AI Foundations: Prompt Engineering Practice Questions

Prompt B explicitly instructs the model to reason step by step (list facts → identify constraints → derive answer), which is the core of chain-of-thought prompting. This also represents decomposing the task into structured sub-steps, a related best practice that breaks a complex problem into manageable stages. Role prompting assigns a persona and is not present here. Self-consistency requires multiple runs and majority voting, which is not described. Few-shot prompting requires examples, which are absent. Temperature reduction is a parameter change, not a prompting technique, and is not mentioned.

The instruction 'You are a helpful assistant' defines the model's persona. The instructions to never discuss competitors and to respond in formal English are behavioral constraints that restrict the model's actions. Output format constraints refer to things like 'respond in JSON' or specific structural rules—'150 words' is a length constraint that can overlap, but the clearest two distinct categories here are persona and behavioral constraints. Chain-of-thought activation is not present (there is no 'think step by step' instruction). Few-shot examples are not included anywhere in this prompt.

A low temperature (0.1) minimizes randomness and keeps the model focused on high-probability tokens, producing conservative and consistent outputs. A moderate top-p (0.85) still allows some nucleus sampling flexibility while excluding very low-probability tokens. Temperature 1.5 / top-p 0.99 maximizes diversity and creativity—the opposite of what is needed for a precise financial assistant. Temperature 0.8 / top-p 0.5 restricts the token pool aggressively with top-p while also introducing moderate randomness—an inconsistent and suboptimal combination. Temperature 1.0 / top-p 1.0 is the default setting (no restriction), which allows the full range of the probability distribution and is inappropriate for conservative financial use.

Few-shot prompting provides the model with a small number of input-output examples to demonstrate the desired task format before presenting the actual query. This is exactly what the developer is doing with two translation examples. Zero-shot would provide no examples. Chain-of-thought asks the model to show its reasoning steps. Iterative prompt refinement is a process of evaluating and improving prompts over time, not a structural technique.

Research, including Google's original CoT paper, demonstrated that appending 'Let's think step by step.' reliably elicits step-by-step reasoning from large language models. Requesting JSON format controls output structure, not reasoning process. A low top-p restricts token sampling diversity but does not induce reasoning. Providing final answers without reasoning (few-shot without CoT examples) does not teach the model to show its work.