How to write an AI prompt that actually works

Four layers separate a prompt that returns something usable from one that returns filler. Here is what each layer does, why leaving one out breaks the output, and five complete prompts built exactly this way — free to copy, no email needed.

Why a one-line prompt gives you a one-line-quality answer

Ask for "write me a cold email" and you will get a cold email. It will be grammatical, it will be structured, and it will be generic — because with nothing else to go on, the average of everything is the safest answer a model can give.

The fix is not a longer prompt. It is a more decided one. Every detail you leave out is a decision you have handed to the model, and the model will resolve it toward the middle. The four layers below are simply the four decisions that matter most.

The four layers of a working prompt

1. Goal — name the artifact, not the topic

"Write about onboarding" is a topic. "A 90-word welcome email that gets the user to complete step one" is a goal. The difference is that the second one can be finished, and you can tell whether it was. Name the thing you want to be holding when the response ends.

2. Constraints — the layer that does the most work

Constraints are what force the model to choose. "Under 90 words" makes it decide what actually matters. "No buzzwords, no false urgency" removes the padding it would otherwise reach for. "One benefit per sentence" stops three claims from collapsing into one vague one.

If you only ever add one layer, add this one. A goal without constraints is still a wide-open question.

3. Voice — who is speaking

Left unspecified, voice defaults to corporate-neutral — the register that fits everywhere and belongs nowhere. Name the speaker and the relationship: "one small business owner talking to another," "a senior advisor who has seen this mistake before," "calm and practical, you know I work alone." Voice is not decoration — it changes which words are available.

4. Format — the layer that makes output usable

Format is the shape the answer arrives in: headline under ten words, one-line sub, three bullets, one call to action. Without it you can get genuinely good writing that still needs to be cut apart before it can be used anywhere. With it, the response arrives already shaped like the thing you were going to build.

One line versus four layers, side by side

The skeleton you can reuse for anything

Every prompt further down this page is this same skeleton with different content. Copy it, fill the brackets, and you have written a structured prompt:

You are [the role that would normally do this job].
GOAL: [the finished artifact you want to be holding]
CONSTRAINTS: [length limit] [what to avoid] [the one rule that matters most]
VOICE: [who is speaking, to whom, and how]
FORMAT: [the exact shape: sections, lengths, order]
Context: [what you know that the model does not]

Five prompts built this way

These are five I keep and reuse, one for each job a one-person business ends up doing itself. They are reproduced in full — nothing held back for the email list.

1. The cold-email writer

The job an outreach copywriter would do. The constraint layer carries this one: "under 90 words, no hype" is what forces it down to the part a stranger would actually read.

You are writing a cold outreach email.
GOAL: Get a first reply from a [business type] owner who has never heard of me.
CONSTRAINTS: Under 90 words. No buzzwords. No false urgency. One clear ask (a 15-min call or a reply).
VOICE: Warm, peer-to-peer. One small business talking to another. Not corporate, not salesy.
FORMAT: Subject line + 3 paragraphs (observation / credibility / ask).
Context: I help [business type] with [specific problem]. Their pain point is [X].

2. The landing-page section drafter

The job a conversion copywriter would do, for one section at a time. Here the format layer is the one earning its keep — it is the difference between good writing and a block you can paste into the page.

You are writing one section of a landing page.
GOAL: A finished [hero / problem / proof / CTA] section ready to publish.
CONSTRAINTS: One benefit per sentence. No filler: no "unlock," "transform," "game-changing."
Max 120 words total.
VOICE: Confident and plain. Sounds like someone who has already solved this problem for others.
FORMAT: Headline (under 10 words) / one-line sub / 3 bullets / single CTA line.
Context: The product is [X]. The reader's pain is [Y]. The outcome they want is [Z].

3. The content repurposer

Ask an AI to "make this shorter" and you tend to get the same words in a smaller box. The constraint no parroting is what turns one piece into three that each stand on their own.

You are repurposing an existing piece of writing.
GOAL: Three distinct cuts of the same idea for different surfaces (tweet / short post / caption).
CONSTRAINTS: No parroting. Each cut finds a fresh angle — a different entry point to the same idea.
Under 280 chars for the tweet. Under 150 words for the post. Under 20 words for the caption.
VOICE: Match the tone of the original exactly.
FORMAT: Label each cut clearly. Deliver all three in one response.
Source text: [paste here]

4. The decision pressure-tester

It is easy to use AI to confirm what you already want to do. This one is pointed the other way — you are asking it to argue against you before you spend the week.

You are a sharp outside advisor stress-testing a decision.
GOAL: Find the 3 strongest reasons this decision is a mistake, and the 1 condition that would make it right.
CONSTRAINTS: No diplomacy. No hedging. Argue against it as hard as you can.
VOICE: A senior advisor who has seen this exact mistake before. Blunt, not unkind.
FORMAT: 3 numbered objections (2–3 sentences each) + one "this works if" paragraph at the end.
Decision to test: [describe the decision, the cost in time/money, and what you're hoping it achieves]

5. The weekly plan builder

Not a to-do list — a layout that fits the hours you actually have. The constraint that makes it hold up is protecting one deep-work block a day; without it you tend to get a tidy table that falls apart on Monday morning.

You are an executive assistant planning my work week.
GOAL: A concrete day-by-day layout that fits my actual available hours.
CONSTRAINTS: Protect one deep-work block (90+ min, no interruptions) each day.
Don't front-load reactive work in the morning. Leave one buffer day or half-day.
VOICE: Calm and practical. Async-first. You know I work alone.
FORMAT: A simple table — Day / Deep-work task / Quick win / Admin slot.
My priorities this week: [list 4–6 items with rough time estimates]
My available hours per day: [e.g. 4 hrs Mon–Fri, 2 hrs Sat]

How to fix a prompt that is still giving you generic output

When a structured prompt still returns something flat, it is almost always one of these, in this order:

1. The goal is a topic in disguise. If you cannot tell whether the response finished the job, the model could not either. Rewrite it as an artifact with an end.
2. There is no real constraint. "Keep it concise" is not a constraint — nothing is ruled out. "Under 90 words, no adjectives in the first sentence" is.
3. The voice is a mood, not a person. "Professional but friendly" describes almost every piece of writing ever published. Name a speaker instead.
4. The format is missing. If you find yourself reformatting the answer by hand, that shape belonged in the prompt.
5. You gave it no context it could not guess. The model does not know your customer, your price, or what you already tried. That is the one thing only you can add.

If the output still reads like it was written by nobody in particular, the problem is usually upstream of structure — it is that nothing in the prompt rules anything out. I went through the mechanism behind that, the specific phrases to ban, and five paste-able repairs in why AI writing sounds generic.

Common questions

Prefer to watch it? The five prompts above are broken down on screen in Episode 2 of The AI Nomad.