What you’ll be able to do

You’ve asked ChatGPT for something. It gave you back three paragraphs of nothing, or a bullet list you didn’t want, or an answer that’s technically correct but useless to you. So you closed the tab thinking it just doesn’t get me. Fair enough. But the fix isn’t a better AI. It’s a better ask, and this guide walks through exactly how to write one, with real prompts you can copy right now.

By the end of this you’ll be able to write a request that gets you something usable on the first or second go, on the free tier of whatever assistant you’ve got open. No paid plan needed, no plugins, nothing to install. Just a different way of typing the question.

Before you start

Here’s what usually happens. You type something like “write me an email about the delay” and hit send. The assistant, having absolutely no idea who you are, what delay, or who’s reading this email, does its best guess. It writes something bland and safe because bland and safe is the only thing that fits every possible version of “an email about the delay.”

That’s not the AI being thick. That’s the AI being sensible given what you gave it, which was basically nothing.

It’s tempting to think the tool is the problem and switch from ChatGPT to Claude, then to Gemini, expecting one of them to just know what you meant. They won’t. None of them can see your screen, your inbox, or the three other conversations you had about this exact delay yesterday. Unless you tell it, it doesn’t exist to them.

So the vague answer isn’t a flaw in the assistant. It’s a mirror. Vague in, vague out. The good news is that once you see it that way, the fix is entirely in your hands, and it’s not complicated.

Get set up

You don’t need anything special for this. Any of the big free assistants will do: ChatGPT, Claude, Gemini, or Copilot. All of them cover everything in this guide on their free tier, so there’s no need to pay for anything before you’ve even tried.

One thing worth checking before you start, though: your privacy settings. By default, the consumer versions of ChatGPT, Claude and Gemini may use your chats to help train their models, unless you’ve opted out. Each one has a settings toggle for this, usually somewhere under data controls or privacy. Have a quick look, decide what you’re comfortable with, and move on. You only need to do this once.

Right, that’s the admin done. Onto the actual thing.

Try it yourself

The prompting guides from OpenAI, Anthropic and Google all agree on more or less the same four ingredients. Clear instructions. Context. An example. And the shape you want back. Get these four in, roughly, and the quality jump is immediate.

Clear instructions just means say the actual job. Not “help me with this email” but “rewrite this email to sound more apologetic and less defensive.”

Context is the bit almost everyone skips. Who are you? What’s this for? Who’s going to read it? An assistant that knows you’re a nurse writing to a patient’s family writes very differently to one that thinks you’re a solicitor writing to a client. Same task, wildly different output, and the only difference is what you told it about the situation.

An example is exactly what it sounds like. If you like a particular tone or layout, paste one and say match this. Don’t describe the tone you want in adjectives, that’s guesswork. Show it.

The shape is length, structure, tone; the practical bits. Do you want three bullet points or two flowing paragraphs? Under 100 words or does it not matter? Formal or the way you’d actually talk to your mate?

Here’s a prompt that bundles all four into one go:

I'm [who you are]. I need [the job] for [who will read it / what it's for]. Format: [length, structure, tone]. Leave out: [what you don't want]. Here is the material: [paste]

Fill in the brackets and paste the actual thing you’re working from, the email you’re replying to, the paragraph you drafted, the job advert you’re applying against. Not a summary of it. The real thing. The assistant can’t see it otherwise, and it will fill the gap with a guess.

There’s a second trick that works on every one of the big assistants, and it takes almost no effort:

Before you answer, ask me the three questions you need answered to do this well. Then wait for my answers.

This flips the whole thing round. Instead of you guessing what context matters and maybe missing something, the AI tells you what it’s missing. Then you answer, and it works from an actual brief instead of a guess. It’s the one to reach for whenever you’re not sure what to include.

And if you’ve got an example of the tone or shape you want, this one does the job:

Here is an example of the tone and shape I want: [paste]. Match it, but for this: [paste]

Three prompts, three different jobs: one that front-loads everything, one that makes the AI ask first, one that copies a style. Between them they cover most situations you’ll actually be in.

