Practical AI for non-AI people
Let's talk about how to best talk to AI to help you with complex practical problems.
In this post I’m going to show you a simple and practical strategy to get more out of working with AI. AI aficionados who’ll say it’s “old news” can scroll past — this post is for the rest of us who aren’t playing with the latest AI models every day.
Let’s start with theory before getting to practice.
Theory: what to know about AI models
What I often hear is a sentiment like, “Oh, everyone is saying AI is now PhD-level smart, but I asked it to count ‘r’ in ‘strawberry’ and it said ‘five’, so it’s clearly overhyped”.
No, it’s not. It’s a bit like speaking to a dyslexic PhD student and focusing on their not being able to spell instead of their research abilities.
AI can be incredibly powerful, but only if it’s used correctly, and here’s where many non-professionals make two mistakes.
One mistake is not to give the AI model enough context.
Another mistake is to assume the AI model will ask follow-up questions instead of making assumptions.
This is not a problem with humans. If a friend asks me, “My son is skipping school, what do you think I should do,” we can make two fair assumptions. First, I probably already know a lot about my friend, her son and his school, so I have some context. Second, I will be naturally curious and ask questions before volunteering with advice.
This is not how AI models work.1 AI models always start from scratch, as if you were talking to them for the first time.2 Even more importantly, they don’t ask follow-up questions unless you specifically tell them to.
This has two implications. First, the more context you give them, the better. Second, if you tell them to ask for relevant data, they will produce better results.
Practice: what does it mean for me?
Let’s suppose you have a long-term challenge you don’t know how to accomplish. What inspired me to write this post is that I used this approach to help me develop my concentration skills and the only reason I don’t want to make those chats public is that they contain some private data.
So instead let’s imagine you always wanted to develop a consistent exercise practice but never managed to do it.
A mistake 90% of people make they don’t even think about asking AI for help.
A mistake further 5% of people make is that they ask AI something generic like “how do I establish an exercise routine”.
If you do this, AI will give you a generic answer that you could have written yourself and that is completely useless. Why? Because it knows nothing about you. It knows nothing about your goals, your challenges, your exercise history, about how irresistible you find Nutella, about how running first thing in the morning outside is a non-starter because of childcare, about how uncomfortable you find most gyms because of weightlifting bros there, about how much you loved diving, but you’re not living by the ocean anymore… and many other details! These details matter!
So instead, we’re doing to use a different approach and be like the final 5% of AI users who are doing it right.
1. Prepare the prompt
The first step is preparing a prompt. So before we ask the AI to do something complex for us, let’s make sure we give it all relevant context. We’ll have a separate conversation with your AI of choice (ChatGPT, Claude, DeepSeek, Gemini, etc) just to design the prompt. Here’s what it could look like:
I would like you to design a prompt for AI to build a custom strategy for me that’s grounded in latest scientific research to help me establish an exercise routine that I will be able to stick to long-term. I want it to be customised to my specific circumstances, so I will give you as much as I can about my situation. Then, I want you to ask me follow-up questions to improve the prompt. Here is the context: I’ve never been able to stick to a consistent exercise routine because I’m working full-time and have two young kids, but I know it’s important…
In the rest of the prompt, tell it absolutely everything that could possibly be relevant. What you like and don’t like about exercise. What you tried and what worked. What didn’t work. Tell it every little detail that comes to your mind when you think about establishing an exercise routine. The more you give it, the better result you will get.
Now, it’s probably a lot of typing, but the good news is that you can use Wispr Flow (referral link) to dictate it to the computer. Sadly, built-in voice recognition (on Macs at least) isn’t great, despite all the recent AI advances, but third-party applications like Wispr Flow are killing it. It even gets my accent without mistakes and removes filler words. Since we all speak faster than we type, I highly recommend using it.3
After you send this very long question to AI, it will ask you a few follow-up questions because you asked it to in your prompt. Answer them in full, and then tell it to ask follow-up questions again. Do it until it starts asking about the same or irrelevant things and you start repeating yourself. Then, ask it to produce you a prompt.
What you will get is a prompt that includes key information from everything you told it to. Now you’re ready to ask AI to produce you a custom strategy.
2. Try a few models
Now, give your prompt to a few different models. If you’re only using ChatGPT, like many people, try at least 4o (default) and o3-mini (by selecting Reason in the composer). Alternatively, try Google’s Gemini (various models), Anthropic’s Claude or DeepSeek.4 All of them will give you slightly different results, but each proposed plan will be detailed and tailored specifically to your circumstances, not some generic advice. Choose the response that you like best for whatever reason and stick to that model.
3. Iterate the strategy
If there’s anything in that plan that doesn’t seem realistic, inspiring or accurate, share this with the model. I know it may seem like work, and it is, but the more you share with the model, the better your strategy will be. So if you look at something and it feels like “no, I can’t stand the idea of doing this”, tell it and it will rebuild the strategy.
4. Give it a go
Now, whatever your plan is, put it into practice. After a while, you’ll get some data. Maybe you’ll make some progress with establishing exercise habits. Maybe you’ll try to and fail. Maybe you won’t even try.
It doesn’t matter.
Whatever happens, it’s a valid learning and can be used to refine the strategy. So in a few days, you can go back to the strategy chat and say, for example:
Look, I was really optimistic about trying this approach to establish an exercise routine but in the three days since we spoke I didn’t exercise once despite a clear intention to do so. How do you think we need to update the strategy to make it more likely that I actually do it? Ask me several clarifying questions before making a suggestion.
It doesn’t matters that you didn’t exercise. This means the approach that was suggested wasn’t right and this will be taken into account when updating the strategy. Notice that again, we’re telling it to ask clarifying questions.
Conclusion
All of this sounds so simple if you think about it. Plus, it’s available literally for free on free tiers of every AI model. So there’s really no excuse not to think about literally every single problem you’ve been trying to solve and working with AI on making more progress:
developing a business
improving your management skills
navigating relationship difficulties
changing a job
making money
fighting an addiction
writing a book
dealing with difficult neighbours
learning a new skill
living with a chronic condition
The only thing that’s required from you is a willingness to leverage the power of modern AI by remembering two things: give it plenty of context, as much as you can, then tell it to ask follow-up questions for a few iterations and then feeding the results of that strategy back into AI, even if the result is “no result” — that doesn’t matter one bit. It will make the strategy better.
PS If you enjoyed this post, check out my post on what are AI-first startups if you haven’t read it yet:
This is changing with AI agents that are, for example, specifically learning to ask questions before thinking, e.g. OpenAI’s Deep Research, but regular AI models jump straight into answering — that’s what they were designed to do.
Except if you give them additional info through custom instructions in ChatGPT, Claude Projects and similar tools. And, we’re not talking here about fine-tuned models — this post is not for people fine-tuning their models!
For example, right now I'm speaking this paragraph instead of typing it, and it's a completely different experience. I don't have to worry about filler words or pauses or really anything else. I can just speak as if I were talking to a friend, and I can get high-quality transcription as the result.
Given that it's a Chinese model, assume that it will probably remember everything you tell it and associate it with your email address. So you've been warned. I’ve no idea how big of a deal that may turn out to be, but I personally assume that Google knows too much about me anyway from my email and docs, but the Chinese Communist Party should probably know less about me.