attacks.

Why it matters: Evaluates practical understanding of Prompt Engineering concepts and problem solving for Prompt Engineer at the senior experience tier.