Check the result

So how do you know if it worked? Simple test: could you send this with light edits, or did you end up rewriting most of it? If it’s the second one, the ask was missing something, usually context, an example, or a clear shape.

A slightly more revealing test: run the exact same prompt on two different assistants. Paste it into ChatGPT, then paste the identical thing into Claude or Gemini. If they both come back with a similar misunderstanding, that’s not a coincidence, that’s your prompt telling you where it’s weak. If one nails it and the other doesn’t, fine, that’s just normal variation. But if both go wrong in the same way, the problem’s sitting in your ask, not in either tool.

It’s the most useful diagnostic going. It stops you blaming the assistant and pushes you to look at what you actually typed.

If it doesn’t work

Right, so it still didn’t land. A few common ways this goes wrong, and what to do about each.

Too generic. The answer reads like it could apply to anyone, because it basically could. Go back and add who it’s for, and paste an example if you’ve got one. Vague brief, vague answer, every time.

Wrong facts. If it’s stated something confidently that’s just not true, that’s the AI guessing and dressing the guess up as fact. Give it the actual source material instead of asking it to recall something from memory, or tell it directly to say “I don’t know” rather than make something up. There’s a whole guide on this exact problem if it keeps happening: “How to get AI to stop making things up”.

Too flattering, not honest enough. Some assistants have a habit of telling you your plan is great when it’s really not. If you want it to push back properly, see “How to get AI to be honest with me”.

It’s forgotten what you said before. This one catches people out constantly. Unless you’ve turned memory on, or you’re inside a saved project or Gem, the assistant genuinely cannot see your earlier conversations. Each new chat starts from nothing. If you’re repeating the same context every single time and it’s driving you mad, that’s what “How to get AI to remember context” covers, it’s about setting up standing context once so you stop retyping your job title and preferences every conversation.

Too long. Give it a word limit. Assistants default to thorough, which often just means long. “Under 150 words” fixes this instantly.

There’s one more move worth having in your back pocket, and it’s arguably the single best two questions you can ask after any answer:

What did you assume that I didn't tell you? What would make this answer wrong? Then give me a revised version.

This forces the assistant to show its working, basically. It’ll tell you what gap it filled in with a guess, which tells you exactly what to correct next time. Run it after any first draft that feels slightly off; it catches more problems than anything else on this page.

One thing to flag before you paste anything anywhere: never hand an assistant your passwords, card numbers, someone else’s personal details, or confidential work documents. Check your employer’s policy on this before you paste anything work related, some are stricter than others and it’s worth knowing where the line is before you’re mid-conversation.

Questions people ask

Do I need to pay for ChatGPT to get better answers?

No. Any of the big free assistants will do: ChatGPT, Claude, Gemini or Copilot. All of them cover everything here on their free tier, so there's no need to pay for anything before you've even tried.

How do I write a prompt AI actually understands?

Include four ingredients: clear instructions, context, an example, and the shape you want back. Say the actual job, explain who you are and who'll read it, paste an example of the tone you like, and state length, structure and tone.

Why does AI keep giving me generic answers?

Because the answer could apply to anyone, as the brief could too. Vague in, vague out. Go back and add who it's for, and paste an example if you've got one.

How can I tell if my prompt is the problem or the tool?

Run the exact same prompt on two different assistants. If both come back with a similar misunderstanding, the problem is sitting in your ask, not in either tool. If one nails it and the other doesn't, that's just normal variation.

What should I never paste into an AI assistant?

Never hand an assistant your passwords, card numbers, someone else's personal details, or confidential work documents. Check your employer's policy before pasting anything work related, as some are stricter than others.

Keep exploring

If the honesty and made-up-facts issues keep coming up for you specifically, “How to get AI to be honest with me” and “How to get AI to stop making things up” both go deeper into those exact problems than there’s room for here. And if retyping your context every single conversation is wearing thin, “How to get AI to remember context” walks through setting it up once, properly, so it sticks.

Sources and review notes

https://developers.openai.com/api/docs/guides/prompt-engineering https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview https://ai.google.dev/gemini-api/docs/prompting-strategies

Review date: 2026-09-22