The School of Prompt

A prompt is a brief. You already know how to write a good one.

Ten lessons on talking to AI so that what comes back is worth your time. No tricks and no magic words: the same habits that make a good brief, a good debrief and good feedback. They work with any model and any tool.

10 lessons · about 25 minutes · free, no sign-up

Before lesson one

A model only knows what's in the conversation.: It hasn't read your brief, met your client or seen last year's campaign. Everything it doesn't get from you, it fills in with the average of everything it has read.
Clear beats clever.: There are no secret phrases. What works is what works with a good freelancer on day one: the goal, the material, the standard, and room to ask.
You stay the one who decides.: AI is good at the first layer: reading, sorting, options. The judgement, the taste and the sign-off are still the job.

01Lesson

Start with the outcome, not the task

"Write three headlines" is a task. "Three headlines for a launch post that makes 18–26s curious enough to open the app" is an outcome. The model can only aim at what you name.

What to do

Say what the result is for: a deck, a client email, a post, a decision.
Say who will read it and what they should think, feel or do after.
Say what good looks like: two or three criteria you'd judge it by.
Say what it is not: the obvious version you've already seen.

Why it works

Without a goal the model optimises for something that sounds plausible. With one, it can trade things off the way you would: shorter or warmer, safer or bolder.

Weak

Write three headlines for Nordic Oat's new barista oat drink.

Better

Write three headlines for Nordic Oat's new barista oat drink. They open a LinkedIn post aimed at café owners in Scandinavia. Goal: make them want to ask for a sample. Good means: specific to cafés, no health claims, under 9 words. Not wanted: anything about 'creamy' or 'plant-based goodness', which every competitor says.

Copy and adapt

I need [output] for [who reads it].
After reading it they should [think / feel / do].
Good means: [criterion 1], [criterion 2], [criterion 3].
Not wanted: [the obvious version].

Watch out: If you can't fill in "good means", the brief isn't finished yet, for the model or for a person.

↑ Back to the lessons

02Lesson

Context is the craft

Most weak answers aren't the model's fault. They come from a thin request. Context is the brief, the brand, the audience and the history, given in full rather than from memory.

What to do

Paste the source, not your summary: the brief, the client's email, the brand guidelines.
Add the brand's voice and the words it never uses.
Describe the audience in the client's own words, plus anything you know that they don't.
Add the history: what was tried, what was rejected and why.
Put the material first and the request last, with clear labels for each part.

Why it works

The model fills every gap with the average. Each piece of context replaces a guess with a fact, and the answer moves from generic to specific to your client.

Weak

We're doing a campaign for an oat drink brand, they want something fresh for young people. Ideas?

Better

BRIEF (full text from the client): [pasted brief] BRAND: calm, dry humour, never preachy. Never says 'guilt-free', 'superfood' or 'dairy is bad'. AUDIENCE: 20–30, city, already buy oat milk for coffee but not for anything else (client's research, 2025). HISTORY: last spring's 'Oat of Office' idea was rejected as too jokey for retail partners. TASK: three campaign routes that get this audience to use oat drink beyond coffee.

Copy and adapt

BRIEF:
[paste the full brief]

BRAND (voice, never says):
[...]

AUDIENCE (in the client's words):
[...]

HISTORY (tried, rejected, why):
[...]

CONSTRAINTS (legal, channel, length, budget):
[...]

TASK:
[what you want, and what good looks like]

Watch out: More isn't automatically better. Twenty pages of unrelated decks bury the three lines that matter. Give everything relevant, label it, and cut what isn't.

↑ Back to the lessons

03Lesson

Give it a role, a real one

"You are an expert marketer" does almost nothing. A role helps when it brings a point of view: what this person cares about, what they'd push back on, who they answer to.

What to do

Name the discipline and the seniority: a senior strategist, not "an expert".
Say what they care about and what they're sceptical of.
Say who they work for and what's at stake for them.
Use opposite roles to test an idea: the client's CFO, a tired café owner, the legal team.

Why it works

A role changes which details the model pays attention to. A planner looks for the problem behind the request; a copywriter listens for rhythm; a client service director thinks about how the client will read it.

Weak

You are a world-class creative expert. Give me ideas.

Better

Act as a senior strategist at an independent agency. You're sceptical of briefs that ask for 'viral' without saying why, and you always look for the business problem behind the request. You'll present to the client's marketing director, who cares about the retail listing in Q2 more than awards.

Copy and adapt

Act as a [seniority] [discipline] at [type of company].
You care about [...] and you're sceptical of [...].
You'll present this to [who], who mostly cares about [...].

Watch out: A simulated customer is a way to find questions, not research. Never put "consumers say" in a deck because a model played one.

