AI Has Changed Completely: Here's What Matters in 2026

A year ago, getting real results from AI was hard. That has completely changed. But there's a few big mistakes people are still making that are holding them back. It's not about finding the best tool, prompt engineering, or learning technical skills. I'll show you the few things that matter that will get you ahead of 99% of people that are using AI.

A trap I see all the time is people trying to test every new tool or chase every benchmark. In 2026, the bigger mistake isn't choosing the wrong tool. It's constantly switching before you've learned what any one of them can actually do. A beginner who learns Claude deeply will be far ahead of someone bouncing between ChatGPT, Claude, Gemini, Perplexity, and five other tools at a shallow level. These platforms get better the more you understand them. They learn your preferences. You can build in knowledge, and they start to automate parts of your workflow. That only happens if you go deep into one ecosystem. And I say ecosystem because I'm including more than just the chat interface. This includes the builder tools, memory systems, and agentic tools that are all connected. I'll cover that all later. The three ecosystems worth choosing from are ChatGPT, Claude, and Gemini. Just pick one of those three, whichever one you like. If you don't have a preference yet, it'll become clear as we go. If you feel like you're already behind because these have been out for a while, it's really only been the past few months where they've gotten good enough and easy enough at all the things I'll cover for anyone to learn. So, pick an ecosystem and go deep using what I'll cover here. That gets you 90% of what you need. The remaining 10% is a short list of specialized tools I'll cover at the end. First, we need to go through the part most people are getting wrong and that people are still giving out bad advice on prompting.

A lot has changed with prompting. A bad input still gets a bad output. But since the models have gotten so much better at understanding, you can keep it much simpler. So complex prompt engineering is usually a waste of time. Here's what actually matters. ICC: instructions, context, constraints. Give clear instructions. Define the task and the action you want taken. Include relevant context. Your role, objectives, background, anything relevant to the task. And set proper constraints. Rules, style, tone, length, output format. When it applies, include an example of the output you're looking for and ask it to match that. The order of these doesn't matter as long as you include all three. Think of this like giving a task to a new employee. They need the same things a person would. Clear instructions, relevant context, and proper constraints. If you give a person a vague task with no background, they'll give you a bad result. AI is no different.

Instructions and constraints are fairly straightforward, and these models have gotten much better at working with simple or even vague inputs. The part people tend to struggle with is context. It's also the most important because this is what personalizes the outputs to your situation. Giving more context is usually better, even a full context dump. These models can literally intake a full book of context. You're not going to overwhelm them, just as long as what you include is relevant. But it can be hard to think of all the context you need up front. You'll leave things out that would have helped. So, I have a technique that makes it easy called a context interview. At the end of your prompt, ask it to ask you questions to gather any additional context it needs to best accomplish the task. It will know what it needs better than you do. It asks, you answer, and the output is tailored to your specific situation rather than a generic response.

That context interview is a massive shortcut to better, more personalized responses. But that doesn't mean the outputs will be perfect on the first try. Treat the first output like a draft. It might not even be good. Iterate until it's perfect. Get used to using these tools as a collaborator. You assess the output, identify what needs to be improved, and ask again, just like working with a person you've given a task to. Combine ICC with the context interview, and your first output will be dramatically better. A few iterations after that, and it can be perfect. We'll revisit this as well. That context interview can become reusable knowledge. And these iterations can become reusable processes. Even with all of that, there's one thing that can still go wrong. Hallucinations.

AI models will hallucinate. That's where it confidently makes things up. It can sound super convincing about something that is completely wrong. This is just part of how generative models work, but that could mean getting a wrong answer you confidently send to a client or citing a study that was never written. I have a few quick tips that will dramatically reduce or catch hallucinations. Ask it to indicate confidence levels for each claim. Then it will flag anything it's uncertain about. Ask it to cite its sources. Then go check them for yourself. Ask it to find an expert that disagrees with what it just said. Another great option is to copy the response and paste it into a different model and get a critique. The models are often better at critiquing than generating from scratch. Those tips cut hallucinations down or give you a way to check for accuracy easily. So, especially for high stakes work, use those techniques. I mentioned that context is often the hardest part, but it's not the only place you can get stuck. Knowing what questions to ask or tasks to assign in the first place can be hard sometimes, too. So, we've compiled a list of the top 30 prompts that we've tested and gotten results from. Everything from the God Mode research prompt to the AI business architect prompt to a fix my thinking prompt. There's 30 of them you can copy paste across research, marketing, operations, writing, and productivity with placeholders where you can swap in your own context to personalize them. That's completely free. Just click the link in the description. Now, we've covered prompting, just a few simple practices you can use right away. But there are many layers beyond just sending a prompt, and that's where it gets really powerful.

And this is where picking one ecosystem and alerting it starts to really separate you. The features inside these platforms are where the real leverage is. Once you've picked one, you can start learning how to give it memory and repeat tasks. And nearly every important feature has a parallel across each platform. If you choose Claude, it's projects, artifacts, skills, and eventually Claude Code. If you chose ChatGPT, it's projects, canvas, custom GPTs, and eventually Codex. Gemini has notebooks, canvas, gems, and anti-gravity. They're not identical, but for most people, any path is more than enough. The difference between them matters less than utilizing the full depth you can build inside one ecosystem. Either way, here's how to actually use them.

