Track 3 · Get Help

Treat it as a fast, well-read, over-confident new colleague.

Not an oracle, not a search box, not a friend: someone brilliant who started this morning, knows nothing about your job, and never admits to being lost. Get that picture right and the technique follows.

The mental model

Why that picture produces better results than any prompt trick

These are habits you already have with a new colleague — and most of what people call "prompt engineering".

Because it's…	You would naturally…
New here	Explain the situation before the task. Say who it's for and why it matters. Never assume they know your context.
Well-read but brand new	Expect solid general knowledge and zero knowledge of your specifics. Hand over the actual documents.
Fast and tireless	Ask for three versions instead of one. Iterate rather than agonising over the perfect first request.
Over-confident	Check their work before it goes anywhere with your name on it. See Trust.
Not accountable	Keep the decision yourself. Delegate the drafting, never the judgement.

The test

Before sending a request, ask: could a capable stranger do this task from what I've written? If the answer is no, the AI can't either — and it will guess rather than tell you it's stuck. Nearly every disappointing answer traces back to a failed version of this test.

The craft

The five parts of a request that works

When an answer disappoints, one of these five is almost always the missing piece.

The first two are where most requests fail — a capable stranger can't work without them.

Context — who, for whom, why

Who you are, who this is for, what's already happened, what constraints exist. This is the part beginners skip and experts over-deliver on.
Task — the actual verb

Not "help me with my email" but "rewrite this so it declines without closing the door." Vague verbs produce vague output.
Constraints — length, tone, limits

Length, tone, what to avoid, what must be included, what you've already tried and don't want repeated.
Format — the shape you want back

A table with these columns. Five bullets. An email under 150 words. Plain English, no jargon. Shape is free to ask for and enormously improves usability — and makes checking easier.
Example — one sample of "good"

The most powerful and least used ingredient. One example of what right looks like beats three paragraphs describing it.

The same request, before and after

Before

Write an email to my team about the new process.

Result: something generic, corporate, and unusable — because every choice was left to a stranger who knows nothing.

After

I manage a team of 9 in customer support. From Monday we're moving from email to a shared ticket queue. The team is tired of tool changes — the last two were dropped after a month, so trust is low. Write an email to them announcing it. Constraints: under 200 words. Acknowledge the last two failures directly instead of pretending they didn't happen. No corporate cheer. Say plainly what's in it for them, and give one concrete thing they can do on Monday. End with an open offer to hear objections.

Nothing clever happened there. The second version simply told a new colleague what they'd need to know. That's the whole skill.

The habit that matters most

Round one is not the answer. It's the opening move.

The single biggest difference between people who find AI useful and people who find it disappointing isn't prompt quality. It's that one group stops after the first reply and the other doesn't.

Your first message is a guess about what you want, answered by a guess about what you meant. The value is in rounds two to five, where you say what was wrong with round one. That is not a failure of the tool — it's how the tool is meant to be used.

Round 1

You see the shape of the answer, and discover what you actually wanted — which you often didn't know until you saw the wrong version.

Round 2–3

"Too formal." "You missed the cost angle." "Cut it in half." "Keep paragraph 2, redo the rest." Direct, blunt, specific.

Round 4–5

Now the material is genuinely yours. This is where the output stops being generic and starts being usable.

Say what's wrong, not "try again"

"Do it better" gets you a random reshuffle. "The second paragraph is too vague — give me the actual numbers, and drop the last line entirely" gets you the next draft. Be as direct as you'd never dare be with a person. It genuinely does not mind.

The biggest single upgrade

Stop describing the material. Give it the material.

Most people type a description of the thing they're working on. Then they're disappointed that the answer is generic — but a description is all the AI had.

Before you paste anything, one question

Is this yours to share? Work documents, customer data, colleagues' details and anything under a confidentiality agreement are covered by the never-paste list — and a sales spreadsheet or a client report is almost always your employer's, not yours. Use the tool your employer approved, or don't use one.

Describing

"I have a long report about my local council's budget and I want to know what's going on."

