Bad brief, bad result: why AI doesn't read minds

There's an old engineering rule: the result is only as good as the brief. With AI it bites harder than with people — a person would at least ask a follow-up question, while a model silently hands you a confident answer to the task it invented on your behalf.

You already know the RTCF formula — role, task, context, format — from the lesson "Your first prompt". This lesson answers the question that one left off-screen: why it works, what exactly breaks in the answer when one letter drops out, and which fifth element beginners skip most often.

AI doesn't read minds — and that's not a figure of speech

Remember the lesson on LLMs and tokens: a model predicts the likely continuation of text. Everything it knows about your task is the characters you sent. It has no access to your company, your deadline, your client's temper, or the way your team writes emails.

Meanwhile the task in your head is three-dimensional. You know the client is on edge, that this is the second delay, that the apology should be warm but promise nothing, and that your manager will read it too. Then you sit down and type: "write an email to a client". Out of all that volume, five words made the trip.

What happens next is worth understanding once and for good: the gap doesn't stay empty. A model can't decline to decide — the prediction mechanism is obliged to produce the next word, always. So everything you left unsaid gets filled in by someone. The only question is whether that someone is you or the statistics.

A pause to think

Before you read on, answer this for yourself: if the model is obliged to fill every gap, what will it fill it with? Hold your answer for one paragraph.

The most common thing. The average thing. Whatever showed up most often in similar texts across the internet. That's why answers to empty prompts all taste the same: smooth, faceless, politely about nothing. It isn't laziness or weakness in the model — it's honest statistics doing exactly what they're built to do.

Which gives us the line worth copying down somewhere: a bad brief doesn't produce a bad answer. It produces a good answer to a different task — the average one. That's precisely why people say "AI just writes fluff": they got a high-quality average when they wanted their own specific thing.

The five parts of a task

You know four of them; the fifth is new:

1. Role — who the model should "be". Picks the shelf of knowledge and the tone.
2. Task — what exactly to do. One verb: draft, find, shorten, compare.
3. Context — what it needs to know about the situation: who it's for, what already happened, what matters.
4. Format — what the answer should look like: a table, 5 bullets, under 100 words.
5. Constraints — what it must not do: don't invent numbers, don't promise dates, no jargon, no longer than one screen.

Constraints are the fifth element, and the most underrated one. The first four tell the model where to go. Constraints build the fence it must not cross. Without a fence, models routinely do the thing that comes back to bite you: promise a client a refund you never authorised, drop in a plausible number that doesn't exist (hello, the hallucinations lesson), write three screens where you needed a paragraph.

What breaks when a part is missing

• No role → no tone. Text "from the internet in general": stiff officialese one moment, breezy blogger the next.
• No task → the model recaps the topic instead of doing something with it. The single most common source of the "it's all fluff" feeling.
• No context → the model invents facts. The most dangerous hole: the answer looks confident while half the details were fabricated for plausibility.
• No format → a wall of text you then mine by hand for the useful bit. You didn't hand off the work, you postponed it.
• No constraints → extra promises, extra length, extra invention. The most expensive to fix, because you don't notice it right away.

See for yourself: two buttons

Two prompts follow. Hit Run under the first one and read the answer all the way through — the point is to feel it, not just to agree with me.

Write an email to a client about a delay.

Got it? Most likely it's polite emptiness: "dear customer", "we apologise for the inconvenience", "we value your trust". Not one detail of your actual situation — because you gave none. And notice: the model almost certainly decided for itself how long the delay is and what you're offering in return. It filled the gap, exactly as advertised.

Now the same intent, with all five parts:

You are a customer support manager at a small furniture workshop.
Task: write an email telling a client the delivery of their sofa moves from 14 to 21 March.
Context: this is the second delay in a row, the client already sent an angry message, the cause is a fabric delay at our supplier. This one is genuinely our fault.
Format: under 120 words, a normal email with no bullet lists, warm human tone.
Constraints: do not promise compensation or any date beyond 21 March, do not invent causes I haven't named, no corporate officialese, and don't open with "Dear valued customer".

Compare the two answers. The machinery under the hood is identical — same model, same day, same button. The whole difference is your text. That is the skill people get paid for.

The intern heuristic — the rule to take with you

How do you spot a hole in your brief before you get the bad answer? There's a ten-second test:

Imagine you sent this text to a new intern, they silently did all of it, and left for holiday. What did they get wrong?

The intern doesn't know the client is already annoyed — that's context, write it in. The intern might promise a discount — that's a constraint, write it in. The intern will write three pages — that's format, write it in. Everything you just mentally told the intern is the missing part of your brief. It works with no AI at all, on paper.

A common misconception: "longer prompt = better prompt"

No. Length is a side effect, not a goal. A prompt is good when it has no gaps the model would fill differently from what you want — that's the whole test. "What is 15% of 2400" needs no role and no constraints: there's nothing to fill in. An email to a furious client with no context and no constraints, though, is a landmine.

The rule is simple: the more an error costs, and the more of the task is personal to you and unguessable from the internet, the fuller your brief has to be.

Do this now

Take any real task from this week and write its brief across the five parts. Run the intern heuristic on it: what would they get wrong? Write those answers in. Next up in the course we take five such briefs apart piece by piece — and then learn to check what the AI hands back.