↑ Back to the lessons

04Lesson

Teach it your style

Adjectives like "punchy", "premium" or "human" mean something different to everyone. Examples don't. The fastest way to get your voice back is to show it.

What to do

Collect three to five pieces you're proud of: emails, copy, a deck page.
Ask the model to describe the style as rules, not to copy the content.
Read the rules, correct them and keep the result as your style card.
Next time, paste the style card plus one example at the start.

Why it works

Models are very good at imitating patterns they can see. Rules extracted from your own work are more precise than anything you'd write from scratch, and the card makes the style reusable.

Weak

Write it in my style. I like it short and witty.

Better

Here are four client emails I wrote and liked (below). Describe my writing style as 8–10 rules: sentence length, how I open and close, vocabulary I use and avoid, how I handle bad news, how formal I am. Don't summarise the content. Then rewrite the draft at the bottom using those rules.

Copy and adapt

Below are [3–5] pieces I wrote: [type].
Describe my style as 8–10 concrete rules:
sentence length, openings and endings, words I use and avoid,
tone with clients, what I never do.
Don't copy the content. Then show me the rules before writing anything.

Watch out: The same goes for a brand voice. Build one style card per client from their best published work, and never mix them in one conversation.

↑ Back to the lessons

05Lesson

Say what to avoid, and why

Constraints are where good work comes from: the length, the legal lines, the words the brand never uses. A reason next to each rule helps the model apply it in cases you didn't list.

What to do

List hard limits: length, format, mandatory lines, claims that need legal sign-off.
List the clichés of the category so they're off the table from the start.
Give the reason: "no exclamation marks, the brand is calm" works better than "no exclamation marks".
Say what to do instead, not only what not to do.

Why it works

A bare rule gets followed literally. A rule with a reason gets understood, so the model also avoids the things you forgot to list.

Weak

Don't make it cheesy.

Better

Avoid the category's clichés: 'creamy', 'plant-powered', 'guilt-free', sunrise shots, milk splashes. The brand's tone is calm and a bit dry, so no exclamation marks and no puns on 'oat'. Health claims need legal approval, so none in this round. Instead, lean on small, specific moments from real kitchens.

Copy and adapt

Hard limits: [length, format, mandatory lines].
Avoid: [category clichés], because [reason].
No [claims / words], because [reason].
Instead: [what you want more of].

Watch out: A long list of "don'ts" with no direction gives you safe, flat work. Always end with what you want instead.

↑ Back to the lessons

06Lesson

Let it ask before it answers

A model always answers, even when the brief has holes. Asking it to question you first turns those holes into a short list you can fill in or take to the client.

What to do

Start with: "Before you start, ask me the questions that would change the result most."
Limit the number, five at most, ordered by impact.
Answer what you can; mark the rest as unknown.
Ask it to list its assumptions at the top of the answer.

Why it works

Guesses that read like facts are the most expensive kind of mistake: the work gets built on them. Questions bring the guesses into the open while they're still cheap to fix.

Weak

Here's the brief. Give me the strategy.

Better

Here's the brief. Don't write anything yet. Ask me up to five questions that would most change the strategy, ordered by impact. For each, say what you'd assume if I don't know the answer.

Copy and adapt

Don't start yet.
Ask me up to [5] questions that would most change the result,
ordered by impact.
For each one, say what you'd assume if I can't answer.

Watch out: The questions are useful even if you never use the answer. Often the best three go straight into the debrief with the client.

↑ Back to the lessons

07Lesson

Ask for range, then choose

Ten variations of one idea is one idea. Ask for options that differ on purpose, say along which line, and make each one name the bet it's making.

What to do

Say how the options should differ: insight, tone, mechanism, channel.
Ask each option to state its bet in one line: "this works if...".
Ask for the obvious ideas first, then cross them out and ask for routes that avoid them.
Pick one, then develop it in a separate step.

Why it works

A model's first idea is the most common one, because it learned from what has already been done. Forcing range moves it away from the average; naming the bet makes the options easy to judge.

Weak

Give me 10 ideas for the campaign.

Better

First list the five most obvious ideas any agency would bring for this brief. Then give me three routes that avoid all five and differ in the insight they start from. For each route: the insight in one sentence, the idea in one sentence, and the bet it makes ("this works if...").

Copy and adapt

List the [5] most obvious ideas for this brief.
Then give me [3] routes that avoid all of them
and differ in [insight / tone / mechanism].
For each: insight, idea, and the bet it makes ("this works if...").

Watch out: Don't ask for 30 options to feel safe. Three that really differ are easier to judge and better to present.

