What a Prompt Actually Is
A prompt is a brief. Treat it like one. When you hand a freelancer a vague request — "write me something about climate change" — you get back something vague. When you hand the model the same thing, you get the average of everything it has ever read on climate change: confident, generic, forgettable. The model isn't reading your mind. It's filling a gap with the most statistically likely text, and "most likely" is another word for "most boring."
The fix isn't a longer prompt. It's a specific one. A good writing prompt narrows the model's options until the only sensible output is the one you actually want. Five levers do most of that narrowing: role, audience, format, constraints, and examples.
The Five Levers
Role
Tell the model who it is. This sets the vocabulary, the assumptions, and the register before it writes a word. "You are a science journalist who explains hard ideas to curious teenagers" pulls the output toward concrete analogies and short sentences. "You are a peer-reviewer for an academic journal" pulls it toward hedged claims and citations. Same topic, completely different draft.
Skip the flattery. "You are a world-class, award-winning, genius writer" does nothing — there's no corpus of award-winning genius writing tagged as such for the model to imitate. Roles work when they're functional, not when they're impressive.
Audience
Who reads this changes everything about how it should sound. A draft for your professor and a draft for your group chat share zero sentences. Name the reader and their starting knowledge:
Audience: first-year students who have never taken economics.
Assume they know what money is but not what "inflation" means.
That second line is the one most people forget, and it's the one that stops the model from either talking down to your reader or burying them in jargon.
Format
If you don't specify structure, the model defaults to its house style: an intro paragraph, three or four headed sections, a tidy conclusion, and a closing line that starts with "Ultimately." You can feel the template. Break it by asking for the exact shape you want — number of paragraphs, whether to use headings, sentence-length targets, what to put first.
Constraints
Constraints are where good prompts get their personality. They're the rules that keep the model from drifting back to the average:
Rules:
- No words like "delve," "tapestry," "moreover," or "in today's world."
- Maximum 18 words per sentence.
- Open with a specific example, not a definition.
- British spelling.
Negative constraints ("don't do X") work, but positive ones ("do Y instead") work better — telling the model what to write beats telling it what to avoid, because "avoid" still puts the forbidden idea in front of it.
Examples
This is the strongest lever and the one people use least. Show the model one or two samples of the voice you want and it will match the pattern faster than any adjective could describe it. A paragraph of your own writing, a headline you admire, a paragraph you're rewriting — paste it in and say "match this tone." This is also how you teach the model your voice, which is the whole subject of the next chapter. If you want to go deeper on structuring multi-part instructions, the Advanced Prompt Engineering course is worth your time.
Putting It Together
Here's a weak prompt and its rebuilt version. The weak one:
Write a blog intro about productivity apps.
The rebuilt one:
Role: You are a tech writer for a student newsletter.
Audience: busy undergrads who are skeptical of productivity hype.
Task: Write the opening 2 paragraphs of a blog post reviewing
productivity apps.
Format: Paragraph 1 is a 2-sentence hook about a real frustration.
Paragraph 2 sets up what the post will cover.
Constraints:
- No phrases like "in our fast-paced world" or "game-changer."
- Concrete and slightly skeptical in tone.
- Under 120 words total.
Match the dry, direct tone of this sample: "Most habit apps die in
your phone's third screen. Here's what actually stuck."
The second prompt isn't longer for the sake of it. Every line removes a way the model could go generic. That's the real skill — not writing more, but closing exits.
Build Templates, Not One-Offs
You write the same kinds of things over and over: essay intros, email replies, summaries, social captions. Don't rewrite the prompt each time. Build a template once, with blanks, and reuse it:
Role: [who the model is]
Audience: [reader + their knowledge level]
Task: [what to produce, with length]
Format: [structure / sections / sentence rules]
Constraints: [banned words, tone, must-dos]
Sample to match: [paste a voice example, optional]
Keep these in a single notes file. Over a semester you'll collect a handful that consistently produce drafts you barely have to fix — an essay-outline template, a "make this email less stiff" template, a "summarize this reading in plain language" template. That library is worth more than any single clever prompt, because it turns a good result from a lucky accident into something you can repeat on demand.
One warning: a template gets you a strong first draft, not a finished one. The model will still occasionally invent a fact, misjudge your reader, or slip into its default cadence. Your job is to spot that and push back — change one lever, regenerate, compare. Prompting is a conversation, not a vending machine. The people who get the best output aren't the ones with the secret magic words. They're the ones who treat the first draft as a starting position and keep negotiating.