How to Write Better AI Prompts in Under 10 Minutes (With Examples)

Why are your AI outputs so generic?

You have been there. You type a question into ChatGPT, read the response, and feel nothing. It is technically correct. It is also completely useless, the kind of paragraph that could have been written about any company, any product, any audience. You close the tab thinking AI is overhyped.

The problem almost never lives in the model. It lives in the prompt.

Large language models are prediction engines. They generate the most statistically likely next word given what you gave them. If you give them a vague three-word question, the most likely output is a vague, generic answer. The model is doing exactly what you asked. You just asked for the wrong thing.

Think of it this way: if you hired a new freelancer and handed them a sticky note that said "write me some marketing copy," you would get back something forgettable. Give that same freelancer a one-page brief with your product, your target customer, three competing examples, a word count, and the tone you want, suddenly the first draft is usable. The model works exactly the same way.

The fix is not complicated. There are five things every useful AI prompt needs, and most people include one or two. Add the other three and your outputs will change inside the same conversation.

What makes a great AI prompt? The 5-part formula

This framework applies whether you are writing ChatGPT prompts, working with Claude, or using Gemini. The underlying principle is the same: give the model enough signal to stop guessing. Here are the five parts.

Part 1. Role

Tell the model who it is. Not just "you are an expert," but a specific professional with a specific background. The role activates a narrower slice of the model's training and pushes it toward domain-specific vocabulary, reasoning patterns, and conventions.

Without role: "Write a subject line for my email."

With role: "You are an email copywriter who has written B2B cold outreach
for SaaS companies for ten years."

The role does not need to be long. One sentence is enough. What matters is that it is specific rather than general.

Part 2. Context

Context is the background information the model cannot know unless you provide it. Your product, your customer, the situation you are in, competing examples. OpenAI's prompt engineering documentation notes that "the more information you give the model, the better it can tailor its response to your needs." That is not a suggestion. It is how the system works.

Without context: "Help me write a product description."

With context: "My product is a $29 prompt template pack for solo consultants
who charge $150/hr and bill 20 hours a week. They are not technical.
They want to save time on client deliverables without sounding like a robot."

The context section is where most people leave the most output quality on the table. Paste in real text. Paste in a competitor description you admire. Paste in three past emails you wrote that sounded right. The model will pattern-match from those examples.

Part 3. Task

One clear verb. Write, summarize, list, compare, rewrite, extract, critique. If you cannot state the task in a single sentence starting with a verb, the task is not clear enough yet.

Weak task: "Can you help me think about my landing page?"

Strong task: "Write three alternative headlines for my landing page.
Each headline should be under 10 words and focus on time saved,
not features."

Vague task language ("help me with," "think about," "explore") produces exploratory outputs. If you want something specific, ask for something specific.

Part 4. Constraints

Constraints are the boundaries. Word count, what to avoid, the reading level, banned phrases, the tone. Without constraints, the model defaults to the most statistically average version of what you asked for. Constraints force it out of average.

Without constraints: "Write a cold email."

With constraints: "Under 120 words. No subject line question marks. No emojis.
No mention of 'AI-powered.' Write as if the sender is a person,
not a company. End with one specific question, not a call to action."

Anthropic's prompt engineering documentation makes a similar point: "Be clear and direct" is the single most useful piece of advice they offer. Constraints are how you enforce clarity.

Part 5. Format

Tell the model what the output should look like. Numbered list, bullet points, a table, a three-paragraph email, an FAQ, a script. When you skip the format spec, the model guesses. It often guesses wrong, handing you a five-paragraph essay when you needed a six-item checklist.

Without format: "Give me some ideas for social posts."

With format: "Give me 5 LinkedIn posts. Format each post as:
Line 1: One-sentence hook (no questions)
Lines 2-4: Three bullet points with specific examples
Line 5: One call-to-action sentence
Each post should be under 150 words."

Format specs are especially important for outputs that go directly into other tools, spreadsheets, email clients, CMSs. Specifying the structure once saves you from reformatting every output manually.

What does a bad prompt look like vs. a great one?

Theory is easier to absorb with concrete examples. Here are three side-by-side comparisons covering common use cases. Read both versions and notice what information the great prompt provides that the weak one leaves out.

