Same model, wildly different results, depending entirely on how you ask. These six habits are the difference — each one takes a sentence to learn and a lifetime of better answers to enjoy.
Treat it like a smart new hire on day one, not a search box. Give it context before you ask, say specifically what output you want, iterate instead of settling for the first draft, let it ask you questions, hand it a role, and show it an example of "good" whenever you can.
The single biggest lever, and it costs nothing but a few extra sentences. As covered in the previous guide, the AI only knows what's on the desk in front of it. A bare question forces it to guess at everything you didn't say — who you are, what you're trying to achieve, what "good" looks like to you — and it will guess with confidence, which is worse than guessing hesitantly.
"Write an email to my landlord about the heating."
"I rent an apartment and the heating's been out for four days despite two calls to the building manager. Write a firm but polite email to my landlord referencing those two calls, asking for a repair date in writing, and noting this is my third attempt at contact."
The second version doesn't just produce a better email — it produces the right email on the first try, instead of a generic one you'd have spent five minutes correcting.
"Help me plan this" is a request with no shape. Length, format, tone, and audience are all things the AI has to invent if you don't supply them — and it will invent something, confidently, that may not match what you had in mind at all.
"Help me plan a trip to the coast."
"Give me a 3-day coastal trip itinerary for two adults, budget-conscious, no driving more than an hour a day, as a day-by-day bulleted list — nothing prose-y, I want to scan it fast."
Naming the format ("bulleted list, not prose") is a small thing that saves a large amount of re-asking.
A first response is a first draft, not a verdict — and treating it as the final answer is the single most common way people undersell what these tools can do. The AI will happily revise, sharpen, cut, or completely redo something in seconds; the cost of a second pass is nearly zero, so there's rarely a reason to settle for a first pass that's only 80% right.
(Gets a decent but wordy draft of a small-business policy update, copies it as-is.)
"Good structure, but cut it by a third and make the tone less formal — imagine explaining it to a new employee in person, not issuing a memo."
Naming exactly what's wrong — not just "make it better" — gets you a genuinely better second draft, because now there's something concrete on the desk to react to.
When a request is genuinely underspecified — you don't fully know what you want yet either — the fix isn't to guess harder yourself. Ask the AI to interview you first. This flips the dynamic: instead of you trying to anticipate every detail it might need, it tells you what it's missing.
"Help me reorganize my paperwork filing system."
"I want to reorganize my paperwork filing system but I'm not sure what the right structure is. Ask me questions one at a time until you have enough to propose a structure — don't propose anything yet."
This is especially valuable for anything with real stakes: a plan, a policy, a structural decision. A few rounds of good questions up front save several rounds of "no, not quite" after the fact.
Assigning a role isn't a magic incantation — it's a fast, efficient way to load the desk with a whole cluster of relevant context in a single sentence. "Act as a skeptical editor" pulls in a different set of instincts than "act as an encouraging first reader," even for the exact same piece of writing.
"What do you think of this pricing plan for my small business?"
"Act as a skeptical potential customer who's price-sensitive and has three competitors' sites open in other tabs. React to this pricing plan the way you actually would — what would make you close the tab?"
The role doesn't need to be elaborate. It needs to point at the specific lens you actually want applied.
Description is slow and lossy; a real example is fast and precise. If you have a past piece of writing, a report, or a message that hit the tone or structure you're after, hand it over and say "match this" — it's a shortcut past dozens of words of explanation, and it works because the AI is fundamentally a pattern-matcher. Showing it the pattern is more efficient than describing the pattern.
"Write a weekly update for my team, keep it short and upbeat."
"Here's last month's weekly update, which landed really well: [pasted example]. Write this week's update in the same structure and tone, covering: [this week's points]."
One good example, pasted in full, will usually out-perform a paragraph of instructions trying to describe the same thing.
None of these six require special tools or technical setup — they work in any AI chat, today, with the account you already have. Stack a few of them in a single request and the difference is dramatic: context plus a role plus a concrete example will usually get you most of the way to a finished result on the first try, with iteration doing the rest.
The habits here make any single conversation better. The next guide covers what happens when you stop treating each conversation as disposable — and start giving the AI a memory, a method, and a set of rules that carry across all of them.