A common complaint about AI is that it gives you generic junk. Look at the prompt behind that junk and it's often one line. "Write a blog post about my business." "Make this email better." "Give me marketing ideas."
That's usually not an AI problem. That's a prompt problem.
I lean on AI a lot. I call my AI agent Punk, and Punk runs a small team of specialist bots that help me with content, websites, and the daily grind across my businesses. Every one of those bots runs on written instructions. When the instructions are clear, the work is good. When they're vague, I get vague work back. Same as hiring a person.
Here's how I think about writing prompts, and what the companies that build these models actually recommend.
What is an AI prompt?
A prompt is everything you give an AI model before it answers. Your question, your instructions, any text you paste in, and any examples you show it.
OpenAI's prompt engineering guide describes prompt engineering as writing effective instructions so a model consistently produces content that meets your requirements. That word consistently matters. Anybody can get one lucky answer. A good prompt gets you a good answer every time you run it.
You don't need to be technical. If you can write clear directions for a new employee, you can write a good prompt.
Why most AI prompts get bad results
The AI doesn't know you. It doesn't know your business, your customer, your voice, what you already tried, or what "good" looks like to you. If you don't tell it, it fills in the blanks with the most average answer it can come up with.
Anthropic, the company behind Claude, says it well in its prompting best practices. Think of the model as a brilliant but new employee who lacks context on your norms and workflows. Their test is simple. Show your prompt to a colleague with minimal context and ask them to follow it. If they'd be confused, the AI will be too.
Read your prompt like you just walked in the door with zero background. Would you know what to do?
The five parts of a good AI prompt
Good prompts usually cover the same five pieces.
- The goal. What are you trying to get done, and why? "Write a follow-up email to a lead who went quiet after a sales call, so they reply and we can book a second meeting."
- The context. Who is this for, what's the situation, what do you already know? Paste in the notes, the old email, the product details. Don't make it guess.
- The rules. Tone, length, what to include, what to avoid. "Friendly, not pushy. No exclamation points. Under 120 words."
- The format. How do you want the answer shaped? A bulleted list, a table, three options, a subject line plus body.
- An example. If you have one, show it. "Here's an email I wrote that I liked. Match this style."
You don't need all five every time. A quick question doesn't need a novel. But when you care about the result, cover as many as you can.
Anthropic also points out that giving the reason behind a rule helps. Their example is telling the model not to use ellipses because the text will be read aloud by a text-to-speech engine. The model can then generalize from the why, not just the rule. Same with people. "Keep it short because they read on their phone" beats "keep it short."
A simple AI prompt template you can copy
Here's a basic structure for anybody starting out. Fill in the brackets and delete what you don't need.
You are helping me with [the job, like writing a sales email or planning a blog post].
Goal
[What I want to get done and why it matters.]
Context
[Who this is for, what's going on, and anything you need to know. Paste notes or documents here.]
Rules
- Tone is [friendly / direct / professional].
- Length is [a specific length, like 3 to 5 sentences or under 150 words].
- Do [the things you want].
- Instead of [the thing you don't want], do [what to do instead].
- If you don't know something or it isn't in the context, say so instead of guessing.
Format
[Exactly how the answer should look.]
Example
[Optional. Something you liked that it should match.]
The labels aren't magic. They just make it obvious where the instructions end and the context begins. OpenAI's guide recommends Markdown headings, lists, and sometimes XML-style tags to mark the sections of a prompt, and Anthropic recommends tags when a prompt mixes instructions, context, examples, and pasted documents. Plain labels work fine for everyday use.
Before and after prompt examples
This is where it clicks for most people.
Before. "Write a product description for my candle."
After. "Write a product description for a hand-poured soy candle that smells like cedar and orange. It's for my online shop, and my customers are mostly people buying gifts. Use a warm, simple tone. Write one 3 to 5 sentence paragraph, then three short bullet points with the scent, the burn time placeholder [X hours], and the size placeholder [X oz]. Don't make up any facts that aren't in this prompt."
Notice the specific length. OpenAI's best practices for prompt engineering call this out. Instead of something fluffy like "fairly short, a few sentences," say "use a 3 to 5 sentence paragraph."
Before. "Give me blog post ideas."
After. "I run a lawn care business for homeowners. Give me 10 blog post ideas that answer questions my customers search for before hiring a lawn service. For each one, give a working title and one sentence on who it helps. No news or trend topics. I want posts that stay useful for years."
The second version tells the AI who you are, who the reader is, what you want, and what to skip. That's the whole game. If you want to take those ideas further, I walk through how to write a blog post outline that keeps the post on track.
Show examples when the format matters
If you need answers in a specific shape, show it instead of describing it. This is called few-shot prompting, which just means giving the model a few examples of what right looks like.