Example 1: Blog post intro

WEAK:
"Write an intro for a blog post about productivity."

GREAT:
"You are a productivity writer who covers tools and systems for freelancers.
Write an opening paragraph for a blog post titled 'Why Your Task Manager
Is Making You Less Productive.' Target audience: freelancers with 3+ years
of experience who already use tools like Todoist or Notion and are skeptical
of adding more systems. Tone: direct, slightly contrarian. No rhetorical
questions. Under 80 words."

Example 2: Customer email reply

WEAK:
"Write a reply to an angry customer email."

GREAT:
"You are a customer support specialist for a small e-commerce brand
that sells handmade goods. A customer emailed to say their order arrived
damaged and they want a refund. Our policy is to offer either a replacement
or a full refund. Write a reply that acknowledges their frustration,
apologizes without admitting legal fault, and offers both options clearly.
Under 100 words. No corporate-speak. First-person singular, warm but
professional tone."

Example 3: Social media caption

WEAK:
"Write a caption for an Instagram post about my new product launch."

GREAT:
"You are a social media copywriter for a direct-to-consumer skincare brand
targeting women 28-42. Write a caption for an Instagram post announcing
the launch of a new SPF moisturizer priced at $38. The product's main
differentiator is that it does not leave a white cast on darker skin tones.
Tone: confident, inclusive, factual. No exclamation points. No hashtags
in the caption body. Under 60 words. End with a single clear CTA to
click the link in bio."

Notice what every great prompt does: it names who the model is, gives enough background to make decisions, states exactly what to produce, and draws a box around the output with constraints and a format. That structure is repeatable across any task.

How do I get AI to write in my voice?

This is the question I get most often from content creators and solopreneurs. The good news is that the fix is simple. The bad news is that most people skip it.

The fix: paste samples of your own writing into the prompt. At least three examples. Then ask the model to extract a style description before writing anything new.

"Read these three emails I wrote. Extract a list of style observations:
sentence length patterns, vocabulary level, structural habits, phrases I
use, things I avoid. Then write a new email in that style about [topic]."

[Paste Email 1]
[Paste Email 2]
[Paste Email 3]

The extracted style list is something you can save and reuse. Drop it into any future prompt as a context block. Over time, your AI outputs start to sound less like a LinkedIn thought-leader and more like you.

A few other techniques that help: tell the model what you do not sound like ("not corporate, not academic, not casual-bro"), give it a reference writer whose style is similar to yours, and ask it to flag any phrases in its output that feel off-brand so you can address them in the next iteration.

If you want a head start on the style-extraction process, the AI Starter Pack includes pre-built voice-extraction prompts you can run against any writing sample in under five minutes.

How do I get AI to do research without making stuff up?

Hallucination is real. Models are trained to produce fluent, confident-sounding text. They will invent a citation, a statistic, or a case study rather than admit uncertainty. This is not a bug you can patch with better hardware. It is a property of how these systems work, and it requires a workflow response.

Two techniques reduce fabrication significantly.

Source-grounded prompting. Paste in the source material and instruct the model to work only from what you gave it.

"Based only on the text below, summarize the three main arguments.
Do not add information from outside this text. If a claim is not
supported by the text, say so rather than inferring it.

[Paste article, report, or transcript here]"

Explicit uncertainty permission. Models default to confidence because confident outputs feel like better service. You can override this.

"If you are not certain of a fact, say 'I am not certain about this'
rather than stating it as fact. I would rather have a caveat than
a wrong answer presented with confidence."

Google's Gemini prompt guidance specifically recommends adding grounding context when you need reliable, source-based outputs. The same principle applies across every major model. Treat unsourced AI outputs as draft-quality claims that need verification before you publish or send them.

For anything that involves numbers, recent events, or named people, run a quick verification pass. The five minutes you spend checking two facts is faster than managing the fallout from publishing an invented statistic.

Does this work with Claude and Gemini too?

Yes. The five-part formula is model-agnostic because it addresses what every large language model needs: a persona to adopt, background to work from, a task with a single clear verb, guardrails on the output, and a structural template.

That said, there are practical differences worth knowing.

