Your first conversation with AI: the complete guide for absolute beginners

An empty box, a blinking cursor, and underneath it the line “How can I help you today?” Most people type “hi,” get a polite reply, don't know what to do next, and close the tab feeling like the whole fuss probably isn't for them. But the problem isn't the tool, and it isn't you — you're just missing one piece of information nobody says out loud: what to actually ask AI for, and how to tell whether what it hands back is worth using.

The whole guide fits into one sentence. Don't treat AI like a search engine or a magician — treat it like a very fast, well-read colleague who's occasionally, confidently wrong, someone you hand work to and whose output you always sign off on yourself. A search engine finds you the page where the answer lives. A model just writes the answer straight out — including the times it doesn't actually know it. The difference between someone AI saves hours for and someone it costs their credibility comes down to exactly this: the first person knows what to check.

This guide walks you from zero: choosing a tool, your first conversation, ten concrete tasks depending on what you do, and finally the checks that expose good-sounding nonsense. Every phase has copy-paste prompts — just fill in the brackets. You don't have to read it in one sitting. Get through the first two phases today, the third tomorrow, and come back to the fourth once you have a first answer you're not sure about.

A typical scenario

Jana runs operations at a distribution company, has twenty-five people reporting to her, and around sixty emails a day. She tried AI last year: typed “write me something about team motivation,” got five paragraphs about how motivation is key and requires an individual approach, and decided it was a toy for marketers. She closed the window and didn't come back for a year.

The second time she tried it differently — not with a topic, but with actual work sitting on her desk. She took the notes she'd typed up during a meeting and had them turned into a task list with owners and deadlines. It took two minutes, and the output was usable after fixing two names. The next day she had AI draft the first version of an unpleasant reply to a supplier. On the third day she had it pull the deadlines and penalties out of a fifty-page contract, then read those spots herself.

A month in, her day looks like this: the morning hour spent on email has shrunk to roughly a third, because AI writes the first draft of her replies and she edits them. Prepping for Monday's meeting takes ten minutes instead of fifty. Twice, the model confidently handed her a number it had made up — she caught both, because she checks every number and every reference against the source. That's the entire difference between her first attempt and her second: the tool didn't change. What changed was the request, and the habit of checking. The rest of this guide is the long version of what Jana does.

Phase 1: choosing a tool without a religious war

The most common reason people never start is, ironically, the fear of picking wrong. For a beginner, the differences between the major tools are far smaller than the difference between “using it” and “not using it.” All of them can answer questions, write and edit text, summarize documents you paste in, read images and scans, and search the web. If reading an hour of comparison tables tonight delays you, that hour has cost you more than a better choice would ever save.

Three big options and one specialist

ChatGPT is the most widely used, so you'll find the most guides and the most coworkers you can ask for help. Claude (at claude.ai, on web, mobile, and desktop) tends to be strong with long documents and following instructions closely; it has Projects with persistent context, Artifacts (small apps and documents built right inside the chat), and connectors to Gmail, Calendar, Drive, or Notion. Gemini has the edge if you live in Google Docs, Gmail, and Drive. And NotebookLM is a different kind of tool altogether: it answers strictly from the material you upload to it, which makes it the safe choice when you need certainty the model isn't inventing anything from outside sources.

Practical advice for week one: pick one and stick with it. The habit is worth more than the optimization. A month from now, once you know what you actually want from AI, you can try the same request elsewhere and compare — before that, you have no real basis for judging.

Free or paid

Start free. The free tiers are enough to find out whether a tool fits how you work. A paid account starts making sense the moment you hit one of three walls: you run out of daily message limits, you need to work with longer documents and files, or you need to feed the tool data you actually care about protecting. Sensitive information — employee personal data, health information, unreleased contracts, access credentials — belongs only in a paid account with contractual data protection, and even there it's better to anonymize it before you paste it in. Don't put it in a free chat at all.

A ten-minute comparison test

When you're ready to make the choice properly, don't read reviews — give both tools the exact same piece of your own work and compare the results with your own eyes. Use a task where you can judge quality yourself.

I want to compare two AI tools on my real work. I'll give both
the same task and compare the outputs.