Giving

Paste or upload the actual document. Then: "What stands out? What would you check before believing it? What's missing that I'd want for a decision?"

Same for the transcript, the error message, the rejection letter, the job description, the manual nobody can find. The more real the document, the bigger the gain — and the more the privacy question above applies.

Copy these

Ten patterns that work in almost any situation

Each is a role you can give it. Copy one, paste it, replace the bracketed part.

01The tutor

Teach me [topic]. I know nothing about it. Don't lecture — go one small step at a time, and after each step ask me a question to check I followed. If I get it wrong, explain what I misunderstood before moving on.

02The critic

Here is [my work]. Don't tell me what's good. Tell me the three weakest things about it, why each is weak, and what would fix it. Be blunt — I'd rather hear it from you than from my audience.

03The translator

Rewrite this for [audience], who [what they care about and what they don't]. Same meaning, nothing added or softened. Then tell me which parts I'm likely to be challenged on.

04The explainer

Explain [thing] to me three times: once as if I'm ten, once as if I'm a smart adult with no background, and once as if I work in the field. Then tell me which one I probably need.

05The devil's advocate

I'm about to [decision]. Argue against it as forcefully as you can. Then tell me which of your arguments is strongest, and what evidence would change my mind.

06The rehearsal partner

You are [the other person: interviewer / my landlord / a sceptical client]. I'm going to [what I want]. Play them realistically, including being difficult. One exchange at a time — wait for my reply before continuing.

07The extractor

Here is [messy text]. Turn it into a table with these exact columns: [columns]. If a value isn't stated, write "not stated" — do not guess or fill it in. List anything that didn't fit the table underneath.

The "do not guess" instruction is the important half. Without it, gaps get quietly filled with plausible inventions.

08The pre-mortem

A technique from decision researcher Gary Klein (Harvard Business Review, 2007): imagine the failure has already happened, then explain how. The idea is his; the wording below is ours.

It's six months from now and [my plan] has failed badly. Write the story of how it failed. Then rank the causes by how likely they are, and tell me which one I could cheaply protect against this week.

09The question-finder

I'm about to meet [a doctor / lawyer / builder / recruiter] about [situation]. What should I ask them that someone in my position usually doesn't think to ask? What would I regret not asking?

10The unsticker

I'm stuck on [problem]. I've already tried [what you tried]. Ask me questions until you understand the situation properly — don't propose anything until you do.

Making it interview you is the most underused move there is. It surfaces the thing you left out because it seemed obvious.

Save yourself the time

Seven beginner mistakes, in order of how much they cost

Mistake	Fix
Stopping after one reply	Treat round one as a draft. The value is in the corrections.
Describing instead of pasting	Give it the real document. Generic input, generic output.
No context about you	Say who you are, who it's for, and what you're going to do with it.
Asking for facts, not thinking	It's much stronger at structuring, criticising and explaining than at recalling. Use search for lookups.
Shipping the first draft	Its default voice is recognisable and bland. Rewrite in your own words — that's also how you catch its errors.
Being polite instead of clear	"Could you perhaps make it a bit shorter?" → "Cut to 120 words. Drop the third paragraph."
One giant request	Break it into steps and check each. A wrong step four is invisible inside a wall of output.

When to move up a gear

Three levels of tool. Most people never leave the first.

Level 1

Chat

Type a question, get an answer. Fine for most things, and where everyone starts.

Move on when: you're pasting the same background into every new conversation.

Level 2

Chat with your files

Upload documents, keep a persistent set of instructions, work over your own material rather than general knowledge.

Move on when: you want it to do things, not just produce text about them.

Level 3

Agents and coding tools

It takes multiple steps on its own — searching, running code, editing files, checking its own work.

The catch: more autonomy, more to verify, and far more it can break in one go when it's wrong.

Rule for level 3

Never let it act unsupervised on something you couldn't undo. Work on copies, keep backups, and read what it did before you accept it. Autonomy is a multiplier — it multiplies mistakes just as efficiently as it multiplies work.