Track 1 · The Map
You cannot ask for help you can't imagine.
The biggest blocker isn't bad prompting — it's not knowing what's on the menu. This page is the menu, plus five habits for finding the rest yourself.
The real bottleneck
Most people use a fraction of it, and never find out
Ask someone who has had an AI chatbot for a year what they use it for, and you'll often get two answers: writing emails, and looking things up. Both work. Both are a rounding error on what the tool does.
This isn't a failure of intelligence or effort. It's structural. A search engine teaches you its limits quickly — you type something, you get nothing useful, you learn. An AI chatbot will refuse some things outright, but it almost never says "I am out of my depth on this one." On an ordinary question it produces something either way. So you rarely bump into the edges, and you never discover the rooms you didn't walk into.
Your blind spot isn't a wrong belief you could argue with. It's an absent question. The fix isn't a better answer — it's exposure to the range, and a habit of going looking.
The menu
Ten things it is genuinely good at
A list of shapes of work, not features — most useful requests are one of these ten with different names on them. Open any you haven't tried.
01Turning one form into another
Messy notes into a clean summary. A transcript into minutes with action items. A table into a paragraph. A paragraph into a table. An email chain into a timeline. A spec into a checklist.
This is the single most reliable category, because the answer is already contained in what you gave it. Nothing has to be invented, so there's much less to get wrong.
02Explaining anything at any level
The same idea explained to a ten-year-old, to a smart adult with no background, and to a specialist. You can say "I got lost at the third sentence — start again from there" as many times as you need, and it never gets impatient.
Underrated: paste in the thing you don't understand — a legal clause, an error message, a letter you can't parse — and ask what it means in ordinary words. Take the names and numbers out first, and treat the answer as a way to understand the document, not as a ruling on it. See what not to paste.
03Beating the blank page
A first draft of anything: an awkward email, a complaint letter, a job description, a speech, a policy, a plan. Not because the draft is good — usually it's mediocre — but because most people find reacting to a mediocre draft far easier than starting from nothing.
Mediocre is the risk. If you ship the first draft, you ship generic. The value is in what you change.
04Criticising work — including yours
"What's the weakest part of this argument?" "What would a hostile reviewer attack first?" "What am I assuming here without saying so?" "What's missing that an expert would expect to see?"
This is where a lot of the real value hides, and almost nobody uses it. It's more useful than asking it to write for you, because a critique you disagree with is still information.
05Pulling structure out of mess
Fifty customer emails into five themes with counts. A pile of receipts into categories. A long thread into "who is asking for what, and by when." An unstructured page of notes into a table with consistent columns.
Ask for the output as a table or a list with fixed fields — structure in the request produces structure in the answer, and structured answers are far easier to check.
06Being a tutor with infinite patience
Not "explain X" but "teach me X. Ask me a question, wait for my answer, tell me what I got wrong, then go one step further." A tutor that never sighs, never runs out of time, and doesn't mind that you asked the same thing twice.
For anyone who left school long ago, or never had access to good teaching, this may be the most valuable thing on this page.
07Rehearsal for something real
"You are the hiring manager. Interview me for this job, one question at a time, and be tough." Or the difficult conversation with a colleague. Or the negotiation. Or explaining your diagnosis back to your doctor so you know you understood it.
Nobody sees you get it wrong. That's the whole point.
08Translation — between languages, and between worlds
Between human languages, yes. But also: engineer to executive, legal to plain English, doctor to patient, angry to professional. "Rewrite this so it says the same thing but doesn't sound furious" is a genuinely valuable service.
09Writing and fixing code — even if you can't
A spreadsheet formula. A script that renames 400 files. A small web page. A fix for an error message you don't understand. You do not need to be a programmer to get real use out of this — you need to be able to test the result.
Never run code you can't test the effect of, and never run something that deletes or overwrites without a copy of the original. Code that runs is not the same as code that is correct.
10Thinking out loud with something that answers
Half-formed idea, no audience, 11pm. "Here's what I'm thinking, poke holes." "Give me three ways to look at this that I haven't considered." "What would someone who disagrees with me say?"
Not because it's wise. Because articulating a problem to something that responds is how a lot of people find out what they actually think.