My profession: [profession]. A typical task I do every week:
[description of the task, e.g. answering customer questions].
Here's the specific assignment plus the source material:

[paste source material — an email, notes, a piece of text]

Do the work, then add three things:
1. Everything you had to guess because I didn't tell you.
2. Which parts of the output I should verify before using it,
   and exactly where.
3. One question whose answer would most improve the result.

No intro, no closing summary — start straight in on the work.

You'll get the finished task back, plus three points that reveal how each tool thinks. Points 1 and 3 are the interesting ones — a tool that quietly fills in half your assignment and doesn't admit it will cause you problems long-term. Paste the same brief into the second tool and decide based on which output you'd rather edit.

Phase 2: the first conversation — what to write and what to expect

Your first message decides whether you get work or an essay. The rule is simple: don't hand it a topic, hand it a task. “Something about team motivation” is a topic, and it returns filler. “I have a Monday meeting with five people after we lost a bid — build me a structure so it doesn't turn into a blame session” is a task, and it returns something you can use.

The anatomy of a first message

A request needs four things, and you can write all four without any preparation. Who you are and what situation you're in (one sentence is enough). What exactly you want done. What you have to work with — text, numbers, notes, pasted straight into the message. And what the output should look like: bullet points, a table, a ten-line email, three options. Anyone who wants to take this skill further will find it laid out in the tip on role, context, and format; for a first conversation, answering those four questions is enough.

Don't be afraid to write long, in plain English. There are no magic words or secret syntax. A longer, specific message works better than a short, general one, because the model knows nothing about your situation — until you tell it, it's guessing.

Let it ask you questions

The fastest way to turn an average answer into a good one is to let it ask questions before it starts working. Most people never think of this, because they talk to AI like a vending machine instead of a colleague.

I want you to [describe the task, e.g. draft an email to a
client who's a month overdue on an invoice].

Don't write anything yet. First ask me 5 questions whose
answers would most change the result — the kind you'd
otherwise have to guess. Be specific, not generic ("what
tone" is a bad question; "should it sound conciliatory, or
already like a final warning" is a good one).

Once I answer, write the result, and below it list the
assumptions you still had to make on your own.

You'll get five questions back — usually two you could've answered yourself, and three you honestly hadn't thought of. Watch for one classic failure: the model sometimes answers its own questions and just keeps going. When that happens, say “hold on, I'll answer those” — and next time, put the line “don't write anything yet” on its own line in the prompt.

The answer is raw material, not a finished result

Treat the first answer like a colleague's first draft: usually about eighty percent right, with the rest needing polish. Polishing rarely takes more than one sentence — “cut it in half,” “add the exact numbers from the source material, nothing else,” “shift the tone to something an angry customer can tolerate.” That's the entire trick behind the conversation, spelled out in the tip talk to AI, don't shoot blind. Anyone who walks away disappointed after the first answer misses most of the value.

What to expect from a first conversation: a fast first draft, structure where you had chaos, explanations of things you don't understand, and the nerve to actually start. What not to expect: a finished piece ready to send unread, reliable numbers, knowledge of your company, or exact citations without verification.

Phase 3: ten first tasks by profession

Theory ends here. The fastest way to get AI actually saving you time is to take work you have to do today anyway and do it with AI. The following ten tasks are chosen so you can tell immediately whether the output is good — you don't need an expert to check it, you'll know yourself.

1. A first draft of an unpleasant email. A rejection, a chaser, an apology. The best ratio of time saved to risk taken.
2. A summary of a long document with pointers to where each claim comes from in the text — and then you actually read those spots.
3. An explanation of something you don't understand, three ways: simply, with an analogy, and through an example.
4. Meeting notes turned from your messy scribbles into decisions, tasks, who owns them, and by when.
5. A constrained brainstorm — fifteen ideas, five of them cheap, five fast, and five that nobody else would suggest.
6. Cleaning up text you dictated or typed in a hurry, without changing the content.
7. Prep for a hard conversation: what the other side will likely push back on, and what to say to it.
8. Turning chaos into a table — proposals scattered across paragraphs, turned into a side-by-side comparison.
9. Quiz questions on material you're supposed to know, to test whether you actually understand it.
10. Breaking a big task into a first small step you can finish in ten minutes.

