How to Write Better ChatGPT Prompts

Most people who say "ChatGPT is overrated" are getting bad results because they're writing bad prompts. The model is fine. The brief is the problem. After running thousands of prompts across ChatGPT, Claude, and Gemini, almost every disappointing answer traces back to one of seven mistakes — and each one has a clean fix. This guide is the seven rules, in the order they matter most. Skip ahead to whichever one looks most useful.

1. Be ruthlessly specific

Vague inputs create vague outputs. ChatGPT can't read your mind, your audience, or the context you're carrying around in your head. The fastest way to upgrade any answer is to swap general words for concrete details. Replace "small business" with "12-person bookkeeping firm in Texas serving solo founders." Replace "make it engaging" with "make the first sentence a question and use one specific number in each paragraph."

Before: "Write a sales email about our new feature."
After: "Write a 120-word sales email to existing free-tier users introducing our new bulk-export feature. Tone: helpful, low-key, no urgency. Mention that exports are available in CSV and JSON. End with a one-line CTA to enable it in account settings."

2. Set the role explicitly

Telling ChatGPT who to be steers the entire response. A prompt that starts "You are a senior product designer who has shipped at Stripe and Linear" produces noticeably different output than the same task with no role at all. Roles act like a filter — they pull the model toward the vocabulary, standards, and reasoning patterns of that profession. Pick a specific role, not a generic one. "Marketer" is weak. "Lifecycle email marketer at a Series A B2B SaaS" is strong.

Before: "Help me write a tagline."
After: "Act as a brand strategist who has named consumer products at frog and IDEO. Give me 10 tagline options for a meal-kit aimed at busy parents. Each should be under 8 words, sound human, and avoid food puns."

3. Provide context, not just task

Context is the single biggest difference between an amateur prompt and a professional one. Before you tell ChatGPT what to do, tell it the situation. Who is the audience? What's the goal? What have you already tried? What does success look like? Five extra sentences of context can eliminate three rounds of revisions.

Before: "Write a LinkedIn post about leadership."
After: "I'm a first-time engineering manager writing for an audience of senior ICs and other new managers. Last week I had to give critical feedback to a senior engineer for the first time and it went well. Write a 200-word LinkedIn post sharing what I learned, in a humble, observational tone — no listicles, no bold takeaways, just one honest reflection."

4. Define the output format

If you don't tell ChatGPT what shape the answer should take, you'll get a wall of prose. Decide up front: a table, a numbered list, three labeled options, a 300-word paragraph, a single tweet, JSON. Format defines what you can actually use. A "list of 10 ideas" with bolded names and one-sentence explanations is paste-ready. Ten ideas buried in flowing paragraphs are not.

Before: "Compare these three CRMs."
After: "Compare HubSpot, Pipedrive, and Close in a markdown table with these columns: best for, pricing tier we'd start on, biggest strength, biggest weakness, deal-breaker if you're a 5-person team. Add one row at the bottom recommending one for our use case (B2B SaaS, $2k ACV)."

5. Give an example (one-shot prompting)

The fastest way to lock in tone is to show, not tell. Paste in a sentence, paragraph, or post you like and ask ChatGPT to match the voice. This is called "one-shot prompting" — one example turns abstract instructions ("punchy, conversational") into something the model can actually pattern-match against. Your example doesn't need to be related to the topic; it just needs to demonstrate the voice or structure you want.

Before: "Write a casual product description."
After: "Write a product description in the same voice as this example: '[paste a description you like here]'. The product is a $24 reusable ice pack that stays cold for 12 hours. Match the rhythm, sentence length, and energy of the example. Don't copy phrases."

6. Set constraints

Constraints are how you eliminate the bland defaults ChatGPT falls back on. Word counts. Banned words. Required structure. Things to avoid. The more rules you give, the more the model has to actually think instead of reaching for stock phrases. A useful trick: add a "do not" list — "no exclamation points, no bullet points, do not use the words 'leverage,' 'unlock,' 'powerful,' or 'seamless'." Watch the writing improve immediately.

Before: "Make this email more engaging."
After: "Rewrite this email to be more engaging. Constraints: under 100 words, no exclamation points, no questions in the subject line, banned words: 'excited,' 'thrilled,' 'reach out,' 'circle back.' Open with a specific observation, not a generic greeting."

7. Iterate, don't restart

Beginners scrap a draft and rewrite the prompt from scratch. Pros keep the conversation going. ChatGPT remembers everything inside a single thread, so the second prompt is almost always shorter than the first: "Make it 30% shorter." "Cut the third bullet — it sounds salesy." "Rewrite paragraph two from the customer's POV instead of ours." Each round teaches the model what you actually want. By round three or four, the output is usually exactly right — and you've learned something about your own taste.

One specific iteration trick: when an answer is close but off, ask ChatGPT to critique its own draft first. "Before rewriting, list 3 weaknesses of the current version." Then have it rewrite based on those weaknesses. The self-critique step almost always produces a stronger second draft.

Putting it all together

Here's the same task written two ways. First, the throwaway version most people send.

Bad prompt: "Write a LinkedIn post about why remote work is good."

You'll get a generic, listicle-flavored post you've already seen a hundred times. Now the optimized version.

Optimized prompt: "You are a former management consultant turned remote-team operator who has run a 40-person distributed product team for 5 years. Write a 220-word LinkedIn post arguing that remote work isn't 'good' or 'bad' — it's a management problem disguised as a location problem. Audience: senior managers and founders, not junior employees. Tone: dry, observational, slightly contrarian. No bullet points. No bolded takeaways. Open with a specific scene or moment, not a thesis. End with a question that pushes back on a common assumption. Banned words: 'leverage,' 'unlock,' 'productivity,' 'flexibility,' 'work-life balance.'"

The optimized prompt locks in role, context, audience, length, format, tone, and constraints — and bans the five most-overused words in that genre. The output won't read like every other LinkedIn post. It'll read like something a real person with a real point of view wrote.

Common traps to avoid

1. Asking for "the best" without defining best. "Best for whom? Best at what?" If you don't define the criteria, ChatGPT will pick its own — and they probably won't match yours.
2. Stuffing too many tasks into one prompt. "Write the post, then critique it, then translate it, then turn it into 5 tweets." Big multi-task prompts come back shallow. Break them up.
3. Confusing "long" with "good." A great answer can be three sentences. Setting word counts down (not just up) often produces sharper output.

Where to go from here

Pick one rule above and apply it to a real prompt today. The fastest way to build the muscle is to do it on something that actually matters — not a practice exercise. Two stops that will accelerate the next twenty prompts you write:

• Prompt Improver — paste a weak prompt and we'll add the missing role, context, and format for you.
• Prompt Library — 200+ free, copy-pasteable prompts already written using all seven rules above.