How To Ask An AI Assistant Better Questions For Better Answers

Most frustration with AI assistants is not actually about the assistant. It is about the question. Two people can ask what sounds like roughly the same thing and get wildly different quality answers back, and the gap almost always comes down to how much the question actually tells the assistant about what is needed, rather than any difference in the tool itself.

Vague Questions Get Vague Answers

Asking "write me an email" gives an assistant almost nothing to work with, so it fills the gaps with generic assumptions: a neutral tone, a medium length, a vague recipient. The result reads fine but rarely fits the actual situation, because the assistant had no way of knowing the situation in the first place. Compare that to "write a short, direct email to a client explaining a two day delay, no apology beyond one sentence, keep it under eighty words," which removes almost every guess the assistant would otherwise have to make.

This is not a special trick, it is closer to how a question gets answered by a colleague who was not in the room for the context either. A person asked to "write me an email" with no other detail would ask follow up questions before starting. An assistant usually will not, unless the question invites it to, so it just picks reasonable defaults and moves ahead. The fix is the same either way: put the context in the first message rather than waiting to be asked for it.

Give the Assistant a Role, Not Just a Task

Framing a request around a role changes the kind of answer that comes back, even when the underlying task is identical. Asking "explain how compound interest works" produces a general, textbook style answer. Asking "explain compound interest the way you would to someone opening their first savings account, no jargon, one simple example" produces something noticeably more usable, because the audience and the tone have both been specified up front instead of left for the assistant to guess at.

The same pattern shows up constantly in getting more useful answers out of an assistant generally: the people who report the best results are rarely using a more advanced tool, they are simply giving the assistant a clearer picture of who the answer is actually for.

Break Big Questions Into Smaller Ones

A single sprawling question that bundles three or four separate asks tends to produce an answer that handles one part well and skims over the rest. Asking an assistant to plan a trip, suggest a packing list, draft an itinerary and recommend restaurants all in one message usually gets a shallow pass at all four rather than a thorough answer to any single one.

Splitting that same request into a short sequence, starting with the itinerary and only asking for the packing list once the itinerary is settled, almost always produces better results at each step, because each question gets the assistant's full attention rather than a quarter of it. This mirrors something trusting an assistant with real work depends on more broadly: breaking a task into checkable pieces rather than handing over one enormous ask and hoping the whole thing comes back right.

Show, Rather Than Just Describe

When the goal is a specific tone or format, an example usually works better than a description of that tone or format. Saying "write in a friendly, casual style" is a vague instruction that gets interpreted differently depending on the assistant and the day. Pasting two sentences of the actual voice that is wanted and asking the assistant to continue in that style removes almost all of the ambiguity, because there is no longer anything to interpret, only something to match.

This matters even more for structured output. Asking for "a table comparing these options" leaves the columns and format up to guesswork. Sketching out even a rough version of the table, with the column headers already filled in, tells the assistant exactly what shape the finished answer needs to take.

Tell It What a Good Answer Does Not Look Like

Constraints are just as useful as instructions, and they are frequently left out entirely. Saying "summarize this article, but do not just repeat the headline in different words" heads off one of the most common weak summaries before it happens. Saying "suggest three options, but nothing that requires buying new equipment" saves a round of back and forth that would otherwise be spent filtering out suggestions that were never realistic to begin with.

Negative constraints work because an assistant without them will often reach for the single most statistically common answer to a question, which is exactly the generic, unhelpful response most people are trying to avoid in the first place. Ruling out the obvious default up front pushes the answer toward something more specific and more useful.

Treat the First Answer as a Draft

The best results rarely come from a single perfect question. They come from a short back and forth, where the first answer gets refined with a follow up like "shorter," "more specific to a beginner," or "remove the third suggestion and replace it with something cheaper." Each of these follow ups is a small, cheap correction, far faster than trying to write one flawless question that anticipates every possible issue in advance.

People who get consistently good results from an assistant tend to treat the first response as a rough draft rather than a final answer, and the small adjustments that follow usually take less time than the original question did to write. That habit, more than any particular phrasing trick, is what separates a frustrating exchange from a genuinely useful one.

The Simple Habit That Changes Everything

None of this requires learning special commands or memorizing a long list of techniques. It mostly comes down to treating the assistant the way a reasonable person would treat a capable but completely new colleague: explain the goal, give the relevant context, say what to avoid, and expect to clarify once or twice along the way. Assistants that seem unreliable are often just being asked questions that leave too much room for guessing, and tightening up the question is usually a faster fix than switching tools.