↑ Back to the lessons

08Lesson

One step at a time

"Read this brief and write the campaign" asks for five jobs at once, and you can't check any of them. Split the work the way an agency would, and look at each layer before the next one goes on.

What to do

Split it: understand the brief, name the problem, routes, then execution.
Check and correct each step before moving on.
Carry the approved output of one step into the next as its input.
Start a new conversation when the task changes or the chat gets long and muddled.

Why it works

Each step is small enough to judge. A mistake in step one gets fixed there, instead of quietly spreading through the copy, the visuals and the deck.

Weak

Here's the brief, write the full campaign with copy and visuals.

Better

Step 1 of 4: summarise the brief in five lines and separate what the client asked for from the business problem behind it. Stop there. Once I've approved that, we'll do the problem statement, then routes, then copy.

Copy and adapt

We'll do this in [4] steps: [1. ...], [2. ...], [3. ...], [4. ...].
Do step 1 only: [...].
Stop and wait for my corrections before step 2.

Watch out: Very long conversations drift: old instructions fade and earlier versions creep back in. Copy the approved pieces into a fresh chat and carry on from there.

↑ Back to the lessons

09Lesson

Give feedback like a good creative director

"Make it better" and "more premium" give the model nothing to work with. Good feedback says what to keep, what to change and the effect you're after, the same feedback you'd want from your CD.

What to do

Say what works and should stay.
Quote the exact line or part you want changed.
Describe the effect you want, not just the fix: "it should feel like a friend's tip, not an ad".
Ask it to explain a choice before you reject it.
After three rounds of "no", rewrite the original instruction instead.

Why it works

Specific feedback narrows the search; vague feedback makes the model change random things and lose what was good.

Weak

Not quite. Make it more premium.

Better

Keep route 2 and the second headline. The body copy reads like a supermarket leaflet: 'great taste' and 'perfect for every day' are the problem. I want it to feel like a barista's tip to a regular. Shorter sentences, one concrete detail, no adjectives about taste. Change only the body copy.

Copy and adapt

Keep: [what works].
Change: ["exact quote"], because [what it does wrong].
The effect I want: [...].
Change only [this part]; leave the rest as it is.

Watch out: "Change only this part" matters. Without it, a model often rewrites everything, including what you liked.

↑ Back to the lessons

10Lesson

Make it check its own work, then check it yourself

Fluent isn't the same as right. AI writes clean sentences around invented numbers and claims nobody approved. Ask it to review the work against the brief, then read it the way you'd read a junior's first take.

What to do

Ask it to tag each claim: in the brief, an interpretation, or an assumption.
Ask it to review the answer against your "good means" criteria and the constraints.
Ask what a sceptical client would object to first.
Check every number, name and quote against the source yourself.

Why it works

A second pass with a different job (reviewer, not writer) catches a lot. But the model can't know what it doesn't know, so the final check is always a person's.

Weak

Is this good?

Better

Review your answer against the brief. 1) Tag every claim as STATED (with where in the brief), READING or ASSUMPTION. 2) Check it against the criteria and constraints I gave you and list anything it breaks. 3) Tell me the first thing a sceptical marketing director would push back on.

Copy and adapt

Review your answer:
1. Tag each claim: STATED (where), READING, or ASSUMPTION.
2. List anything that breaks my criteria or constraints.
3. What would a sceptical [client role] object to first?
Don't rewrite yet.

Watch out: If you show the client anything AI helped with, be ready to say what it did. "AI laid the groundwork, the idea and the decisions are ours" is an honest answer.

↑ Back to the lessons

Build your kit

The lessons pay off most when you stop starting from zero. Keep these somewhere you can paste from, and most good prompts become five minutes of work.

A context file per client: Brand voice, never-says list, audience in their words, history of what was rejected. Update it after each project.
Your style card: The rules from lesson 4, plus one example you're proud of.
Your templates: The goal, questions-first, range, feedback and review prompts from these lessons, adapted to the way you work.
One conversation per client and task: Never two clients in one chat, not even to compare. Start fresh when the task changes.

Most AI tools let you save project instructions or reusable files. Use them for the context file and the style card, so every new chat starts informed.

The whole school on one card

01Name the outcome and what good looks like.
02Give the source material, labelled, request last.
03Give a real role with a point of view.
04Show your style; get it back as rules.
05Constraints with reasons, and what to do instead.
06Let it ask questions before it answers.
07Ask for range along a line, with the bet named.
08One step at a time; fresh chat when it drifts.
09Feedback: keep, change, effect, only this part.
10Review against the brief, then check it yourself.

Written by people with 25 years on the agency side. Models change every few months; these habits don't. Groundwork: AI. The idea: ours.