How to write AI prompts that work: 7 simple patterns anyone can use

Most disappointing AI answers are not the model's fault. They come from prompts that leave too much to guesswork. After 14 years analyzing technology and retail tools, I have watched the same pattern play out with every new assistant: people type a short request, get a generic reply, and decide the tool is overrated. The fix is rarely a better model. It is a better prompt.

A prompt is the instruction or question you give a large language model (LLM), the type of AI behind tools like ChatGPT, Claude and Gemini. An LLM predicts a useful response based on the words you supply, so every detail you leave out gets filled in with the most average possible assumption. Learning how to write AI prompts is mostly learning which details matter. The seven patterns below are the ones I return to every week, whether I am drafting product descriptions, summarizing supplier contracts or comparing pricing data.

What makes an AI prompt good?

A good prompt removes ambiguity. It answers the questions a capable new colleague would ask before starting a job: who is this for, what do you already know, what exactly do you want, and what should the result look like.

In practice, strong prompts share four traits:

• Specific task. One clear verb, such as summarize, compare, rewrite or draft, rather than a loose request to help with something.
• Relevant context. Facts the model cannot know, like your audience, your brand tone or the data you are working with.
• Constraints. Length, format, reading level and things to avoid.
• A defined output. A table, a bulleted list, three options, a draft email. The model should know what finished looks like.

Length is not the goal. A 40 word prompt with these four traits will usually beat a 300 word prompt that rambles.

What are the 7 prompt patterns and how do you use them?

A prompt pattern is a reusable structure for an instruction, much like a template for a business letter. You swap in your own details, and the structure does the heavy lifting.

1. The role pattern

Assign the model a role and, just as important, name the audience. Example: Act as an experienced retail buyer. Explain these supplier terms to a first time shop owner. The role sets the expertise and vocabulary. The audience sets the depth and tone. Without an audience, a role alone often produces jargon.

2. The context first pattern

Lead with background before the request. Paste the relevant notes, data or previous email, then ask your question. Models read the whole prompt, but placing context first keeps your actual request clear and makes it easy to reuse the same context with different tasks. A useful habit: write a short paragraph describing your business, product or situation once, save it, and paste it at the top of related prompts.

3. The constraints pattern

Constraints are the rules the answer must follow. Good ones are measurable: under 120 words, no more than five bullets, written for a 12 year old reading level, avoid technical terms, use British spelling. Negative constraints help too. Telling the model not to include an introduction or not to repeat the question removes filler you would otherwise delete by hand.

4. The examples pattern

Show, then ask. Paste one to three examples of the output you like, then request a new one in the same style. This approach is often called few shot prompting, where each example is a shot. It is the fastest way to match a brand voice. If you write product listings, give the model two of your best listings and ask for a third for a new item. Keep examples consistent, because the model will copy their flaws along with their strengths.

5. The output template pattern

Spell out the exact structure you want back. For instance: Return a table with columns for Product, Price, Main benefit and Main drawback. Or give a fill in skeleton with headings. Templates are especially valuable when you plan to paste the result into a spreadsheet, a content system or a report, because consistent structure saves cleanup time.

6. The step by step pattern

For tasks that involve reasoning, such as comparing options, checking math or planning, ask the model to work through the problem in steps before giving a final answer. Researchers call this chain of thought prompting. A simpler version is to break a big job into stages yourself: first list the key points in this report, then group them by theme, then write a summary. Each stage is easier to check than one large output.

7. The critique and revise pattern

Ask the model to review its own work against criteria you set, then improve it. Example: Review your draft for vague claims, missing numbers and sentences over 25 words. List the problems, then rewrite it. This turns a single response into a short editing cycle. It works well for emails, sales copy and summaries where tone and precision matter.

How do the patterns compare?

Each pattern solves a different problem. The table below shows where each one tends to help most, based on my own use rather than formal testing.

How do you combine patterns into one prompt?

You rarely need all seven at once. Two or three patterns cover most everyday tasks. Here is a prompt that combines role, context, constraints and an output template for a common retail job:

You are a copywriter for an independent kitchenware shop. Our customers are home cooks aged 30 to 55 who value durability over trends. Write a product description for a 26 cm cast iron skillet, preseasoned, oven safe to 260°C, priced at 49 pounds. Keep it under 90 words, avoid the words premium and luxury, and use plain language. Return a headline of no more than eight words, then the description, then three short bullet points listing key features.

Compare that with the typical first attempt: write a description for a cast iron pan. The short version forces the model to guess the audience, length, tone, price point and format. The longer version leaves almost nothing to guess, and in my experience it usually produces a usable draft on the first or second try.

A simple order to follow when you build a prompt:

1. Role and audience.
2. Context and source material.
3. The task, stated with one clear verb.
4. Constraints and things to avoid.
5. Output format or template.
6. Optional: examples, or a request to critique and revise.

What prompt mistakes should you avoid?

Most weak results trace back to a handful of habits:

• Stacking several tasks in one request. Asking for a summary, a pricing analysis and a social post at once tends to produce three shallow answers. Split them.
• Vague quality words. Asking for something good, engaging or professional gives the model nothing measurable. Describe what those words mean to you.
• Assuming the model knows current facts. Many models have a training cutoff and may not know recent prices, launches or policies. Supply the facts or verify them afterward.
• Trusting numbers without checking. LLMs can produce confident but wrong figures, sometimes called hallucinations. Check any statistic, date or quote before you publish or act on it.
• Sharing confidential data carelessly. Read the privacy terms of the tool you use, and remove customer names, account details and internal financials unless your company has approved that tool for such data.

How do you test and improve a prompt over time?

Treat prompts like any other business asset. When one works well, save it in a shared document with a short note on what it is for. When an answer disappoints, change one element at a time, such as the constraints or the example, so you can tell which change made the difference.

Follow up messages are part of prompting too. Instead of starting over, reply with targeted corrections: make it shorter, use a warmer tone, replace the second bullet with something about warranty. Most chat tools remember the conversation, so small adjustments are cheaper than a full rewrite.

As a practical estimate, a team that keeps a library of 10 to 20 tested prompts for its repeat tasks can cut a lot of the back and forth that eats time in the first weeks of using AI. The patterns themselves stay the same across tools, which means the effort you put into learning them carries over when you switch from one assistant to another.