How to Ask

Andrej Karpathy posted a tweet in December 2025 making a simple point: stop treating LLMs as entities with opinions, and start treating them as simulators.

Here’s what he actually wrote:

Don’t think of LLMs as entities but as simulators. … don’t ask “What do you think about xyz”? There is no “you”. Next time try: “What would be a good group of people to explore xyz? What would they say?”

That tweet turned into an actual tool. A Claude Skill called best-minds, barely a few dozen lines of code, states its own principle plainly:

Don’t ask the AI “what do you think.” Ask: who in the world understands this best? What would they say? Then simulate that person.

The rules underneath it: how many people you bring in is decided by the question itself — one is enough when one is enough, more only when the problem calls for friction between viewpoints; the bar is finding whoever genuinely understands it best, not just someone “suitable”; the simulation has to be grounded in what that person has actually published, written, or said, not an invented persona; and wherever possible, it quotes them directly instead of paraphrasing.

The author also compares it to another skill of theirs called lifecoach-ai: that one picks from a fixed roster of 13 preset sages by keyword match, while best-minds has no roster at all — it searches the whole world based on what the question actually is. One picks from a menu. The other goes looking.

What does this look like in practice? When you’re stuck on a headline, instead of stacking up a prompt full of rules — needs a number, needs contrast, needs an emotional hook — just ask who should be the one vetting this headline. The AI pulls up how someone who actually understands headlines thinks, instead of assembling a passing-grade answer inside the rules you handed it. The same applies when you’re stuck on code or can’t think through a product decision: the question becomes who should be looking at this kind of problem.

The problem with long prompts is that they cap what the model can simulate: the more rules you add, the more it’s boxed into producing an “optimal answer” within your own framework. Asking “who should look at this” doesn’t cap anything — it’s the question that actually activates the simulation. One is instruction, the other is simulation — instruction has a ceiling, simulation doesn’t.