You asked for a LinkedIn post. You got four paragraphs that could have been about anything, opening with "In today's fast-paced digital landscape." So you added "make it more engaging and detailed." Now you have six paragraphs of the same thing.

That loop is the single most common way people use AI, and it's the reason most people quietly conclude the model isn't that good. The model is fine. The prompt gave it nothing to be wrong about.

The short answer

A prompt that works does four things: it assigns a role with real expertise, it gives context the model couldn't guess, it states a task narrow enough to fail at, and it specifies the output — format, length, and what to avoid. Vague prompts don't produce bad writing; they produce average writing, because averaging is exactly what a language model does when nothing in your prompt rules anything out. Every technique below is a different way of ruling things out. Length is not the variable. Constraint is.

Why "be more detailed" makes it worse

A language model predicts the most probable next chunk of text given everything before it. "Write a LinkedIn post about productivity" has millions of plausible continuations, and the model lands somewhere in the dead centre of all of them. That centre is what we recognise as AI voice: fluent, structurally correct, and saying nothing.

"More detailed" doesn't move you off the centre. It just asks for more of it. What moves you is information that makes most of those continuations wrong — a specific reader, a specific claim, a number, a thing you refuse to say. Every constraint you add deletes a chunk of the probability space. Delete enough of it and the only thing left is the post you actually wanted.

This is also why copying someone's 400-word "mega prompt" off Twitter usually disappoints. It's dense with instructions and empty of your specifics. Ten adjectives about tone constrain less than one sentence about who's reading it.

The four parts, and which one you're skipping

Almost everyone writes the task and skips the other three.

  • Role — "You're a direct-response copywriter who's written for DTC skincare brands." This isn't a personality costume. It's a filter that pulls the model toward a narrower slice of its training.
  • Context — who the reader is, what they already know, what happened before, what you've already tried. This is the part with the highest return and the one people skip hardest, because it's the only part you can't copy from a template.
  • Task — one verb. "Write" is not a task. "Rewrite this paragraph so the first sentence contains the objection, not the benefit" is a task.
  • Output spec — format, length, structure, and a ban list. Word counts are soft suggestions to a model; "three bullets, under 12 words each" is enforceable and it'll usually comply.

If a prompt of yours is underperforming, it's almost always missing context. Check that before you rewrite anything.

Examples beat adjectives, every time

"Make it punchy and conversational" is a request the model has to interpret. Two examples of punchy and conversational is a pattern it can copy. This is the highest-leverage move in prompting and the most underused, because pasting examples feels like more work than typing an adjective.

It isn't, really. Two or three samples of writing you like — yours or anyone's — will get you closer to a voice than any amount of tone description. And the samples don't need to be about your topic. You're demonstrating rhythm, sentence length, and how the thing opens, not subject matter.

Six prompts, annotated

Copy these. The line under each one tells you what it's doing and what to change.

1. The base template

You are a [specific role with a niche, e.g. "B2B SaaS onboarding writer
for technical users"].

Context: [who the reader is, what they already know, what's happened
before, what you've already tried and rejected].

Task: [one specific verb and object].

Output: [format]. [Length limit]. Do not use: [ban list].

Why it's built this way: it forces you through all four parts in order, so you can't skip context. Swap the role for something narrow — "copywriter" is too broad to filter anything; "copywriter for indie mobile apps" actually steers. The ban list is the sleeper feature: "no rhetorical questions, no em-dashes, don't open with a definition" removes more AI-smell than any positive instruction.

2. Plan before you write

Before writing anything, outline your approach in 5 bullets: the angle,
the reader's main objection, the one claim I need to prove, the structure,
and what you'll deliberately leave out. Stop there and wait for my go-ahead.

Why it's built this way: it splits thinking from drafting. You catch a wrong angle in 5 bullets instead of 800 words, and "what you'll deliberately leave out" forces the model to commit to a scope rather than covering everything shallowly. Use it for anything longer than a paragraph. Drop the "wait for my go-ahead" if you're in a hurry — you'll still get a better draft.

3. Make it interview you

I want [outcome]. Before you produce anything, ask me the 5 questions
whose answers would most change what you write. Ask them one at a time.
Don't produce the deliverable until you've asked all five.

Why it's built this way: this is the fix for "I don't know what context to give." The model knows what's missing better than you do; this makes it say so. "One at a time" matters — ask for five at once and you'll get five and answer two. Swap in a bigger number for genuinely complex work, like a strategy doc or a pricing page.

4. Match a voice with samples

Here are three things I've written: [paste 3 samples].

Describe my voice in 6 specific mechanical rules — sentence length,
how I open, punctuation habits, words I never use. Then write [new piece]
following those rules exactly.

After writing, list any rule you broke and why.

Why it's built this way: you're making the model extract the pattern explicitly instead of absorbing it vaguely, and the self-audit at the end catches drift. "Mechanical rules" is doing real work — ask for a description of your voice and you'll get "warm and authoritative," which is useless. Swap the samples, not the instruction.

5. The critique loop

Score that draft 1-10 on: specificity, whether a competitor could publish
the same thing word-for-word, and whether the first line earns the second.
For anything under 8, rewrite only that part. Show the rewrite, not a
summary of it.

Why it's built this way: self-critique only works when the criteria are things that can actually be failed. "Is it good?" gets you a 9 and a compliment. "Could a competitor publish this?" is a genuine test. Swap the three criteria to match what your work is judged on — for an ad, use "does it name the objection in the first line."

6. Kill the tells

Rewrite this with these rules: no sentence over 20 words unless the one
after it is under 8. No list where a sentence works. Cut every hedge
(somewhat, often, generally, can help). Cut every sentence that would
survive deleting it. Return only the rewrite.

Why it's built this way: it's all negative constraints, which are enforceable in a way that "sound more human" isn't. The alternating-length rule is the single biggest fix for AI rhythm, which is uniform in a way human writing never is. Run it as a second pass on any draft, including your own.

Which pattern to use when

Pattern Use when What it fixes Cost
Base template Any first draft Generic output from no constraints 2 min to fill in
Plan first Anything over 300 words Wasted drafts on the wrong angle One extra turn
Interview me You don't know what context matters Missing context you couldn't name 5 questions
Voice samples Output has to sound like a person AI cadence and vocabulary Pasting 3 samples
Critique loop Draft is fine but forgettable Averageness, no edge One extra turn
Kill the tells Final pass, every time Hedges, bloat, uniform rhythm 10 seconds

Most work needs two of these, not six. Base template plus kill-the-tells handles the majority of everyday writing. Add plan-first when the stakes go up.

The part nobody tells you

Your first prompt is a draft. Professionals don't write one perfect prompt; they run one, read what's wrong, and fix the specific thing that's wrong — usually by adding one constraint, not by rewriting from scratch.

So when the output disappoints, resist retyping the whole prompt. Ask instead: what would I have had to say for this specific failure to be impossible? Then add that one line. Three rounds of that beats an hour of prompt-crafting, and you end up with a template you can reuse.

Keep the ones that work. That's the whole discipline.

Ready to stop writing prompts from scratch? Our Advanced Prompt Engineering collection has the structured, tested versions of everything above — role-context-task-output templates, critique loops, and voice-matching chains, ready to paste.

Related guides