Which one should you try first? Here are three full prompts for three typical situations. Take the one closest to your life and replace the brackets.

First task for a manager: a meeting with an actual output

You're an experienced team lead. I manage [number] people in
[field], and on [day] I have a [length]-minute meeting.
Topic: [meeting topic].

Context that I know and you don't:
- what happened: [description of the situation]
- what I need to come out of the meeting: [decision / division
  of work]
- where I expect pushback, and roughly from whom: [description]

Prepare me:
1. An agenda broken down by minutes so it actually fits.
2. Three questions to open the meeting with that get others
   talking, not me.
3. Two objections that are likely to come up, and how to
   respond without putting anyone down.
4. A format for the notes: what I should walk away with
   written down.

No generic advice about leading people — just things tied to
this specific meeting.

You'll get an agenda you can walk into the room with. Check the minute allocations especially — models plan optimistically and love cramming six items into thirty minutes. And rewrite points 2 and 3 in your own words, since you'll be saying them out loud.

First task for a student: material you don't understand

I'm a student of [field], year [year]. I don't understand
[topic], and in [number] days I have [an exam / a test /
a paper] on it.

Explain it to me in three passes, back to back:
1. Simply, as if I'm hearing it for the first time — no
   jargon, 10 sentences max.
2. With an analogy to something from everyday life, and note
   where the analogy breaks down.
3. Through one concrete example, worked or walked through
   step by step.

Then give me 5 check questions, easiest to hardest, but don't
answer them yet — I'll answer, and you tell me where my
understanding is off. Ask about understanding, not
definitions.

This prompt does three things at once: it explains, it shows the limits of the analogy (typically where misunderstandings are born), and it tests you. One warning: for numbers, dates, and formulas, check the explanation against your course notes or textbook — the model can explain beautifully and be confidently wrong at the same time.

First task for a freelancer: the proposal you don't want to write

I'm a freelance [profession]. A [type of client] client
reached out about [description of the job]. Here's what I know
from their inquiry:

[paste the inquiry text or call notes]

Write me a first draft of a proposal email:
- scope of work broken into phases, so it's clear what they get
- what's NOT included in the price, so it doesn't get assumed
  (no amounts — I'll add those myself)
- assumptions on the client's side: what they need to deliver
  and by when
- what happens if the scope expands mid-project
- one closing sentence prompting the next step

Tone: businesslike, confident, no apologizing and no
superlatives. Under 300 words. Then, separately, list 3 things
from the inquiry that are unclear and that I should nail down
before sending.

You'll get the skeleton of a proposal and — more valuable — a list of unclear points you'd otherwise only discover halfway through the job. Don't put prices in the prompt, and don't have it suggest any — those are yours to set. And a rule that recurs throughout this guide: AI proposes, the human sends. No email goes out without you reading it in full.

Phase 4: how to tell a good answer from plausible-sounding nonsense

This is the most important phase of the whole guide, and the one you can't skip. Language models are machines for generating the most probable continuation of text. That means when they don't know an answer, they don't say “I don't know” — they generate text that looks like a correct answer. A made-up citation gets a credible author, a made-up clause gets a credible number, a made-up statistic gets a credible value. It's dangerous not because the model gets things wrong, but because it gets things wrong in the exact same tone it uses when it's telling the truth.

Three questions to ask of every answer

First: could the model actually know this? Anything about your company, your client, yesterday's prices, or the contents of your hard drive — the model doesn't know it unless you gave it that information or it found it on the web. If it shows up in the answer anyway, it's made up.

Second: is this verifiable, or just nicely phrased? Split the answer in your head into two buckets — phrasing (the model is genuinely good at this) and facts, numbers, names, and citations (these get verified). The sentence “I'd recommend starting with a pilot in one department” is an opinion and rests on your own judgment. The sentence “according to a 2024 study, this increases productivity by 37 percent” is a claim you either track down or cut from the material.

Third: what happens if this is wrong? A badly worded brainstorm costs nothing. A wrong number in a client proposal costs you the job. Scale how carefully you check based on the consequences, not on how confident the answer sounds.

A prompt for auditing its own answer

The model can't reliably fact-check itself, but it's reliably good at sorting out which part of an answer is which — and that saves you most of the work.

