The gap between thinking it and building it There was a period where I could imagine software I had absolutely no idea how to build. I could picture the interface. I could imagine the workflow. I knew how I wanted it to behave. I could even explain, in detail, what I wanted the finished product to do. But there was a problem. Knowing what something should be and knowing how to build it are two completely different abilities. For a long time, that gap was enormous. You could have a genuinely good idea and still be unable to do anything with it because you didn't have the technical knowledge, time, or resources to turn the idea into a functioning system. Then AI changed something important. Not the ability to have ideas. The ability to cross the gap. Suddenly, I could take something that existed almost entirely in my head and start describing it. "Make it work like this." "Change this." "That part isn't right." "Add this." "Now connect these two things." And instead of the conversation ending at the limits of my technical ability, I could start iterating toward something real. That changes how you think about building. You don't necessarily need to know everything required to construct the final system before you begin. You need to understand what you're trying to construct well enough to guide the process. The bottleneck moves. It becomes less about: "Do I know how to build this?" And more about: "Can I understand what I'm trying to build clearly enough to direct its construction?" That's a very different skill. Because the people who can imagine sophisticated systems now have a new option. They can start building. Not perfectly. Not magically. And not without learning. But they can finally begin closing the distance between imagination and execution. And sometimes, that distance was the only thing stopping the idea from existing.

You don't actually understand something because an expert explained it to you. You understand it when you've broken it, changed it, tested it, and discovered why it works. That's why I think more people need to learn to experiment. Not just with prompts. With coding. With design. With building. With the way you work. We spend so much time asking: "What prompt structure do the experts use?" "What's the best framework?" "What's the correct design system?" "What's the industry standard?" "What does this AI expert recommend?" And those answers can be useful. But there's a problem. You're learning their conclusion, not necessarily their reasoning. If someone tells you that Structure A produces better results than Structure B, you know A worked for them. But if you actually test A against B yourself, you learn something much more valuable. You discover: Why A worked. Where B failed. What happens when you combine them. What happens when you remove a component. What happens when you deliberately do the "wrong" thing. And sometimes... You discover the advice was wrong for what you're trying to build. That's the part I think we overlook. Your favourite prompt framework doesn't need to become your religion. Your design system doesn't need to become a prison. An acronym isn't a law of physics. A guru's workflow isn't a law of nature. Unless something is genuinely set in stone, experiment with it. Change the prompt. Break the structure. Try the ugly design. Remove the step. Add the step. Build it completely differently. Then compare the results. Because there's a huge difference between: "Someone told me this works." and "I tested this, and now I understand why it works." The first gives you instructions. The second gives you judgement. And judgement is much harder to copy. So learn from the experts. Steal the frameworks. Study the best practices. But then close the book for a minute. Go into the lab. Change something. See what breaks.

Everyone is teaching AI to sound like them. Same tone. Same phrases. Same writing style. But here’s the problem: If AI needs you to tell it what to think every time, you’ve personalised the voice, not the intelligence. You didn’t build a digital version of yourself. You built a very convincing parrot. 🦜 The real shift in AI personalisation isn’t cloning yourself. It’s building a system that understands how you operate. Most people personalise AI by giving it examples: Your writing. Your tweets. Your prompts. Your tone. Your favourite phrases. Then they wonder why the output still needs constant correction. Because they’ve copied the surface, not the system. Think about what actually makes an experienced operator effective. It’s not just how they communicate. It’s how they: • interpret context • make decisions • recognise patterns • challenge assumptions • handle constraints • know what matters • know what to ignore • decide when something isn’t good enough That is the part most AI personalisation completely misses. I’ve seen this distinction repeatedly in my own work building systems like Lisa. The goal isn’t: “Make AI sound like me.” It’s: “Make AI understand how I work.” That requires something much deeper than feeding it more examples. You need: Context. Boundaries. Decision frameworks. Working principles. Relevant memory. Operational rules. Feedback loops. And enough situational awareness to understand that the same answer isn’t always the right answer. Because here’s the uncomfortable bit: More context doesn’t automatically create better intelligence. If you throw 500 examples into a system without telling it how those examples should influence decisions, you’ve created a bigger pile of information, not a better assistant. That’s why a simple AI clone can actually make your workflow slower. It produces something that looks right. You review it. You find the problem. You rewrite the prompt. You correct the output. You try again. And suddenly the thing that was supposed to save you time has become another job.

You might be getting worse at using AI by trying to get better results from it. Not because optimisation is bad. Because if you analyse every interaction, reverse engineer every result, and chase the “perfect” prompt, you eventually stop experimenting. And experimentation is where some of the most interesting AI discoveries happen. Getting better results from AI is useful. Better prompts and thoughtful testing can improve the way you work with it. But… Optimising every interaction can strip away the best part of using AI: experimentation. When people get an interesting result, they often stop to reverse-engineer it: “What was the exact prompt?” “How did you test it?” “What variables did you change?” “How can I make this more reliable?” Those questions matter when you’re building a repeatable process. But they don’t need to lead the way every time. Sometimes, keep exploring. Change the prompt. Ask the ridiculous question. Give it terrible instructions. Follow an unexpected answer down an interesting path. Experimentation can take you somewhere optimisation never would. You don’t have to turn every AI interaction into a prompt-engineering case study. When the only goal is the “best possible response”, you stop experimenting and start optimising. And optimisation isn’t always the point. Something to think about… Sometimes, the point is simply to press the button and see what happens. AI should help you think, create and explore not turn every moment of curiosity into a performance review.

Here is a tip before I head to bed. I just built a prompt that doesn’t simply improve the prompts I already use. It evolves them. Instead of asking AI: “Make this prompt better.” I give it something I already use and make it work through different layers: 1. INTENT What am I actually trying to accomplish? 2. ASSUMPTIONS What is my current prompt taking for granted? 3. STRATEGIES What completely different approaches could solve the same problem? 4. STRUCTURES How could those approaches be organised differently? 5. VARIATIONS What new prompts emerge from those approaches? 6. PRESSURE TEST Where does each version break? 7. SYNTHESIS Which ideas are actually worth combining? 8. EVOLUTION What should the next version become? The important part is that I never automatically keep the original. And I never automatically keep the AI's version either. AI can generate the possibilities. I decide what survives. Because I can't give you exactly what you want. Only you know what that looks like. What I can do is give you a better way to explore the space between what you have now and what you could build next. That’s the part I’m interested in. Not just making prompts better. Making it faster to discover what "better" could actually mean.