AI Engineering & Prompt Engineering Interview Questions – Practice Quiz

Prompt engineering is the practice of designing and iterating on the text inputs (prompts) sent to an LLM to steer its responses toward a desired outcome. It does NOT involve altering model weights (that would be fine-tuning) or changing the architecture. The other options describe fine-tuning, architecture design, and model compression — all distinct from prompt engineering.

"Let's think step by step" is a classic zero-shot chain-of-thought (CoT) trigger, popularized by Kojima et al. (2022). Appending this phrase instructs the model to reason through intermediate steps before giving a final answer, which has been empirically shown to significantly improve accuracy on arithmetic, logical, and commonsense reasoning tasks.

"Few-shot" prompting refers to including a small number of worked examples (demonstrations) in the prompt itself so the model can infer the desired pattern. "Zero-shot" prompting provides no examples — the model must generalize from its pre-training alone. These are two of the most fundamental prompting paradigms introduced alongside GPT-3.

System prompt vs user turn is not cosmetic — it determines instruction authority, caching behavior, and scope. Constraints and persona live in the system prompt; per-request context belongs in the user turn.

Prompt iteration follows a scientific loop: measure failures on the eval set first (no root cause without data), diagnose the failure pattern, draft a targeted fix, re-evaluate to quantify the improvement, and promote only when the score delta is positive and no other metric regresses. Skipping the eval step and shipping an intuition-based rewrite is the most common mistake — you end up trading one failure mode for another.

Chain-of-thought (CoT) prompting consistently improves reasoning on multi-step arithmetic and symbolic tasks. However, it does NOT reliably improve — and can sometimes harm — performance on tasks that require only simple factual recall (where extra reasoning steps add noise), highly creative open-ended generation (where structured thinking can constrain diversity), or tasks where the model lacks the background knowledge needed to reason correctly (garbage-in, garbage-out). Retrieval-augmented generation addresses knowledge gaps; CoT alone does not. The key insight is that CoT helps when intermediate steps genuinely decompose a hard problem, not when the bottleneck is elsewhere.

The FILL_BLANK tests precise knowledge of the two canonical advanced prompting strategies. 'Tree of Thoughts' (ToT) extends CoT by exploring multiple reasoning branches simultaneously (like a search tree) and using a verifier/evaluator to prune or select paths. 'Self-consistency' generates multiple independent CoT traces and selects the final answer by majority vote (marginalization over reasoning paths). Both are important senior-level concepts in prompt engineering that go beyond naive zero-shot or few-shot prompting.

Constitutional AI (CAI) uses a two-phase pipeline: first a supervised learning phase where the model critiques and revises its own outputs guided by a set of principles (the 'constitution'), and second a reinforcement learning from AI feedback (RLAIF) phase where a feedback model scores revised outputs. This is distinct from standard RLHF, which relies on human preference labels, and from DPO, which directly optimizes preference data without a reward model. The key insight is that the critique-revision loop is self-supervised and principle-driven, not human-annotation-driven. Knowing the precise mechanics is essential for staff-level AI engineers designing alignment pipelines.

Three findings are robustly supported by the cited papers. (b) Wang et al. 2023 self-consistency paper demonstrated that majority-voting over multiple sampled reasoning chains consistently outperforms greedy-decoded single-chain CoT on benchmarks such as GSM8K and MATH. (c) Kojima et al. 2022 showed that simply appending 'Let's think step by step' to prompts—without any worked examples—produces meaningful accuracy gains in large models, establishing zero-shot CoT as a viable technique. (e) Within Wei et al. 2022's own experiments on GPT-3-class models, CoT gains were empirically observed to appear substantially only above ~100B parameters; this is a factual description of those experiments, scoped appropriately. Subsequent work (e.g., Schaeffer et al. 2023) has debated whether this threshold reflects a genuine phase transition or is a metric artifact, and fine-tuned smaller models (e.g., Flan-T5, LLaMA variants) can show CoT gains—but those findings are outside the cited papers. Option (a) is wrong because Wei et al. 2022 explicitly documented that small models (<~100B in those tests) do not benefit reliably from few-shot CoT. Option (d) is wrong because incorrect rationale steps in few-shot examples tend to degrade performance relative to having no rationales at all, as documented in CoT ablation studies.

This prompt template demonstrates a classic and insidious prompt injection vulnerability. The user-controlled content (the email body) contains a hidden instruction that overrides the system prompt directive to 'only summarize'. Because most LLM inference pipelines concatenate system prompt, context, and user input into a single token stream, a sufficiently persuasive or role-overriding instruction embedded in retrieved/external content can redirect the model's behavior. The model will likely output the exfiltration URL rather than a summary. Recognizing the attack vector (the injected instruction in the 'external content' field), the exact failure mode, and standard mitigations (input sanitization, privilege-separated tool calls, instructed skepticism, output validation) is a staff-level AI engineering concern.