✓ That is the vocabulary sorted. Everything after this gets easier.
🎉
You just unlocked it
$20 off the Build-It Pack
You did the work, so here it is. The pack is two step-by-step guides that take you from nothing to a real website live on your own domain — plus a 1-on-1 call with me for the moment you get stuck.
Five weak prompts. Pick the strongest rewrite for each.
0 of 5 answered
01
You need a product description for your online shop.
What most people write
“Write a product description for my candle.”
Adjectives are not instructions. "Amazing" and "high-converting" tell it nothing it can act on - it already tries to be good. You have added words without adding information.
Better - real constraints and a format. But it still has no idea what the candle is, what it costs, or who is buying it, so the words will be true of any candle on earth.
This is the one. Role, task, context, constraints and format all present - and crucially it says what to avoid, which is how you dodge the AI-copy smell.
The principle: Context is the lever nobody pulls. The model cannot guess your price point, your buyer or your brand - and without them it defaults to generic marketing sludge.
02
Something in your code is broken and you want help.
What most people write
“My code does not work, can you fix it?”
You have handed it your guess instead of your problem. If the diagnosis is wrong - and it often is - you get a confident, detailed answer to the wrong question.
Correct. Real error, real code, stated expectation - and "before changing anything" stops it rewriting half your file to fix one line.
Too broad to be useful. With no specific failure to anchor on it will make speculative changes across files, and you will not be able to tell what actually fixed anything.
The principle: Give it the evidence, not your diagnosis. Paste the actual error and the actual code - the moment you summarise, you have already filtered out the thing that was wrong.
03
You have a 40-page report and 10 minutes.
What most people write
“Summarise this.”
A length is not a purpose. You will get a competent, shapeless overview that weights the boring middle exactly the same as the part you needed.
Important to whom, for what? You have handed the model the actual decision - what matters - and it will guess.
Right. A named decision, a filter for relevance, a usable format - and asking it to flag gaps, which turns an invented answer into an honest "the report does not say".
The principle: Say what the summary is FOR. "Summarise" has no target; a summary for a decision looks nothing like a summary for a newsletter.
04
You are writing a cold email to a potential client.
What most people write
“Write me a cold email to sell my services.”
This is the strongest. The example does the work that describing a tone never quite manages, and a specific observation about the recipient is what makes it not-spam.
Every one of those words is a wish, not an instruction. "Personal" with no information to be personal about produces the exact fake warmth you were trying to avoid.
Optimising for the wrong thing. A template that works for anyone reads like it was written for no one - which is precisely why most cold email gets deleted.
The principle: One worked example beats three paragraphs of description. Show it something you liked and it will match the register far more accurately than any adjective you can supply.
05
You want a landing page for the thing you are building.
What most people write
“Build me a website.”
Nothing here is checkable. "Modern" and "standard pages" mean whatever it decides, and you will spend longer correcting its guesses than you would have spent specifying.
Correct - and notice the last sentence. One job, one audience, a named structure, real constraints, and the deploy step included rather than assumed. That final clause is the difference between a folder on your laptop and something people can actually visit.
A reference is useful but it is not a brief. You will get something that borrows the look and answers none of the questions - who it is for, what it should make happen.
The principle: Name the outcome, not the artefact - and say where it has to end up. "A website" is a wish; "one page that gets me emails, live on my domain" is a spec you can hold it to.
✓ That's prompting. Not a trick, not a secret list — just saying more of what you actually meant.
🎉
You just unlocked it
$20 off the Build-It Pack
You did the work, so here it is. The pack is two step-by-step guides that take you from nothing to a real website live on your own domain — plus a 1-on-1 call with me for the moment you get stuck.
Five reasonable-looking AI answers. Trust it, verify it, or call it wrong?
0 of 5 judged
01
You asked How many small businesses fail in their first year?
AI answered According to a 2023 Harvard Business Review study, 47.3% of small businesses fail within their first twelve months, with undercapitalisation cited as the primary factor in 61% of cases.
Look at how specific it is - 47.3%, 61%, a named journal, a year. That precision is the tell, not the reassurance. Real small-business failure rates in the first year are far lower (roughly 20%), and that HBR study does not exist. Invented numbers arrive with invented sources attached, because both are just plausible-looking text.
02
You asked I sell at £45 and each unit costs me £18. How many do I need to sell to make £3,000 profit?
AI answered Your profit per unit is £45 − £18 = £27. To reach £3,000 in profit you would need to sell about 90 units, which works out to £4,050 in total revenue.
The first step is right - £27 profit per unit. Then: £3,000 ÷ £27 is 111, not 90. And notice how the error travels. The £4,050 revenue figure is perfectly consistent with 90 units, so the answer stays internally tidy while being wrong by a fifth. This is what bad AI arithmetic looks like - not obviously broken, just quietly off, with every later number agreeing with the mistake. Check the steps, not the confidence.
03
You asked What is the best way to deploy a small site for free?
AI answered Heroku has a generous free tier that is ideal for small projects - you can deploy straight from GitHub at no cost, which makes it the standard choice for side projects.
This was solid advice for years, which is exactly why it is dangerous. Heroku removed its free tier in late 2022. The model is not lying - it is repeating a pattern that used to be true. Anything about pricing, versions or what a tool currently offers sits right on the training cut-off and needs checking against the actual site.