Take your previous answer and break it down. Return a table
with columns: claim | type (fact / opinion / estimate / filled
in from my request) | how confident you are | where I should
verify it.

Rules:
- list every number, date, name, law, and source reference as
  its own row
- where you estimated or guessed something, say so openly
  instead of justifying it after the fact
- at the end, list what you'd need from me so you wouldn't have
  to guess at all

Don't rewrite anything, and don't defend the answer.

You'll get a clear map of the answer that immediately shows half the “facts” are actually estimates. The last line of the prompt matters — without it, models tend to slide into defending what they wrote. And a warning: even this audit isn't proof of anything. It's a sorted list of things to verify; the verification itself is yours to do, following the process in the tip fact-checking with AI.

Have it show you where in the source it says that

For document work, there's a simple test that exposes a made-up summary within a minute.

I'm attaching a document called [name].

Summarize it in 10 bullet points. For EACH bullet, note which
part of the document it comes from — chapter, section, or page
number, plus the first few words of that passage, so I can find
it.

Where a claim draws on more than one spot, list all of them.
Where a bullet would be useful but the document doesn't support
it, DON'T write it — instead, add it to a "not found in the
document" list at the end.

Don't add anything from general knowledge, even if it happens
to be correct.

You'll get a summary you can double-check by searching the original file. Spot-check three bullets — if those hold up, the rest usually does too; if even one points to a passage that isn't there, stop trusting the whole summary. Be especially careful with scanned documents: the model reads images and handwriting fine, but on low-quality scans it mixes up digits.

Limits, straight from the source

A useful, slightly fun beginner prompt that'll save you a few disappointments:

Explain your limits to me so that, as a beginner, I know when
to trust you and when not to. For each point, give an example
of a question where it actually comes up in practice.

Cover at least this:
- where your knowledge comes from and how current it is
- what happens when I ask about something you don't know
- how good you are at counting and at long tables of numbers
- what you remember from our earlier conversations and what you
  don't
- where you're most often wrong without me noticing
- what three kinds of tasks I should avoid giving you entirely

Be specific and skip the marketing — self-criticism is fine.

You'll get a surprisingly honest overview worth revisiting. Take it as a rough guide: the model talks about itself in probabilities too, and it doesn't know some details of its own setup. Still, the essentials come through — leave long calculations to a spreadsheet or a script, verify current prices and dates on the web, and never send anything to a customer or into a decision without reading it yourself first.

Phase 5: what to expect and what not to — a realistic map

After the first week, it's worth resetting your expectations. The breakdown below is experience, not theory.

Where AI wins from day one

Work where the input is text and the output is text, and where you can judge quality by eye: first drafts of emails and documents, rephrasing, shortening, tone shifts, turning notes into structure, summaries, explanations, brainstorming, quiz questions, checklists, tone-aware translation. This also covers solo work you'd otherwise have to do alone — the model replaces the colleague you can ask a dumb question.

Where caution is needed

Anything where the output leaves your head before you've checked it: numbers, legal and tax questions, health matters, citations, names, and references. The model isn't necessarily wrong here — but a mistake has a real cost. The rule is simple: sending, paying, deleting, signing, and stamping are always approved by a human.

Where AI fails

Precise arithmetic over a large table. The model isn't a calculator — it generates numbers as text, and with a long table it only ever sees a slice of it. When you need to compute something, have it write a formula or a script that does the calculation, and have it explain what the formula does. Also: knowledge of recent events without web search, anything about your company that you never told it, and memory across conversations, unless you've turned that on via projects and persistent instructions.

A prompt for a realistic first week

I'm a [profession], and I'm just starting with AI. My typical
week looks like this:

[describe 5-8 activities you repeat, and roughly how long each
takes]

Suggest a plan for my first 5 working days: one task per day
to try with AI. For each day, note:
- which of my activities it replaces or speeds up
- exactly what I should paste into the chat as source material
- how I'll know the output is good without having to send it to
  an expert for review
- how much time it can realistically save — be conservative,
  not optimistic

Order the tasks from lowest risk to biggest payoff. Don't
include anything where I'd have to paste in personal data or
non-public documents.

You'll get a plan you'll actually keep three or four days out of — the fifth tends to be ambitious. Cut the savings estimates in half until you've measured them yourself: the first week costs more time than it saves, and that's normal.

