Prompt Engineering Isn't Dead. Most People Just See One Piece of It. AI is getting better at understanding you. But that doesn't automatically mean you've become better at explaining yourself. I've seen countless posts claiming that prompt engineering is dead. You don't need to learn it anymore. Just tell AI what you want. And to a certain extent, they're right. You no longer need elaborate prompts, complicated agent systems, coded clones or multiple personas to get useful results. Sometimes, you can just ask. But here's what people overlook. Imagine asking AI to build you a website. You say, "Make me a professional website for my business." It generates something polished, but it's missing the features you need, the audience you want to reach, and the functionality you had in mind. So you start correcting it. Then correcting the corrections. Eventually, you've spent more time fixing the output than you would have spent explaining what you actually wanted. The problem isn't necessarily the AI. It's the gap between knowing what you want and being able to communicate it. A better model doesn't eliminate the need to communicate clearly. It simply becomes better at working with what you give it. And this is where I think the conversation about prompt engineering misses something important. To me, prompt engineering has multiple levels. Level 1: Understanding what you want to build. This isn't just about learning to speak computer language or providing context. It's about understanding your own intentions, breaking an idea down, identifying what you actually need, and articulating how you want to bring it to life. Sometimes, that means providing detailed instructions. Other times, it means saying, "I need an app that does these six things." The skill is knowing what to communicate, not how complicated you can make the request. Level 2: Understanding how to reconstruct what you know. To me, this goes beyond optimising prompts or collecting frameworks.

Stop talking to AI like it's a computer. Yes, technically, it is. Are you? AI understands natural language. You don't need to speak in weird, perfectly optimised prompts every time. Talk to it like a colleague. Explain what you're trying to do. Give it the relevant context. Tell it what's not working. For example: Computer mode: «"Generate a professional LinkedIn post about AI productivity. Use a conversational tone. Include three key points. Keep it under 150 words."» Human mode: «"I've been thinking about how people overcomplicate prompting. I want to write a LinkedIn post explaining why talking to AI naturally often works better. Make it sharp, not preachy, and something that sounds like me."» See the difference? The second gives the AI intent, context and direction, instead of just a list of instructions. That's why the "talk to it like an intern" advice works. You naturally give an intern background, context and direction because you're focused on getting the outcome, not constructing the perfect sentence. But don't dump your entire life story into the prompt either. Too much context becomes context overload. This is also why I use personas like Lisa, Riley and Ryan. I'm not trying to squeeze the "perfect response" out of AI. I'm trying to get something done. There's a difference. So stop obsessing over optimisation. Build an AI assistant that understands how you work, give it a role behind the scenes, and then talk to it like a person. The prompt does the heavy lifting. You just have the conversation.

Late night pro tip Pro tip: If you need AI to stay within a specific character or word limit, set a lower target than your actual maximum. Leave yourself a buffer of around 10 to 20 percent, or more if the output is particularly complex. Why this works AI models can struggle to produce outputs that consistently satisfy exact length constraints. Even when you specify a maximum, the generated response may exceed it. For example, suppose you need a copywriting prompt that fits within a 5,000 character limit. Instead of asking: "Create a copywriting prompt with a maximum of 5,000 characters." Ask: "Create a copywriting prompt for marketing purposes. Keep it under 4,000 characters." You've now created a buffer of 1,000 characters. If the model exceeds your requested target slightly, you still have a better chance of staying within your actual limit. The same principle applies when creating custom GPT instructions. If the platform allows a maximum of 8,000 characters, aim for perhaps 7,000 or 7,200 rather than targeting the absolute ceiling. For example project instructions for GPT are 8,000 characters and no more The important distinction This is a practical workaround, not a guarantee. A lower target reduces the risk of exceeding your limit, but it doesn't prevent the model from overshooting. For critical limits, count the characters or words after generation and ask the AI to shorten the output if necessary. The takeaway: Don't aim for the maximum. Aim comfortably below it, verify the limit leave yourself room for error. Even if it means you go even less A very expensive tip: This won't work if you're prompt is confusing and your contradicting yourself earlier on in the prompt you still need to structure it so that everything is flowing. This will save you a lot of rewriting rewording and re-clarifying

Your Design System Shouldn't Become Your Design Prison Experimentation Is Part of Good Design ----------- Here's something I've started noticing when I build things with AI. The more I experiment, the more I learn. Not just about AI. About what I actually like. When you find a workflow that works, there's a temptation to keep using it forever. The same sidebar. The same workspace. The same card layout. The same colour system. The same architecture. And eventually, you stop designing. You start repeating. That's where experimentation becomes important. Every time I deliberately build something differently, I discover another possibility. A different way to structure a product. A different way to organise information. A different visual language. A different approach to a logo, interface, workspace, or even the underlying architecture. Sometimes the experiment is terrible. That's fine. The point wasn't to make the perfect thing on the first attempt. The point was to find out what happens when I stop assuming that the thing I've already built is the best possible version. I've started thinking about it this way: Reliable, reusable, and repeatable are excellent qualities for how a product functions. They aren't necessarily qualities you want every product to have visually. A product should feel familiar enough to use. A design should have enough freedom to become something of its own. Otherwise, every new tool slowly becomes the previous tool wearing a different coat. So try it. Take one thing you're building and deliberately break one of your usual rules. Remove the sidebar. Change the navigation. Try a gallery instead of a workspace. Use a colour system you normally wouldn't touch. Build the interface around something other than the architecture you already know works. You might hate it. You might accidentally find something better. Either way, you've learned something. That's the real value of experimenting with AI. It isn't just that AI lets you create faster.

Everybody wants better results from AI until getting better results starts testing how they actually use it. Because eventually, this journey asks a question most people aren't prepared for: "What if getting the best possible answer isn't actually the point?" Not when your prompt works perfectly. Not when your outputs are getting better. Not when everyone is talking about how powerful your new AI workflow is. I mean when you're staring at an interesting result and your first instinct is to dissect it. "What was the exact prompt?" "What variable caused this?" "How can I reproduce it?" "How can I make it more reliable?" "How can I turn this into a framework?" Those questions matter. But here's the problem. Sometimes, you've stopped experimenting. You're optimising. And there's a difference. That distinction became something I've thought about a lot when using AI. Because I've watched people take something genuinely interesting and immediately try to systemise it. They find an unexpected answer. Then they reverse engineer it. They find a great prompt. Then they turn it into a template. They discover a useful workflow. Then they build a framework around it. And before long, every interaction with AI becomes another opportunity to optimise. Better prompt. Better structure. Better instructions. Better variables. Better outputs. Better system. And somewhere along the way, something gets lost. Curiosity. I remember thinking: "What happens if I don't optimise this?" What happens if I just follow the interesting answer? What happens if I ask the ridiculous question? What happens if I deliberately make the prompt worse? What happens if I remove a constraint instead of adding one? What happens if I combine two completely unrelated ideas? What happens if I give AI an instruction that makes absolutely no sense and see how it responds? That's where things get interesting. Because optimisation and experimentation are trying to do two completely different things.