Red Flags That Sound Junior

Interviewers do not mark you down for not knowing every technique. They notice how you talk about the ones you use. Certain phrases signal that a speaker is repeating a slogan instead of reasoning, and they cost points even when the idea behind them is fine. This chapter collects the most common ones and a stronger replacement for each.

A typo in a résumé does not prove someone is careless, but readers treat it as a signal. In the same way, a phrase like "it should be fine" does not prove a candidate is shallow, but listeners treat it as a sign that nothing was checked. The fix is cheap: say what you would measure.

Each red flag skips one of three things: a reason (why this and not another), a number (how big, how slow, how costly), or a check (how you would know). A strong sentence has at least two. When you hear yourself saying a slogan, add the reason, the number, or the check.

• -"We'll just use a bigger model." Say what the bigger model fixes, what it costs in latency and money, and how you would test that it helps on an eval set before paying for it.
• -"We'll measure accuracy." With rare positives, accuracy hides failure. Say precision and recall at the review capacity, or a cost-weighted metric, and name the threshold.
• -"It should be fine." Say how you would find out: an eval, a shadow run, a canary with guardrail metrics.
• -"We need more data." Say which data is missing, what kind of label, roughly how much, and how you would check that it helps.
• -"The model is state of the art, so we do not need to evaluate it." Public benchmarks are not your task. Say you would build a small eval set from real cases.
• -"Use a vector database." Say why retrieval is needed here, whether exact terms matter (hybrid search), and the size of the index.
• -"The LLM will handle the edge cases." Say which edge cases matter, how you would test them, and what rule or fallback covers the ones that fail.
• -"Cache everything." Say what is stable enough to cache, how it expires, and what a stale answer would cost.
• -"Fine-tune it." Say that you would try prompting and retrieval first, and fine-tune when an eval shows a consistent behavior gap.
• -"An agent can do that." Say whether the path is predictable. If it is, a workflow is simpler and more reliable, and if not, say which limits and approvals you would add.
• -"Guardrails will stop it." Say there are layers, that no layer is perfect, and that you would measure attack success and over-refusal together.
• -"We'll add more GPUs." Say which bottleneck you found: prefill, decode, memory, or batching, and what you would try before buying hardware.

The code below is a toy: a list of patterns for several of these flags, and a function that scans a paragraph and prints what it finds. It will not judge a real answer, but it is a good way to practice, since you can paste your own practice answers into it. Press Run, and read the three examples.

The first answer, "just use a bigger model, it should be fine and accuracy will go up", trips four flags: a slogan with no trade-off, a guess with no check, a bigger model with no cost, and accuracy with no mention of rare cases. The second answer starts with a baseline and names precision and recall at the review capacity, and the scanner finds nothing. The third trips three: an absolute, a vague "more data", and another "just". Note the limit of this tool. A pattern match cannot tell whether the second answer is also good. It only finds slogans.