The other half of the map
Eight things it is bad at on its own
Most of these follow directly from how the thing works: it produces likely-looking language, one piece at a time. Tools fix some of them — but only when the tool actually fires. Check that the answer shows a link or shows its working; if it shows neither, assume no tool ran.
| It is bad at | Because | What to do instead |
|---|---|---|
| Knowing today's facts | Its own knowledge stops at a training cutoff date. Beyond that it will refuse or invent. | Use a tool that actually searches the web, and open the sources. No links came back? Then no search happened. |
| Knowing what it doesn't know | It cannot tell its own remembering from its own inventing. Both come out sounding the same. | Assume every specific claim is a claim, not a fact, until you've checked it. |
| Arithmetic and counting | Left to itself it predicts the look of a number, so long sums and counting drift. Running real code fixes this. | Do the maths yourself, or use a tool that genuinely calculates — and check it actually showed you the calculation. |
| Anything about you | Unless the product was built to remember, each conversation starts from nothing. It doesn't know your company, your files, or last week. | Give it the context every time — paste the actual document rather than describing it. |
| Saying "no" | It is built to be helpful. A confident wrong answer is a more likely piece of text than a refusal. | Explicitly permit it: "If you don't know, say so." It helps. It is not a guarantee. |
| Holding a position under pressure | Push back and it will often fold and agree with you, whether or not you were right. | Never use "are you sure?" as your fact-check. Check the fact somewhere else. |
| Genuine novelty | It reflects patterns in what already exists. Ask for something original and you often get a competent average. | Use it to widen your options, then bring your own taste to choose and sharpen. |
| Consequences | It has no stake in the outcome. It will not be there when the plan fails, the money is gone, or the letter is sent. | Keep the decision. Delegate the drafting, never the accountability. |
The part that surprises everyone
It is not uniformly clever. The edge is jagged.
People expect ability to be a single dial: a bit smart, quite smart, very smart. That intuition is wrong here, and it's the reason so many first impressions are either "this is magic" or "this is useless" — both formed from one lucky or unlucky question.
In practice the boundary between what it does brilliantly and what it fails at is ragged and unintuitive. It can produce a genuinely expert paragraph about immunology and then get a basic detail about your own industry confidently wrong. Neither result tells you much about the next task. Researchers studying consultants using AI at work named this shape the "jagged technological frontier" — a useful term, and not ours. It comes from a 2023 Harvard Business School working paper by Dell'Acqua and colleagues, reporting a field experiment with several hundred Boston Consulting Group consultants.
"It got that wrong, so it's useless." Or: "It nailed that, so I can trust it on the next thing." Both are the same mistake — generalising from one point.
Ability is task-shaped, not level-shaped. The only way to know where the edge sits for your work is to test it on your work — and to keep testing, because the edge moves with every model release.
The antidote
Five habits that find your blind spots for you
You can't list what you don't know. But you can run a process that keeps surfacing it. These five cost almost nothing and compound quickly.
-
Keep a friction log for one week
Every time something at work or at home is tedious, repetitive, or makes you sigh — write one line in a note. Don't judge, just log. At the end of the week you're holding a list of the AI-shaped holes in your life. This finds far more than any "100 prompts" article, because it's about your actual days.
-
Ask it what you should have asked
The cheapest trick on this whole site. At the end of any exchange:
What should I have asked you that I didn't? What do people in my situation usually miss here?You are literally asking the map to show you the parts you walked past. Then verify what it says — it's a lead, not a fact.
-
Push one task until it breaks
Take something it did well and keep escalating: longer, harder, more constraints, more data. The point isn't to win — it's to feel where the edge is, so you stop guessing. Ten minutes of deliberate breaking teaches more than an hour of tutorials.
-
Watch a competent person work, once
Not a polished tutorial — someone doing their real job with it, screen shared, including the bits where it goes wrong. Most of what you take away will be things you didn't know were possible, rather than techniques. Ask a colleague. Most people are flattered to be asked.
-
Run the same task in a second model
Once a quarter, take a real task and give it to a different AI. Two things happen: you find out how much of what you assumed was "how AI is" was just how that one is, and disagreement between them flags exactly the claims worth checking. Comparing is not the same as switching — you can do this without moving. It doubles as verification: see Trust.
Self-audit
Sixteen ordinary things people don't know it can do
Tick anything you'd never thought to try. Whatever you leave unticked is your map of what you didn't know you didn't know. Saved on this device only.
Items 1, 4 and 15 have a hard edge. They are about understanding a document and preparing better questions — never about getting a diagnosis, a legal opinion, a tax position, or a deadline you will rely on. Confirm anything with a consequence attached with the sender or a qualified person. The AI helps you walk in informed; it is not the professional, and it carries none of the liability if it's wrong.
Take the personal details out before you paste. Account numbers, names, dates of birth, reference numbers — none of them are needed for the AI to explain the jargon. The rules are on the Trust page.