Phase 6: turning trial into habit

Most people don't stop using AI because it doesn't work — they stop because they forget it exists. The habit gets built with three things, and all three fit into one afternoon.

Custom instructions: write once, apply forever

If you find yourself correcting AI the same way every time (“shorter,” “in English,” “skip the intro”), write that down once, permanently. Most tools have a setting that applies to every new conversation — the how-to is in the tip on custom instructions. To start, a few sentences are enough: who you are, what you do, how you want answers (length, tone, language), and what the model should avoid.

Help me write custom instructions I can set once so they apply
in every conversation.

About me: [profession, field, who you work for].
What I most often want from AI: [3-5 types of tasks].
What annoys me about answers: [e.g. long intros, jargon,
bullet points everywhere, hedge phrases].
How I want answers to sound: [tone].

Write me a draft, under 150 words, phrased as instructions to
you, not as a description of me. Add 3 sentences you'd
deliberately leave OUT, because they'd contradict each other or
constrain the model for no reason.

You'll get finished text ready to paste into your settings. The last paragraph is there on purpose: beginners love overstuffing their instructions and then wonder why the answers sound robotic.

Save what worked

A prompt that once got you a great result is a tool, not a throwaway note. Save them wherever — notes, text expander snippets, Notion. The method is covered in the tip a prompt library. A practical compromise to start: one notes file called “AI prompts,” where you copy only the ones you've used at least twice.

Add complexity one layer at a time

Once you've had a month of basic use under your belt, add one thing: persistent context through projects (upload a company description, writing style, and sample documents once, and stop pasting them in every time), then file handling, then maybe connectors to mail and calendar, or scheduled routines that prep something for you every morning. Don't rush it. Setting up connectors before you know how to write a good request just speeds up how fast you produce unusable output.

Common mistakes

• Handing it a topic instead of a task. “Something about marketing” returns an essay nobody asked for. The request needs to say what you'll actually do with the output.
• Walking away after the first answer. The first draft is raw material. Two sentences of follow-up usually move it from “average” to “usable.”
• Trusting numbers and citations. A made-up number looks exactly like a correct one. Anything that feeds a decision or leaves the company gets checked against the source.
• Pasting sensitive data into a free chat. Personal data, unreleased contracts, and access credentials belong only in a paid account with contractual data protection — and even there, anonymized.
• Letting AI send, sign, or delete something. The draft comes from the machine; the decision is yours — for mail, payments, and deleting files alike.
• Spending the first week comparing tools. The gap between tools is smaller than the gap between a good request and a bad one. Pick one and start.

The best tools

• ChatGPT — the most widely used choice, the most guides, and the most people around you to ask when you get stuck.
• Claude (claude.ai, web, mobile, and desktop) — strong with long documents, Projects with persistent context, Artifacts as small apps built right into the chat, and connectors to mail, calendar, and documents.
• Gemini — the natural first choice if you work mainly in Docs, Gmail, and Drive.
• NotebookLM — when you need certainty that the answer comes strictly from your own uploaded material; good for studying and research.
• Whatever notes app you already have — for a library of prompts that proved themselves. The best tool is the one you actually write in.

What you get out of it

• Time: realistically thirty minutes to two hours a week on text work, already in the first month — mostly on emails, summaries, and prep. Not because AI types faster than you, but because it removes the friction of starting from a blank page.
• Money: indirectly — fewer hours on admin means more hours on billable work or on moving projects forward. Freelancers see this fastest.
• Peace of mind: the procrastination on tasks where you didn't know how to start disappears, and with it, the nagging feeling that you didn't start again today either.
• Quality: a second pair of eyes on a document, check questions before a meeting, and a checklist before hitting send — things that used to be available only to people who had a colleague nearby to ask.

Pro tip

The next time — or the third time — AI hands you something that looks great, try one sentence: “Take what you just wrote and find the three weakest spots — where an experienced person in this field would call it shallow or off.” The model is surprisingly good at criticizing its own output, and you build the habit of looking at answers critically before you rely on them.

And the closing rule that holds from your first conversation to the day you're using AI daily: the machine proposes, the human approves. It's your name under that output, not the model's.