In short
Start with one tool, one real task and about fifteen minutes. This page walks you through seven steps, from understanding why AI makes things up to building a small assistant of your own. Each step has an outcome, a time estimate and three things to do today. Total: about five hours, on a free account.
Key points
- Seven steps, about five hours in total. Spread them over two or three weeks.
- A free ChatGPT or Claude account is enough. No code, no card, no course to buy.
- Every step ends with something you did, not something you read.
- Do them in order. Step 5 only makes sense once you have caught the tool being wrong in step 1.
Before you begin
Pick one tool and stay with it for two weeks. ChatGPT or Claude, it does not matter which. The free tier of either covers everything on this page. People who bounce between tools in their first month end up comparing menus instead of learning anything. I have watched it happen to marketing teams with budget for both.
You do not need to be technical. There is no code here, no API and no credit card. The skill you are learning is describing what you want clearly, which is a writing skill. If you can brief a freelancer, you can do this.
One more thing. Open a note somewhere, a doc or a notebook, and write down every time the tool gets something wrong. You will need it in step 5, and after a month it becomes the most useful document you own.
Step 1 - Understand what it actually is
After this you can explain, in one sentence, why AI makes things up. About 10 minutes.
An AI assistant does not look anything up. It predicts the next likely piece of text, again and again, from patterns in a very large amount of writing. That is the whole trick. It explains most of what surprises people later: why it sounds certain when it is wrong, and why the same question can get a different answer tomorrow.
The word for a confident wrong answer is a hallucination. You will meet one this week.
- Ask it: "Explain how you produce an answer, as if I were ten." Read the reply twice.
- Open a second chat and ask the same question. Put the two answers side by side. They will not match, and that tells you something about what you are dealing with.
- Ask it something small and specific about your own town or your own company. Check one detail. There is a fair chance it is wrong, and that is the point of the exercise.
Practice: this week, catch it once. One confident statement that turns out to be false. Write it in the note. That small pause before you trust an answer is worth more than anything else on this page.
Step 2 - Have your first real conversation
After this you have used AI for something that mattered today. About 15 minutes.
Not a test question. Not a poem about your dog. A real task from your actual day, because the only way to find out what the tool is good for is to hand it something you needed done anyway.
- Paste in a long email you received and ask for a three line summary. Then ask what the sender actually wants from you. The second answer is usually the useful one.
- Ask it to draft a reply in your tone. It will come back too formal or too chirpy. Say exactly what is wrong with it and ask again.
- Pick something from your job you have nodded along to for years without really understanding. Attribution models, say, or how a CDN works. Ask it to explain, then ask a follow up the way you would in a meeting.
Practice: one real task a day for a week. Small is fine. The habit is the outcome, not the tasks.
Step 3 - Learn to brief it properly
After this your prompts include a role, the situation, an example and constraints. About 30 minutes.
Most disappointing answers come from thin requests. "Write a headline" gives the tool nothing to work with, so it reaches for the average headline it has seen a million times. Treat it like a capable new colleague on their first morning: quick, willing, and with no idea who you are or what your company sells.
A good prompt is a brief, and a decent brief covers five things.
| Include | What it looks like |
|---|---|
| A role | "You are a B2B copywriter." |
| The situation | "The audience is finance directors at mid-size Indian firms. They distrust marketing." |
| One outcome | "Six headlines for a webinar invitation." |
| An example | "Here is one that worked last quarter: ..." |
| Constraints | "Under nine words. No puns. No exclamation marks." |
- Find a prompt that disappointed you last week. Rewrite it with all five parts and compare the two answers.
- Add one line at the end: "Before you answer, tell me what you would need to know to do this well." Then answer its questions. This one line fixes more bad output than any other trick I know.
- Ask for three versions at different lengths. Say which you prefer and why. Watch the next answer improve.
Practice: never accept a first draft. Name one specific problem and ask for one revision. Two rounds of that beat ten minutes of polishing the original prompt.
Step 4 - Put it to work on the jobs it is good at
After this you know the five jobs it does well, and what to keep away from it. About an hour.
Summarise. Draft. Rewrite. Extract. Explain. That is the honest list. It is excellent at all five and unreliable outside them. Arithmetic, anything that happened recently, and anything where a wrong answer costs money are not on the list, and it will attempt them anyway with a straight face.
- Summarise a document you have been avoiding. A forty page PDF, an annual report, a contract. Ask for the three things you need to know and one thing that looks like a risk.
- Paste in messy meeting notes and ask for every date, name and action item as a table. Then check the table against the notes. Extraction is where it is most reliable, and it still misses things.
- Take something you wrote for one audience and have it rewritten for another. A technical update rewritten for a client, for instance. Fix what it gets wrong about your client, because it will.
