What you need: Any free AI tool (ChatGPT, Claude, or Gemini) · one task you’ve been putting off · ~15 min

You’ve done this. You typed a real question into ChatGPT, hit enter, and got back something so generic it could’ve been written for anyone on earth. Helpful in the way a fortune cookie is helpful.

That’s not the AI failing you — it’s the ask. Getting something genuinely useful out of these tools comes down to a skill, a small and specific one, and it’s most of what I do with AI all day. I want to hand it to you here, in three parts.

I’ll use a recent example to make it concrete: a bike. My kid had outgrown hers — knees hitting the handlebars — and I was deep in Facebook Marketplace at 9:47pm, a dozen listings open, every photo shot in a dim garage at the one angle that hides exactly what you need to see. Messy, real, kind of a pain — which is precisely the sort of task AI handles well, if you ask right. The bike is just the example. The three parts are the point.

The problem is the ask, not the AI

Most of us prompt like we’re typing into a search bar:

which of these bikes should I buy for my 8 year old kid

And the answer comes back true and useless — “it depends on your child’s height, budget, and how they’ll use it!” — because that’s genuinely all the AI had to go on. It’s not a mind reader, and it’s too polite to push you for more. Vague question, vague answer. Every time.

The fix isn’t a fancier tool or a secret phrase. It’s giving the AI three things it almost never gets enough of.

Every good prompt has three parts

Here’s the whole skill, and it transfers to basically everything you’d ever hand to AI:

A good prompt has context, instructions, and an output.

Picture briefing a junior employee on a task. They need to know your situation. They need to know what you actually want them to do with it. And they need to know what you want handed back at the end. Skip any one of those and you get a shrug. Nail all three and you get something you can actually use.

It’s the same shape every time — only the details inside each part change. Let’s take them one at a time.

Part 1 — Context: tell it your real situation

A lot of AI tools now remember bits about you from past chats, so they’re not totally blank. But “bits” is rarely enough for the decision in front of you — and the AI knows nothing about this specific situation until you spell it out. Your instinct is to keep it short — “a bike for my kid” — and that short version is exactly why the answer feels generic. Give it the real situation, the same details you’d give your junior employee before they started. And if there’s something you could show it — a screenshot, a photo, a flyer — hand that over too. Most AI tools can read images now, and a picture saves you a paragraph.

For the bike, “real situation” looked like this:

Help me pick a used bike for my kid. My daughter is 8 and 4'2". I want a lightweight bike for pavement and gravel paths, a decent brand that'll hold up, and a little room to grow — but not so big she can't put her feet down. Budget's around $150. I've attached screenshots of 5 bikes from Facebook Marketplace that are in my budget and roughly her size.

Part 2 — Instructions: tell it what to actually do

Context sets the scene; instructions are the job. This is where most prompts go thin — we assume the AI will infer what we want done. It won’t, reliably. The move is to picture the steps you’d take yourself if you had a free afternoon, then spell them out. Every little question you’d otherwise Google one at a time goes here. The more of the work you name, the less you’re left doing yourself afterward.

For the bike, the job was: compare them, check fit, sanity-check the prices, and dig the buried details out of each listing — including the “needs new tubes” line that makes you nervous.

Go through the 5 listings and compare them:
- How well each fits an 8-year-old who's 4'2" — wheel size and standover height matter here
- How each does on my priorities: lightweight, good on pavement + gravel, room to grow, solid brand
- Roughly what each costs new, so I can tell if the asking price is a deal
- Anything the seller mentioned that affects the price — "needs new tubes," scratches, missing parts — and whether it's an easy fix or a dealbreaker

Part 3 — Output: tell it what to hand back

Last, name the shape you want the answer in. Ask for “a summary” and you get a wall of text. Ask for the exact thing you can actually use — a table, a checklist, a ranked list, a ready-to-send email — and you get something you can act on without reformatting it in your head.

For the bike, the useful shape was a ranked table I could scan one-handed:

Give me one table ranking the bikes from best fit to worst for my situation. Columns: bike, wheel size, weight, asking price vs. new price, condition notes, and a one-line reason for the ranking. Put the one you'd buy at the top and tell me why.

And the output doesn’t have to stay trapped in the chat. Most tools will hand you a printable document — a .docx or PDF you can save or pin to the fridge — if you just ask. Overkill for a five-bike shortlist, but for something bigger it earns its keep: it’s how the birthday-party plan comes out as a printable checklist you can actually carry around on the day.

Context, instructions, output. That’s the whole move. With all three in place, the answer comes back as a ranked shortlist with a recommendation instead of “it depends!” — because the ask is carrying its weight instead of leaving everything to the AI.

“That’s a lot of typing”

It is — up front. If those three parts look like more work than “which bike should I buy,” that’s because they are, for about one extra minute. But that minute is what saves you the twenty on the back end, cross-checking specs and prices one browser tab at a time.

And you don’t have to type any of it. Every one of these tools takes voice now, so you can just talk it through — “okay, she’s eight, about four-two, I want something light for pavement and gravel…” Out loud is honestly where this structure feels most natural anyway and the AI won’t mind if you ramble a bit.

When you don’t even know what to ask

All of this assumes you already know what you want to ask for. Some nights you don’t. At 9:47pm, fourteen tabs deep, “a bike, I think?” is a completely valid mental state.

When you’re not sure what to ask for, hand that problem to the AI too — tell it to interview you first:

I'm trying to pick a used bike for my 8-year-old, but I'm not sure what I should be comparing. Ask me a few questions one at a time, then use my answers to tell me what matters and recommend the best fit.

It’ll walk you through the decisions you didn’t know you had to make — wheel size, where she’ll actually ride, how much growing room is too much — and you’ll back into a solid prompt without having written one.

A few honest things about asking well

Specific in, specific out — it’s a law, not a tip. “My daughter likes biking” gets you generic. “8, 4’2″, pavement and gravel, room to grow, $150” gets you a real answer. The detail you put in is the usefulness you get out.

Whatever it picks, verify before you buy. AI is reading the same dim garage photo you are, and it’ll confidently fill in a spec or a price it isn’t actually sure about. So once you’ve landed on a top pick, do the two-minute check yourself: Google the exact model, confirm the wheel size, weight, and real price against the listing — and if you can, have your kid straddle it before any money changes hands. The structure of the answer is reliable; the final fact-check is yours.

It’s a conversation, not a vending machine. If the answer’s missing something, just say so — “add a column for hand brakes vs. coaster brakes.” Ten seconds, redone. The back-and-forth is where the good stuff lives.

Learn these three parts once and you stop getting fortune-cookie answers for good. The bike’s the easy part. The asking is what pays off every time after.