Google's prompt design strategies for Gemini go as far as recommending that you always include few-shot examples, and note that clear examples can sometimes replace instructions entirely. They also warn to keep the format of your examples consistent, because the model copies the pattern you show it. Anthropic's guide says examples are one of the most reliable ways to steer format, tone, and structure, and suggests making them relevant and varied so the model doesn't latch onto the wrong pattern.
Want social captions in your voice? Paste three captions you already wrote and liked, then ask for new ones "in the same style as these examples." That beats describing your voice in adjectives. It's also why I think every creator should keep a story library. Your real stories and past posts are the best examples you can feed a model.
Put long documents first and your question last
When you paste in something long, like a transcript, a report, or a pile of customer reviews, order matters.
Both Anthropic and Google recommend putting the long content near the top and your actual question or instructions at the end. Google suggests a transition phrase like "Based on the information above..." to bridge into your request. Microsoft's prompt engineering guide makes the same general point. The order you present information in can change the output.
For shorter prompts, OpenAI's tips suggest putting the instruction first and separating the text with markers like ### or triple quotes. Either way, the big idea is the same. Keep your instructions and your source material clearly separated.
Tell it what to do, not just what not to do
A prompt that's nothing but "don't do this, never do that" leaves the AI guessing what you do want.
OpenAI's help docs give this exact advice. Instead of only saying what not to do, say what to do instead. Their example is a support chatbot. Rather than "DO NOT ASK USERNAME OR PASSWORD," they tell it to diagnose the problem and refer the user to a help article instead of asking for personal info.
"Don't be salesy" is weak. "End with a helpful next step for the reader, not a pitch" is something a model can actually follow.
Give the AI an out so it doesn't make things up
AI models can state wrong facts with total confidence. One of the easiest ways to cut down on that is to give the model permission to say "I don't know."
Microsoft's guide calls this giving the model an "out." Their example is asking a question about a piece of text and adding something like "respond with 'not found' if the answer isn't present." That small line can help the model avoid making up an answer.
This is a hard rule for my own content bots. If a fact or number can't be verified, it gets cut or qualified. It never gets invented. A shorter answer that's true beats a longer one that's wrong. You still need to check anything important yourself.
How to fix a prompt that isn't working
Prompting is iterative. Google's guide says so right at the top. Its tips are starting points, and you refine based on what you get back. Don't expect the perfect prompt on the first try.
When the answer misses, try these, roughly in this order.
- Add the missing context. Most bad answers come from the AI not knowing something you assumed it knew.
- Get more specific about the output. Length, format, audience, tone. Pick one and make it precise.
- Show an example. If describing it isn't working, show it.
- Rephrase it. Google's guide notes that different wording can get different responses even when it means the same thing.
- Break it into steps. Instead of "research, write, and format this," do one step at a time and feed each result into the next.
- Ask the AI to critique your prompt. "Here's my prompt. What's unclear or missing?" is a quick way to find the gaps.
One more thing. OpenAI's guide notes that different models, and even different versions of the same model, can respond differently to the same prompt. If you switch tools and your old prompt gets worse, it's not you. Test it again and adjust.
Save your best prompts
Once a prompt works, don't lose it. Keep a simple document or note with your go-to prompts for the jobs you do over and over, like follow-up emails, social captions, meeting summaries, and blog outlines.
A good prompt is an asset. Write it once, reuse it, and improve it when something changes.
Saved, tested prompts are also the first step toward automation. My bots all run on written instructions and rules. If you're curious where that leads, I broke down what an AI agent is and how it's different from just chatting with a bot. And if you want specific tools to practice on, my list of top AI tools for freelancers is a good place to start.
Final thoughts
Writing good AI prompts isn't a secret skill. It's giving clear directions. Say what you want and why, give the context, set the rules, show the format, and give an example when it matters. Then tell it what to do instead of only what not to do, and give it permission to say "I don't know."
The AI is only as good as the instructions you give it. And just like I said when I wrote about how to use AI to create content, AI helps with the pass. You keep the voice. A great prompt gets you a strong first draft. Your judgment is still what makes it worth publishing.
FAQ
How do I write a good AI prompt?
State your goal, give the background the AI needs, set clear rules for tone and length, describe the format you want, and include an example if you have one. Be specific, and tell it what to do instead of only what not to do.
How long should an AI prompt be?
As long as it needs to be to give the AI the context it's missing. A quick question can be one line. A task you care about, like a sales email or a blog outline, usually needs a few short paragraphs of goal, context, and rules.
Do AI prompts work the same in ChatGPT, Claude, Gemini, and Grok?
The basics work across all of them. Clear instructions, good context, and examples help every model. But different models and versions can respond differently to the same prompt, so test your important prompts when you switch tools.
How do I stop AI from making things up?
Give it the facts it needs in the prompt, tell it to use only that information, and give it an out, such as "say you don't know if the answer isn't in the text." Then fact-check anything important before you use it.
What is few-shot prompting?
Few-shot prompting means including a few examples of the output you want right in the prompt. The AI picks up the pattern from your examples, which is often more reliable than describing the format in words.