Large Language Model (LLMs) based AI is growing in popularity at a phenomenal rate. As a result there is a constant ‘creation’ of new users. We are all new users when we begin, after all.
I’m writing this in late 2025, and as of now the predominant LLMs are Gemini, ChatGPT, DeepSeek, Claude from Anthropic, and Llama. I’m deliberately excluding image generators like Midjourney, but there are similar rules for them.
How to write better AI prompts for any AI model?
Obviously with so many different potential requirements, there can be no single answer. (Unless 42 works here, Douglas?) Here are some well-tested principles that will massively improve the results you get.
Here are my 5 essentials for writing the best possible AI prompt. Given the pace of change I hope this will be useful for at least 18 months. Bear in mind each system will have its own quirks and nuance, so this is general advice and principles.
1. Provide context. Provide instructions.
Most AI systems are silence-fillers. If you don’t tell it why you want something written, or whom the intended audience is, the AI will guess. It will fill that silence with the most generic response likely to meet most people’s needs. That will give you a middling answer at best. Give the full context that you’re operating in, and be really explicit about your objectives. Where possible lay out clear instructions and steps you want the AI to take to achieve what you’re asking. That can be as simple as;
Find and review at least 20 examples, and then derive a guide for best practices.
There are frameworks that help prompt you to consider what should be included, and how to structure that information. More on that in point 5.
2. Don’t use AI to replace yourself!
AI is generating content so quickly there is a huge amount more content flooding platforms. Even some ‘influencers’ and bands have recently been exposed or revealed themselves to be AI generated. Generally this is termed ‘AI Slop’.
Content that has your personal opinion, is probably going to be much more interesting to a majority of people. ‘Your take’ on a subject is likely going to be more interesting than reading the equivalent generic AI piece.
Many people have a strong visceral response to the kinds of generic content AI can be used to produce. There’s a certain generic look and feel to AI images, especially if people put in bland non-specific prompts. The same is true with the text it creates. This piece from The Atlantic about how ‘AI Slop’ is devaluing social media platforms is well worth a read.
3. Ask an AI to help
Often the best prompts you’ll see for image generation with Midjourney have been written by Gemini or ChatGPT. But meta-prompting is also really helpful. Rather than jump straight in trying to figure out a good way to get what you want, why not ask. For instance;
I'm trying to do X, and I want it to achieve Y, how should I write a prompt that gives you enough detail to be able to help?
When you see the AI’s prompt, you’ll likely notice more gaps in the context that you’ll need to fill. So it’s also useful for fleshing out the ideas. And relatedly…
4. Iterate
You just won’t get the best answer first time around. Sorry, that’s just how it is. But it’s not like visiting the Oracle at Delphi where you have to settle for a cryptic or imperfect answer. (I also don’t wish to tarnish the name of the Oracle at Delphi, I believe Alexander the Great liked it.)
Look at the AI response, and work out where and why it’s made leaps of logic. Often you can fix things with a corrective, such as;
Rewrite this to make it appropriate for a school-age audience.
Many AI models have a tendency to be fluffy and verbose. They’ll use too many words to get across relatively simple concepts. You have to read the content to ensure it makes sense and is a sensible way to say it. You know what you want to say, has the AI captured it clearly?
Also bear in mind AI models still have tendency to ‘hallucinate’ – i.e. make mistakes. They can completely make-up facts, and source them to made-up websites. It’s important to carefully check output.
5. Use a framework
To get the best from these systems, with the fewest rounds of iteration, frameworks are very useful. There are many frameworks out there, all touting particular niches that they deliver for. Most of them have a huge amount of crossover. Which is actually a good thing as it shows there’s a limited number of approaches.
That said, I do suspect the authors prioritised nifty acronyms over function. For instance, I used to use R.I.S.E.N., but others like R.O.A.D. and C.R.E.A.T.E. have been useful. If that sounds like nonsense, you can read more detail about these structures in this piece on frameworks for marketers.
Returning to point 3 above – the AI will likely give you the best framework to structure what it needs. So feel free to use these examples if they are memorable, but my AI’s advice was to include the following:
For Problem-Solving and Analysis
This structure is ideal for tasks that require the LLM to process complex information, analyze a scenario, or act as an expert in a specific domain.
- Role: “You are a financial analyst specializing in stock market trends.”
- Scenario: “A client is considering investing in a new technology company, ‘Innovate Corp.'”
- Input: “Here is the company’s financial report and recent news articles.”
- Task: “Analyze the provided information and provide a detailed risk assessment for the investment.”
- Nuance: “Consider both the positive growth potential and the competitive market landscape.”
- Expectation: “Present your findings in a structured report with a clear conclusion and a list of key risks.”
For Creative or Persuasive Writing
Use this structure when the output needs to be highly creative, engaging, or tailored to a specific audience.
- Role: “You are a marketing strategist for a tech startup.”
- Audience: “The target audience is young entrepreneurs and small business owners.”
- Context: “We’re launching a new project management tool called ‘TaskFlow’.”
- Task: “Write a short, engaging blog post about the benefits of using ‘TaskFlow’.”
- Tone: “The tone should be enthusiastic, approachable, and slightly informal.”
- Expectation: “The post should highlight three main features and end with a call to action to sign up for a free trial.”
- Example: “Provide a short example of a paragraph to follow.”
Additional considerations
AI is controversial, so you should be transparent about when you’re using it.
Every technological change since the Spinning Jenny has been controversial, especially when it threatens people’s livelihoods. AI has already begun to replace members of the workforce. Combining it with robotics will replace a wider swath of employees. Therefore some people will take a principled stand and not use it. Therefore morally you should be upfront if you’ve created with AI, so people know how to read it.
A huge drawback to AI LLMs is their extraordinarily high energy and water usage. Arguably this is an even more important reason to ensure your prompts are well-thought out. Above I suggest iteration of prompts as it is important to get to the best result. However you have to bear in mind the environmental cost of every interaction.
Conclusion
There isn’t a ‘correct way’ to prompt an LLM-based AI system. But there are ways to improve the results you get, and it’s not difficult. If the first response you get is disappointing it’s not a reason to give up. You simply need to analyze what context or instructions you’ve missed out. Then think through some of the other types of information we’ve mentioned above, to see what else could help.
If you do all of that, and still find you’re not happy with the result… It’s possible only you know how to create what you want. Also, in the spirit of transparency mentioned above, I will say this was written by me. I don’t have the excuse of blaming AI for the quality of the writing! Hopefully it will be slightly less anodyne than the equivalent written by AI.
Also published as How do I Structure my AI Prompt on LinkedIn.