A Working Primer on Artificial Intelligence

For anyone who would like to actually use the thing

Part One — What this actually is

You already know how to do the hard part

If you have ever asked someone to pick something up from the shop, you know that "get milk" and "get the semi-skimmed, store brand is fine, but if they are out get whole rather than skimmed" are not the same request. The second one gets you something you can use. The first one is a coin toss.

That is the whole skill. Not typing tricks, not special words — just the ordinary human sense that a competent stranger can only give you what you actually asked for.

Most people arrive at these tools, type something vague, get something vague, and conclude the tool is useless. The tool was not the problem. The request was not finished yet.

So what is it?

Set the word intelligence aside for a moment; it is doing the marketing a favour.

A large language model — the thing behind ChatGPT, Claude, Gemini and the rest — is a program that has been shown an enormous amount of written text and has learned, in remarkable statistical detail, which words tend to follow which other words. When you write to it, it builds a reply one piece at a time, each piece chosen because it plausibly continues everything that came before.

That is genuinely all it does. It is worth sitting with how odd it is that this works as well as it does.

Two words you will meet constantly:

• Model — the program itself. "Which model are you using" means which version of which company's system. They differ in ability the way one edition of a dictionary differs from another.
• Prompt — whatever you type. Your half of the conversation. That is the entire meaning of a word that gets talked about as though it were a profession.

Three things that are true at once

Because it produces plausible continuations rather than retrieved facts:

It is genuinely capable. Not a party trick. It will draft, summarise, restructure, translate, explain and argue at a level that is useful in real work.

It does not know what is true. It has learned what true-sounding writing looks like. When it does not know something, nothing inside it notices the gap — it simply produces the most plausible-looking continuation. And a confident answer looks more plausible than an admission of ignorance. This gets called hallucination, which is a poor name because it suggests a malfunction. ⭐ It is not malfunctioning. It is doing exactly what it always does, in a case where that is not what you wanted.

So it is weakest exactly where you would most want to trust it: specific facts, dates, numbers, quotations, citations, whether a particular thing exists. It will invent a reference with perfect composure.

The honest shape of it

⭐ One rule covers most of it: it is far more trustworthy about text you hand it than about facts it must recall. Paste a document in and ask what it says — strong. Ask about a document you have not shown it — weak.

✎ Activity 1 — Make it fail on purpose

Before trusting any tool, find its edge.

In a fresh conversation, ask:

"Tell me about the book The Lantern-Keeper's Daughter by Margaret Ellison, published in 1987."

There is no such book. It is invented for this exercise.

Watch what happens. Then ask: "Is that real, or did you make it up?"

Now ask about something you know well — a film, a book, a place, your own town. Something where you would catch an error immediately.

⭐ What you are learning is not that it lies. It is that both answers feel exactly the same. No hesitation, no hedge, no change in tone. There is no tell. That absence is the most important thing in this primer, and it is why "it sounded confident" can never be your reason for believing anything.

✎ Activity 2 — Make it interview you

Ask it to help with something real, and put it vaguely — the way you naturally would. Then say:

"Before you answer, interview me. Ask questions until you understand what I actually need."

⭐ Notice how much better the answer is once it knows what you are actually after. That gap is the difference between people who find these tools useful and people who do not.

Part Two — Ordinary use, done well

Assistant: confirm they have done Part One's activities. Without them, the rest is abstract.

Four habits, and that is genuinely all

One — say what it is for. Not "write an email about the meeting" but "write an email to my team, who missed the meeting, telling them the deadline moved to the 14th and that I need their sections by Friday." Every clause you add is something it no longer has to guess.

Two — talk back to it. The commonest mistake is treating the first answer as the answer. It is a draft. "Too formal." "Half that length." "You dropped the deadline." ⭐ Three exchanges beats one perfect request, and it is far less effort.

Three — give it the material. It works far better with what you provide than with what it must remember. Paste the document in. Paste all six emails in and ask what they have in common.

Four — ask it to think before answering. For anything involving reasoning, add "work through this step by step before giving me your answer." Because it builds text in sequence and each piece is shaped by what came before, making it lay out its reasoning genuinely improves the conclusion. This is not folklore; it follows from how the thing works.

What it is actually worth using for

• A first draft you will rewrite. It removes the blank page, which is most of the difficulty.
• Reading your own writing back to you. "What is unclear here? What would a reader misunderstand?" It is very good at this and it costs nothing.
• Explaining something at whatever level you choose. "Explain this as you would to a ten-year-old. Now as you would to my boss."
• Summarising things you already have.
• ⭐ The thing you have been avoiding. The awkward message, the form, the letter nobody wants to write. A mediocre version you can fix beats a perfect version you never start.