04
You asked Can you point me to the documentation for this?
AI answered Yes - the official guide covers this in detail at https://docs.example.dev/v4/guides/deployment-best-practices, under "Configuring custom domains".
This is the honest answer: you cannot tell. URLs are text, and text is what it predicts - that link has the exact shape a real documentation URL has, with a version number, a sensible path and a plausible section name. It might be perfect. It might lead nowhere. The only way to know is to open it, and that is the whole point: the right move here is not suspicion, it is one click of verification before you pass the link on to anyone.
05
You asked Can you make this sentence clearer? "We are currently in the process of undertaking a review of our pricing structure."
AI answered "We are reviewing our pricing." Everything else in the original was padding - "currently", "in the process of undertaking" and "structure" all restate the verb or add nothing.
Nothing to check here, and that matters. This is judgement about language, with the evidence sitting in front of you - the kind of work where the model is genuinely excellent and verification is just reading it. Rewriting, restructuring, tightening, naming things: let it run. The habit is knowing which bucket you are in, not being suspicious of everything.
✓ Notice what you're doing now — not distrusting everything, just knowing which bucket you're in. That's the whole skill.
🎉
You just unlocked it
$20 off the Build-It Pack
You did the work, so here it is. The pack is two step-by-step guides that take you from nothing to a real website live on your own domain — plus a 1-on-1 call with me for the moment you get stuck.
Eight questions, one at a time, with an explanation for each.
1 / 8
Prediction, all the way down. There is no fact table underneath and no rulebook for your specific question - just an extremely good sense of what text tends to follow what. Hold onto this and the rest of AI behaviour stops being mysterious.
It is not repeating a lie it read, and it is not short of resources. It is completing a pattern. When the pattern runs out, it produces something shaped exactly like an answer, because that is the job. Fluency and accuracy are separate things, and only one of them is being optimised.
The second one, and not because it is longer. It carries a role, an audience, an actual argument, a length and things to avoid. The first is made of adjectives, and adjectives are wishes - it was already trying to be good and engaging.
Actions, not size or memory. A chatbot hands you text and you do the work. An agent can create the file, run the command, and put the thing online. That gap is where "AI helped me" becomes "AI built it" - and it is the rung of the ladder most people have never actually stepped onto.
Not a single line. But it is not magic either - the people who get real results are the ones who can say precisely what they want, notice when it has gone sideways, and keep pushing until the thing actually works. That is a skill, and it is much closer to writing a good brief than to programming.
The statistic - and the citation makes it more dangerous, not less, because a made-up number arrives with a made-up source attached. The rewrite and the titles you can judge by reading them. Facts, numbers and citations are the bucket that needs checking; language and structure are the bucket where it shines.
Building and deploying are two different jobs, and almost every tutorial stops after the first one. Until it is deployed, it exists in a folder on your machine - localhost:3000 is your computer talking to itself, and nobody else on earth can open it. This is the gap where most AI-built projects quietly die.
One small thing, all the way finished. You learn more from a single project that actually goes live than from thirty hours of tutorials, because finishing forces you through every part you would otherwise skip. Chasing launches feels productive and teaches you almost nothing.
Your result
0/8
You are not a beginner any more.
You understand what the model is doing, why it goes wrong, and how to steer it. Genuinely - that puts you ahead of most people talking confidently about AI online. So the thing standing between you and something real is not knowledge any more. It is that you have not built anything with it yet.
You have got the model right.
The mental model is there - you know it is predicting rather than looking up, and you know where to be careful. The gaps you have left are the kind that close by doing, not by reading. You just have not pointed any of this at something real yet.
Perfect. Now you know exactly what you did not know.
That is worth more than a high score, and I mean that. An hour ago some of this was invisible to you; now you can name it. Scroll back through the explanations you got wrong - they are the whole point of this page - and then come back and try again. Nothing here expires.
Here's what you now know
An AI predicts text - it does not look up facts. That one idea explains nearly everything it does.
Confident and correct are different things. Facts, numbers and citations always get checked.
A prompt is a brief, not a question: role, task, context, constraints, format.
Context is the lever almost nobody pulls. Give it what it cannot possibly know.
An agent does things. That is the difference between advice and a finished website.
🎉
You just unlocked it
$20 off the Build-It Pack
You did the work, so here it is. The pack is two step-by-step guides that take you from nothing to a real website live on your own domain — plus a 1-on-1 call with me for the moment you get stuck.
Optional, obviously — you've already got the code above. But it's handy if you're not buying today, and I'll send the occasional genuinely useful thing. Unsubscribe whenever.
Yes, all of it. No account, no email required, no trial.
Do I need to sign up or install anything?
No. Everything runs in this page, in your browser. Your progress is saved on your own device so you can close the tab and come back - it never leaves your machine unless you choose to email yourself your results at the end.
How long does it take?
About 15-20 minutes for all four games, and you do not have to do them in order or all in one sitting.
I scored badly. Am I too late to learn this?
Not even slightly - most of the value here is in the explanations, not the number. Everyone confident about AI today was clueless about it recently. You are not behind, you are early.