This is actually quite hard to achieve, at least in a one-shot LLM prompt.

You could engineer it by asking for a list of subjects sorted by obviousness, then pass that list to a script that discards the top 30% and randomly picks from the rest.

Even if you prompt an LLM to pick one at random, it usually won't, without an external randomness tool like that - which is curious, isn't it, when non-deterministic-ness is so inherent to LLMs.