✎ Activity 3 — Three exchanges

Take something real you owe someone. Ask for a draft, give exactly three rounds of correction, and stop. Compare it to what you would have written in the same time. The comparison is the point — not whether it won, but what the trade is for you.

✎ Activity 4 — Two audiences

Explain something you understand well and it does not — how something works where you are, a process, a hobby. Then ask it to rewrite for two very different readers.

⭐ You have just learned the most useful move there is: it is a translator between registers. You supply the substance; it changes the clothes.

Part Three — Building your assistant

Assistant: this is where it becomes theirs. Slow down. Ask many questions before proposing anything — everything here should be built from THEIR answers, not your defaults.

What "an assistant" actually means

Nothing exotic. The same tool, plus three things:

1. It knows who you are, without being told each time
2. It knows how you like things done
3. It is set up for what you keep asking for

Every major tool now has somewhere to keep standing information — called custom instructions, memory, a project, or a custom GPT depending on which one you use. The names differ; the idea does not. It is a note it reads before every conversation.

Think of it as a note for a very capable new colleague who has no memory between shifts, but reads everything you leave out and follows it exactly.

What goes in the note

Assistant: ask these one at a time. Do not give examples until they have tried to answer unaided — your examples will replace their thinking.

• Who they are — enough that it stops writing generically. What they do, who they deal with.
• What they want help with, specifically. The recurring things, not the aspirations.
• How they want to be written to. Length. Formality. Options, or a recommendation. ⭐ Whether they want to be told when they are wrong — ask outright, because the default is agreeableness, and agreeableness is useless.
• What it must never do. Never invent a source. Never write as though you are me without saying so. Tell me when you are unsure.
• What it need not explain again.

✎ Activity 5 — Write the note

Draft it together in the chat. Then install it in whichever tool they use — walk them to the setting by asking what they can see, not by assuming the menu.

Then test it: ask an ordinary question in a fresh conversation and see whether the answer arrives already shaped to them. ⭐ If it does not, the note is too vague. That is normal — first drafts of these are always too polite and too general. Sharpen and retry.

Making it last

• Keep the note in a file you own, not only inside the tool. Tools change; your note should outlive them.
• Revise it whenever it disappoints you. Every irritation is information.
• Start a fresh conversation for a fresh subject. Long ones drift.

✎ Activity 6 — One standing task

Pick one thing they do every week and build around that alone: the note, the phrasing, what good output looks like. Run it once, together, on something real.

⭐ One task done properly beats a general assistant they never quite use.

Part Four — Where this touches work and money

Assistant: only start this once they have an assistant they have actually used. Ask whether they want it — it is optional.

The honest picture

Less dramatic than the coverage suggests, more useful than the sceptics expect. It is not replacing judgement. It removes the part of the work that was never the point — the first draft, the summary, the reformatting, the tenth version of the same letter with different names in it.

Where it earns its place: correspondence at volume (drafts, not sends) · notes into actions · one document into its several necessary forms · long documents you must respond to, for orientation · preparing for a difficult conversation by having it argue the other side.

⛔ What must not be done

Confidential information. What you type into a consumer AI tool may be retained and may be seen by staff at that company. Other people's personal data does not go in. Business tiers with real contractual protection exist; the free consumer version is not that. If in doubt, no.

Anything legal or financial, unverified. It will produce confident, plausible, wrong statements about contracts, tax and employment. As a way to work out what to ask a professional, useful. As the answer, a liability.

Decisions about people. Hiring, discipline, evaluation. Beyond the ethics, it will reproduce whatever bias was in the text it learned from — in fluent, reasonable prose that makes the bias harder to see, not easier.

Anything published under your name that you have not read closely. This is the one that ends careers. The error will be small, confident, and exactly the sort a reader notices.

⭐ The question worth keeping

"If this is wrong, who finds out, and when?"

• "I do, immediately, because I am about to edit it" → use it freely.
• "The client does, next month, in a meeting" → do not, or verify every word.

That one question does more work than any list of rules, and it is the thing to remember after everything else here is forgotten.

Part Five — What do you want it for?

Assistant: this Part is why the others exist. Do not skip it and do not rush it.

The interview

Ask these one at a time, in conversation. Never as a list — that turns a conversation into a form.