Claude tends to follow formatting instructions precisely and handles long-form tasks well. It also tends to add caveats and hedges unless you explicitly tell it not to. If you want direct, assertive copy, include "be direct, skip the hedges and caveats" in your constraints. Anthropic's prompt engineering documentation is worth reading if you work with Claude regularly.

Gemini has a stronger real-time web integration than the other models, which makes it better for research tasks that require current information. Google's Gemini prompt guidance emphasizes being explicit about what type of response you want (factual, creative, analytical) at the start of the prompt. That instruction also improves outputs from other models, so it is a good habit regardless of which tool you use.

ChatGPT (GPT-4 and later) responds well to few-shot examples, showing it one or two examples of what you want rather than only describing it. If your format instructions are complex, pasting a sample output and saying "format it like this" often works better than writing a detailed format spec from scratch.

The core framework transfers. The adjustments are small. If you already have a prompt library that works in ChatGPT, most of those prompts will produce comparable outputs in Claude or Gemini with minimal editing. For a deep dive on applying this across business use cases, the post on using ChatGPT as a small business owner covers the workflow side in more detail.

The meta-prompting trick: using AI to write your prompts

Once you understand the five-part formula, you can use it to do something that sounds like a paradox: ask the AI to write your prompts for you.

Meta-prompting works like this. You describe the task you are trying to accomplish, tell the model what the output will be used for, and ask it to write you a high-quality prompt structured around role, context, task, constraints, and format. You then run that prompt.

"I need to write a prompt that will produce a competitive analysis
comparing my SaaS pricing page to three competitors. The analysis
will be used by my marketing team to redesign our pricing page.
Write me a well-structured prompt for this task, including a clear
role, relevant context placeholders, the specific task, constraints,
and output format."

The model knows its own tendencies. It will often produce a prompt with structure and constraints you would not have thought to include. This is particularly useful when you are approaching a task type you have not prompted for before.

If you want to take this further, the Stop Guessing Prompt Builder is a structured tool I built specifically around meta-prompting principles. Instead of writing your prompts from scratch, you answer a short set of questions and it assembles the five-part prompt for you. Useful when you want the framework applied consistently without rebuilding it every time.

There is also a broader collection of tested, ready-to-use prompts in the Business & Strategy Prompt Pack if you would rather start with proven prompts and adapt them, rather than building from scratch.

Frequently asked questions

How long should a good AI prompt be?

Long enough to cover the five parts: role, context, task, constraints, format. That usually runs three to eight sentences. Longer prompts are not inherently better, a focused 80-word prompt beats a rambling 400-word one. If you find yourself writing a page, break the task into steps.

Do these prompting techniques work with Claude and Gemini, not just ChatGPT?

Yes. The five-part framework works across all major models because it addresses what every large language model needs: a persona, background information, a clear task, guardrails, and a structure. Claude tends to follow format specs more precisely. Gemini benefits from shorter context when searching recent information. The core formula does not change.

Why does AI keep giving me generic outputs even when I try to be specific?

Usually because one of three things is missing: context about your actual audience, a constraint on what to avoid, or a format spec. Generic prompts train the model to produce generic answers because there is no signal pointing toward something specific. Paste in a sample of your own writing or a competitor example and the output quality jumps almost immediately.

Is prompt engineering hard to learn?

No. The fundamentals take under an hour to learn and a week of daily practice to internalize. You do not need to memorize jargon or follow any official certification. The five-part formula covers 90 percent of everyday use cases. Advanced techniques like chain-of-thought prompting and few-shot examples layer on naturally once the basics are solid.

How do I stop AI from making up facts?

Two habits help most. First, paste in your source material and tell the model to work only from what you provided. Second, add "if you are unsure of a fact, say so rather than guessing" to your prompt. Models are polite by default and will often invent a confident-sounding answer rather than admit uncertainty. Explicitly giving them permission to say they do not know reduces fabrication significantly.

What is meta-prompting?

Meta-prompting means asking an AI to write or improve your prompts for you. You describe the task you are trying to accomplish and ask the model to produce a well-structured prompt you can then run. It sounds circular but it works well: the model knows its own tendencies and can build in the role, context, and constraints that humans tend to forget. The Stop Guessing Prompt Builder automates this process.