How to talk to AI: what to type when the box is empty

You are not typing into a search box

The cursor blinks. The placeholder text says something like "Message ChatGPT" and you sit there, because nothing you can think of feels like a proper question. This is the most common first experience people have with AI, and it says nothing about you. The box looks exactly like a search bar and behaves like nothing you have used before.

Almost everyone types keywords the first time: three or four words, no verb, no context. That is a reflex trained by twenty years of search engines, and it is the one habit worth dropping on day one. A search box wants keywords because it is matching pages. A chat model wants sentences because it is working out what job you are handing it. Take a real need, cancelling a gym membership without an argument, written both ways:

• As a search query: cancel gym membership letter template
• As a request: I want to cancel my gym membership. Write a short, polite message I can send by email. I have been a member two years, my contract needs 30 days notice, and I do not want to give a reason.

The first version returns a generic template with brackets to fill in. The second returns a message you can send, with the notice period already in it and the reason deliberately left out. Same need, about fifteen seconds more typing, a completely different level of usefulness. Try it now on something small from your own week, an email you owe someone or a form you have been putting off. It will also, at some point, tell you something untrue in a confident voice, which the last section deals with. If all of this is new, what AI actually is covers the ground underneath it, and AI versus Google search explains which questions still belong in a search engine.

Four things worth putting in a first message

There is no formula to memorize. What matters is including the things only you know, because the model cannot see your inbox, your deadline, your job, or the person you are writing to. Think of it as briefing a capable temp on their first morning. They can do the work, but nobody has told them who the work is for.

Four details carry almost all of the weight. Two of them is usually enough to change the answer, and all four is better.

1. What you want. The verb and the thing you want back: rewrite, summarize, plan, compare, explain, draft.
2. Who it is for. Your manager, a client, your landlord, a nine-year-old, yourself in six months.
3. How long. One sentence, a paragraph, five bullets, one page. Left unsaid, most models pick a length nobody asked for.
4. Facts only you know. The date, the budget, the history, the thing you cannot say out loud in the message itself.

Here is one real message, twice. Before: "Write an email to my team about the delay." After: "Write a short email to my six-person team explaining that the client launch moved from March 3 to March 17. Three sentences, calm, no apology spiral. They already know the client changed the brief, so do not re-explain that part." The extra detail is not politeness. It is the six people, the two dates, and the one thing they already know, and it removes about three rounds of correcting afterwards.

The follow-up is where the value is

Most people send one message, read the answer, decide AI is either amazing or useless, and close the tab. That single-shot habit is responsible for most of the disappointment. The first answer is a first draft written by someone who has never met you, and the whole point of a chat is that you get to reply.

The conversation remembers everything you have already said, so you never rewrite the original request to fix it. You can type three words. These are the corrections that handle the usual problems, and you can copy them as they are.

• Too long: Half the length. Cut the wind-up, keep the specifics.
• Too formal: Say it the way I would say it out loud to a colleague I like.
• Wrong angle: You wrote it as a [X]. I need it as a [Y]. Try again with that framing.
• It made something up: Which parts of that did I tell you, and which did you fill in yourself? List the ones you filled in.
• It does not sound like me: Here are two things I wrote myself. Match this voice instead of a corporate one. [paste two of your own paragraphs]

The fourth line is worth turning into a habit. These systems produce fluent text whether or not they know the answer, so asking which parts were filled in surfaces the invented ones faster than checking every sentence yourself. It works on figures too: ask where a number came from and you will often get an honest "I estimated that." You will notice none of these corrections are polite. Politeness is optional here. What decides the next answer is whether the correction names something to change, which is why "make it better" does nothing and "cut the first paragraph" works every time.

Ten opening lines you can copy

Pick the one that matches something actually sitting on your desk today. Each is written to be pasted as it is, with your own material added underneath.

1. Rewrite this email so it is shorter and warmer, and keep the request clear: [paste]
2. Explain this document in plain language, then tell me what I should be worried about: [paste]
3. Plan a [event] for [number] people on [date] with a budget of [amount]. Give me a timeline, not ideas.
4. I am deciding between [A] and [B]. Build a table of tradeoffs, then tell me what I have not considered.
5. Teach me [topic] in five minutes starting from zero, and stop twice to check I followed you.
6. I need to tell [person] [awkward thing]. Draft it kindly, three sentences, no over-apologizing.
7. Summarize this in five bullets for someone who will only read the first two: [paste]
8. Translate this into [language] and keep the tone friendly but professional: [paste]
9. Here is my reasoning for [decision]. Argue the strongest case against it: [paste]
10. I need to write [thing] and I do not know how to start. Ask me four questions first, then draft it.

Number ten is the one for a cursor that is still blinking. Handing the blank page back as a short interview is the fastest way out of it, and the four questions it asks will usually be the four things you had not decided yet. If you want a longer supply organized by the kind of day you are having, AI for everyday tasks has a week of them.

When the answer is bad, and what to do about it

Bad answers are not random noise. They arrive in a few recognizable shapes, and each one has a repair that takes a single message.

• The confident invention. A statistic, a citation, a policy, or a product feature that does not exist, delivered in exactly the tone it uses for things it knows. The repair: ask it to mark every claim it is not certain about, then check those yourself, and never pass on a number you have not seen at its source. Why AI gets things wrong explains where the invented ones come from.
• The bland average. Technically correct, entirely generic, reads like a brochure for nothing in particular. The repair: give it something specific to react to. Your draft, your constraints, your own sentences. Vague input produces vague output, reliably.
• The agreeable mirror. You asked why X is better than Y, and it argued for X, because that is how you framed it. The repair: ask the opposite question in a new message and read both answers side by side before you decide.

If you have repaired the same answer twice and it is still wrong, stop rephrasing. Two rewrites usually fix a wording problem. A fourth rarely fixes a model problem, and models fail in different directions: one invents sources while another refuses to commit to anything, one writes warmly while another stays stiff no matter what you tell it. Sending the identical message to a different model is often a five-second fix for something you were about to spend twenty minutes on. That is worth arranging even if it just means keeping a free account on a second provider; Whizi is the version of that where the switch is a dropdown instead of a second 20 dollars a month. How to choose an AI model covers which one to reach for first, and once correcting answers feels routine, prompt engineering for beginners turns these repairs into a structure you can reuse.