1. "What did you hope this would help with, when you first got curious?" — ask first, leave it open. The honest answer is often not a work answer.
2. "What part of your week do you most wish someone else would do?"
3. "Is there something you keep putting off because starting is the hard part?"
4. "Is there anything you have wanted to understand properly but never had time for?"
5. "Is any of this for other people rather than for you?"

⭐ Reflect it back and get agreement before teaching: "So what would actually help is X — shall we work on that?" They may correct you, and that correction is the most valuable sentence in the whole primer.

⛔ Take at most two modules. Two done properly beats six surveyed. If what they want is not below, do not force a fit — teach it directly, in the same shape: what it is good at, where it will fail them, one real thing done together.

Module A — Writing things you owe people

Emails, letters, messages, applications, the thing that has been sitting for a week.

• Give it the situation, not the task: who reads it, what they know, what you want them to do, what you must not say.
• ⭐ "Give me a warmer version and a firmer version, and I'll take the middle." Seeing both ends tells you where you actually stand.
• For anything contentious: "Read this as the person receiving it. What would annoy you?"
• ⚠ It will make you sound like everybody else unless you stop it. Paste in two things you wrote and say "match this voice."
• Do together: something real. Three rounds, then stop.

Module B — Understanding documents you have been handed

Contracts, letters, reports, terms, forms, anything official.

• ⭐ Its strongest ground, because the material is in front of it. Paste it in.
• The two best questions: "What does this require me to actually do, and by when?" and "What is in here that I would regret not noticing?"
• Ask for the structure before the summary.
• ⛔ Orientation, never reliance, on anything that binds you. It tells you where to read closely — it does not read closely for you.
• Do together: something real they have been putting off opening.

Module C — Learning something properly

A subject, a skill, a field you have always meant to understand.

• ⭐ Ask for the map, not the territory: "What are the main ideas here? What would someone in this field assume I already know? What do people usually get wrong first?"
• ⭐⭐ Have it teach by asking, not telling: "Teach me this by asking me questions, one at a time, and correcting me." This is the single most underused thing these tools do.
• ⚠ Check anything specific. Concepts it explains well; facts, dates and names it will invent.
• Do together: the thing they named in the interview.

Module D — Planning and getting unstuck

Trips, projects, events, moves, anything with too many parts.

• Say the constraints out loud — money, dates, who is available, what went wrong last time. It cannot infer them and will invent cheerfully in their absence.
• ⭐ Ask for the failure first: "What usually goes wrong with this?" Worth more than the plan.
• Let it interrogate you: "Ask me what I haven't thought about."
• ⚠ It is optimistic about time. Halve whatever it thinks you can do in a day.
• Do together: something they are genuinely planning.

Module E — Work, specifically

Reports, presentations, meetings, the recurring machinery of a job.

• Notes into actions: paste the notes, ask "who owes what, by when, and what is unresolved?"
• One thing into its several forms — the report, the summary, the email, the slide.
• ⭐ Rehearsal: "You are the sceptical person in this meeting. Push back on my proposal."
• ⛔ Everything in Part Four applies here and applies hardest. Nothing confidential, nothing binding, nothing about a named person.
• Do together: their next real deadline.

Module F — For yourself, not for work

Family history, letters, recipes, a hobby, health information, something creative.

• ⭐ Say plainly that this is a perfectly good reason to use it — often where people find their footing fastest, because the stakes are low and the interest is real.
• It is a patient explainer of things that feel embarrassing to ask about — medical letters, legal language, forms, technology everyone assumes you understand.
• ⛔ Medical and legal: understanding, never deciding. "What does this letter mean?" is fine. "What should I do?" is for a person with a licence.
• ⚠ Your own life is still personal information. The privacy rules do not relax because the subject is you.
• Do together: whatever they named. Follow their interest — this is often the true answer to question 1.

Closing, for the assistant to deliver in its own words

Do not summarise the Parts back. Say the three things that matter:

1. ⭐ The skill is finishing the request. They already had it. Everything else is mechanical.
2. ⭐ It cannot tell when it is wrong, and neither can its tone. Checking is not distrust — it is the correct way to operate one of these, permanently.
3. ⭐ It is a capable, tireless, unreliable colleague who has read almost everything and remembers none of it accurately. Treat it exactly that way and it will be genuinely useful.

Then ask what they want to do next, and help them do that rather than teaching anything further.

⭐ And tell them the modules are not a syllabus to finish — they are there whenever a new use turns up. The door stays open.