I run compliance training for a financial-services team, which means once a quarter I have to test twelve adults on a new regulation and pretend the test means something. Before I found this prompt, I wrote the quizzes myself, and they were a small humiliation. Not factually wrong — just obvious. Recall questions with one right answer and three answers so clearly wrong they may as well have been written in a different language. My colleagues finished in four minutes flat and high-fived each other on the way out. Nobody learned anything. We all just confirmed we could read.
I had no idea what a good distractor was
Here's the skill I didn't have: writing wrong answers that a half-prepared person would genuinely pick. That's the entire game in multiple-choice. If your wrong answers are obviously wrong, you're not testing knowledge, you're testing whether someone is awake. Good distractors are plausible. They're the things you'd believe if you skimmed the material instead of reading it. Writing them well is a real craft and I had spent exactly zero hours learning it.
So I handed the job to a model that's quick on its feet. I pasted our new anti-money-laundering summary into Grok 4.1 Fast Reasoning, set the level to 'applied reasoning,' and ran the Generate a 10-Question Quiz with a Self-Grading Answer Key prompt. Four minutes later I had a quiz that was meaner than anything I'd ever written, in the best way.
The question that broke the curve
One question read, roughly: 'A client makes ten cash deposits of $9,500 over three weeks. What is the most appropriate first action?' The four options were all things a reasonable person who hadn't carefully read the structuring rules might pick. File a report immediately. Ask the client about the deposits. Do nothing because each is under the threshold. Escalate internally first. They're all defensible if you're fuzzy on the material, and exactly one is correct per the reg. My team averaged sixteen minutes on this quiz. Two people failed it. In four years of running this training, nobody had ever failed anything.
The 'do nothing because each is under the threshold' option is the perfect distractor, and it's the kind I'd never have written, because to write it you have to understand the exact misconception you're testing for: that staying under $10,000 makes you safe, when deliberately structuring deposits to stay under the reporting threshold is itself the thing the regulation is about. A bad quiz writer makes the wrong answers obviously wrong. A good one makes the most tempting wrong answer the exact mistake an undertrained employee would actually make on the job. Grok built that, from the source text, in the time it takes me to make coffee.
- Applied-reasoning questions test whether someone can use a rule, not just recite it
- Plausible distractors separate 'almost understands' from 'actually understands' — which is the whole point of compliance
- The balanced answer spread means nobody can game it by always picking C
- The answer key's one-line 'why the tempting wrong answer is wrong' is what I read aloud in the debrief
- The difficulty note told me which two questions to walk through with the team, and it was right both times
Why the two failures were the best outcome
I want to be clear that the two people who failed are good at their jobs. They failed a hard quiz on a brand-new regulation, which is what's supposed to happen — that's the quiz doing its job. They learned more from getting that structuring question wrong and then arguing with me about it in the debrief than they'd have learned from a flawless score on one of my old toy quizzes. Struggle is where the learning lives. A quiz everyone aces taught me nothing about who needs help; a quiz two people fail tells me exactly where to spend my next session.
It also changed the room. When the quizzes were easy, training was theater — everyone knew it didn't matter, so nobody paid attention, so the actual compliance risk sat there untouched while we performed competence at each other. A quiz that can actually be failed makes people read the material beforehand, which is the entire point of having training at all. I'd accidentally been running an attendance check and calling it education. Now I'm running education, and the difference shows up in the questions people ask me in the hallway weeks later.
The honest caveat: because everything must come from the source text, the quiz is only as good as what I paste in. If my regulation summary is thin, the quiz is thin, and once it wrote a question whose 'most tempting wrong answer' was actually arguably correct — I caught it because I read the key, and now I always read the key before I print. For a compliance quiz where a wrong 'right answer' could genuinely mislead someone about a rule, that proofread is non-negotiable, and I'll occasionally rewrite one option by hand. When I'm building a whole module I pair this with my Turn Messy Lecture Notes into Spaced-Repetition Flashcards prompt so people can drill the key terms before the test. Grab the Generate a 10-Question Quiz with a Self-Grading Answer Key prompt on Prompt Dock, paste your driest policy document, and watch your team finally take it seriously.