Practice: for one week, before you start any writing task, ask whether it is one of the five. If it is, start with the tool. If it is not, do not.
Step 5 - Learn when it is wrong
After this you can check a claim in two minutes, and you know what never to paste in. About 20 minutes.
This is the step that separates people who use AI well from people who end up in a screenshot. The tool is confident about everything, so confidence tells you nothing. Every number, name, date and quotation you plan to put in front of someone else gets checked at the source, and the source is never the tool itself.
- Ask for five statistics on a subject you know well. Check each one. Count how many are wrong, or are real numbers attached to the wrong year.
- Ask it to argue against something it just told you. Watch how fast it folds. That is not honesty, it is agreeableness, and it works in both directions.
- Find the data controls in your tool's settings. Decide, deliberately, whether your conversations may be used for training. Then decide what you are comfortable pasting in at all.
Never paste: customer personal data, anything under NDA, passwords or API keys, unreleased financials, or anything you would be uncomfortable reading back to your client.
Practice: open the note from step 1. Every mistake you catch goes in it. Read it before you trust the tool with anything that matters.
Step 6 - Go beyond the chat box
After this you can give it your own documents and stop re-explaining yourself. About 45 minutes.
Most people use these tools like a search box: one question, from a blank page, every time. The useful version is different. It knows who you are, it has your documents in front of it, and it works from your material rather than from the whole internet.
- Upload a spreadsheet or a PDF and ask questions about the file, not about the world. "Which three rows dropped the most month on month?" is a good first one.
- Set your custom instructions. Who you are, what you do, how you like answers written. Two paragraphs, once, and you never explain yourself again.
- Create a project in Claude, or a custom GPT in ChatGPT, for one recurring piece of work. Put the reference documents in it.
Practice: move one recurring task into a saved project this week. If you write a weekly report, that is the one.
Step 7 - Build one small thing
After this you have made something reusable, and the whole point lands. About two hours.
Pick one task you do most weeks. Build an assistant that does it, with your instructions, your examples and your reference documents. It does not need to be clever. It needs to be yours, and it needs to work twice.
An example from my own week: a project that turns a raw Search Console export into a short client note. It has instructions on tone, two past notes as examples, and one rule, that it never states a number that is not in the export. It is not impressive. It saves me an hour every Monday.
- Write down the task, the inputs and what a good output looks like. If you cannot describe the good output, the tool cannot produce it.
- Build it as a project or a custom assistant, with two or three examples of finished work.
- Run it on a real task next week. Fix the one thing that annoyed you most and leave the rest.
Practice: if you are still using it three weeks later, it was worth building. If not, you learned something about the task for the price of an afternoon.
What comes after this
You now know how to use AI. If you have a website or a brand, the more interesting question runs the other way: when someone asks an assistant about your subject, how does it decide whether to mention you?
That is what the rest of this course is about, and Getting found by AI is where it starts.
FAQ
Where should I start with AI? With one tool and one real task, not with research about which tool is best. Open ChatGPT or Claude, paste in something you are working on today, and ask for help with it. Step 2 on this page takes fifteen minutes and teaches more than a week of reading comparisons.
How long does it take to learn AI? About five hours to become useful at work, spread across the seven steps on this page. Spread them over two or three weeks so each habit has time to stick. There is no finish line, but roughly a month in it stops feeling like a novelty and starts feeling like a tool.
Do I need to pay for AI to learn it? No. Every step on this page works on a free ChatGPT or Claude account. Paid plans give you more usage and faster models, which matters once AI is part of your daily work, not while you are learning.
Which AI tool should a beginner use? ChatGPT or Claude. Either one. Pick one and stay with it for two weeks, because switching while you are still learning means you compare interfaces instead of building skill.
Do I need to be technical to use AI? No. Nothing on this page involves code, an API or a credit card. The skill you are learning is describing what you want precisely, which is a writing skill rather than a technical one.
What is AI actually good at? Five things: summarising, drafting, rewriting, extracting information from messy text, and explaining things. It is excellent at all five. It is unreliable at arithmetic, at anything that happened recently, and at any task where a confident wrong answer costs money.
How do I write a good prompt? Treat it as a brief, not a question. Say who the tool should act as, describe the situation and the audience, ask for one outcome, show an example of what good looks like, and state the constraints such as length, tone and format. Then ask it what it still needs to know before it answers.
What should I never paste into an AI tool? Customer personal data, anything covered by an NDA, passwords or API keys, unreleased financial figures, and anything you would be uncomfortable reading back to a client. Check your tool's data controls and decide whether your conversations can be used for training.
What should I do after these seven steps? Learn how AI systems decide what to say about your brand or website. That is the subject of the rest of this course, starting with Getting found by AI.
- Rewritten. Every step now has a worked example, and the FAQ covers what AI is good at, prompts and what not to paste.
- First published.