Ask ten AI engineering candidates to design a production prompt, and most of them will produce something that works on the happy path. Fewer will produce something that survives an adversarial user, a temperature setting they didn’t choose carefully, or six months of drift.
The interview question is almost always framed the same open-ended way — “walk me through how you’d design a production prompt for a structured extraction task” — and interviewers are quietly checking for five specific things in the answer. Missing one gets you a follow-up question. Missing two or three is usually the difference between advancing and not.
The two structural layers, named correctly
The first thing a strong answer does is separate the system prompt from the user message before saying anything else — role, rules, and output format on one side, the actual task and input on the other. This isn’t pedantry. It signals that the candidate understands why a prompt is structured the way it is, not just that structure exists. Candidates who skip straight to “I’d write a good prompt” without this distinction tend to also skip the reasoning that follows from it.
RCTF, and the failure mode of skipping a field
A strong answer names all four: Role, Context, Task, Format. The interesting part isn’t reciting the acronym — it’s being able to say what breaks when one is missing. Skip Format specifically, and the model doesn’t fail with an error. It silently picks its own output shape, sometimes consistently, sometimes not — which is a worse failure than an error, because nothing tells you it happened until something downstream breaks on it.
Parameters as a separate decision from structure
Candidates who conflate “writing a good prompt” with “choosing good parameters” tend to give shallower answers. Temperature and max_tokens are decisions that sit next to the prompt, not inside it. The strong answer here is specific and situational: temperature 0 for anything that feeds a parser, because that task needs identical output for identical input, not because low temperature is generically “safer.” A candidate who says “I always use temperature 0.7 for consistency” has the mechanism backwards.
Technique escalation, in the right order
Zero-shot first — cheap, low-maintenance, and the right default. Few-shot only once zero-shot’s content is correct but its format or structure isn’t holding up. Chain-of-thought only for tasks with real multi-step or multi-condition reasoning, because it adds tokens and latency that a simple task doesn’t need to pay for. The candidates who reach for the most sophisticated technique first, regardless of whether the task needs it, are usually signaling pattern-matching over judgment.
The beat that actually separates senior from junior
Here’s the one interviewers report catching most often: failure modes, raised without being prompted for them. Prompt injection — user text overwriting developer instructions, because to the model both are just tokens with no privileged boundary between them. Knowledge cutoff — a model answering confidently with stale information because it has no live retrieval and no built-in “I don’t know current data” instinct. Context-gap hallucination — the same fabrication pattern, but for information that was never in training data at all rather than information that’s merely out of date.
A candidate who covers structure, RCTF, and parameters cleanly but never volunteers a failure mode will almost always get asked directly: “what could go wrong with this?” How they answer that follow-up — reactively scrambling versus having already covered it — is often the real signal the interviewer is listening for, more than the first four minutes of the answer.
Why this one question does so much work
It’s a good interview question precisely because it’s hard to fake. A candidate who has only read about prompt engineering can usually name RCTF. A candidate who has actually built and debugged a production prompt volunteers the failure modes unprompted, has an opinion about why temperature 0 matters for a specific task rather than as a rule of thumb, and can explain the cost trade-off between few-shot and fine-tuning without being asked to compare them.
That gap — between reciting a framework and having been burned by skipping part of it — is usually visible within the first ninety seconds.