Projects are where you can save your context and knowledge on a specific area. You teach them who you are and what you prefer. And you do this for each topic. So, what you're actually building is like a personal AI staff with one expert per topic who is already briefed and already trained. I'll demo this in Claude just so I can show a concrete example in one ecosystem. In Claude, these are over on the left panel. Then you can click to create a new project. This will be a self-contained workspace for a given topic. What sets this apart is three things. Instructions, where you can tell Claude context about what you're working on, process instructions, tone, and style preferences, and any other specific requirements. Memory. Claude reads your past conversations inside the project and saves what it finds relevant. You can review it and prompt for changes anytime. Knowledge. These are files, documents, brand guidelines, or reference material Claude can draw on. Just drop them in here. This is where that context interview comes back into play. Instead of doing a context interview for every business related task in separate chats, do one in-depth context interview for your business. Then at the end, you can ask it to format that into a document to upload into the knowledge base of your project. It will know how to do that. Then save that into the knowledge base of your project. Then you'll never have to answer those questions again. Same with custom instructions. Tell it you're creating custom instructions for your project. Give it the context you already have and ask it to ask you any more questions it needs to be able to create those project instructions for you. Now that this is set up, any new chat I start in this project will live here. And instead of copy pasting context into every chat or starting fresh every time, you have a Claude that actually knows you, your goals, your business, and has that built into every conversation. And create a separate project for your content, your client work, your business strategy, your health, travel, every topic. Each one will be ready whenever you need it without starting from scratch. This is the number one thing to do after this video if you've never done it. The change in experience in whichever of these ecosystems after you do this is so dramatic and immediate.

Projects give your AI persistent knowledge. Skills give it persistent processes. In the prompting section, I talked about iteration. You won't get the perfect result on the first try. It takes back and forth, which can take time. If you iterate for any process or task that you're going to do again in the future, in Claude, you can just say, "Package this as a skill." It will analyze the conversation, the process you went through, and package all the details up as a reusable skill. It saves that, and then whenever you do that process again, it will invoke that skill automatically and follow the exact process again. You can save a ton of time with these. Like, I have four skills I use just for helping with YouTube videos. And going back to the analogy of giving a task to a new employee, starting a new chat for each task is like hiring a new employee every day. Starting a new project is like giving that employee the employee handbook, all your SOPs, brand guidelines, and how you want them to work. And they've actually read all of that before their first day. And once you've walked them through a process, you package that up as a skill so they can continue doing it in the future. And just like a new employee, there will be times they don't follow the instructions perfectly or you realize you didn't instruct them properly. But in those cases, rather than training from scratch, you can just make a small adjustment to the project instructions or update the skill, you train them further over time and it sticks. That's why using the same ecosystem regularly gets better over time. And the parallels to these in ChatGPT projects are also called projects. The closest similarity to skills is custom GPTs. And Gemini projects are notebooks and skills are gems. There's some nuance in the details, but the concepts are the same with those. I do have a comprehensive tutorial on each of these tools if you need it.

Another core feature of these tools is the building mode. In Claude, this is called artifacts. In ChatGPT and Gemini, it's called canvas. It works the same in all three. When you ask for something that deserves standalone value, it will build it out in a side panel and that persists while you ask for changes to it in the chat panel. That could be a tracker, an interactive dashboard, a web page, a game. Just ask and it gets built. These can be incredibly useful and time-saving. For example, create a Facebook ad campaign results summary based on this data that I can present on a Zoom call. Then I will drop in a couple files I have with some numbers on them and send it off. That looks just amazing on the first try. But what's cool about this is it will stay on this side panel. So you can easily make changes using natural language without having to download it first or scroll back through the chat history. So I could drop in our logo and say use our brand colors. Then it will find the exact colors from our logo and update it right here. This dedicated panel makes it much easier to work with. This goes far beyond static documents. Artifacts can also build things you interact with directly. Here's an expense tracker where I can drop in a receipt and it will analyze and extract every purchase and give me a full dashboard with it all tracked and calculated. Or an interactive site to better learn how LLMs work. Each of these was built from just a single prompt. You can even share these. It will give you a URL that you can send to other people. Then they can use them and interact even if they don't have their own Claude account. You can build very useful tools right in the chat interface. You're already ahead of 90% of people using AI. But each of these has an even more powerful tool in their ecosystem for building that will get you better than 99%.

What you just saw in artifacts is powerful, but it has limits. You can build things that go way beyond what the chatter interface can handle. This is much easier than you'd expect, and you can create incredibly useful and time-saving tools. This is called vibe coding, describing what you want in plain language and letting the AI write all the code. You never touch a line of it. Everyone should try this regardless of how non-technical you think you are. You don't have to look at any code at all. You just need to learn how to describe the tool you want, test what the AI builds, explain what broke, and iterate. You will get stuck and have bugs. When that happens, you have a built-in tutor. Just ask it what you could have done differently, what the better approach was, and how to fix it. Here's a quick demo in Claude Code. This is using the desktop app which has Claude chat, co-work, and code all on separate tabs, but it's all under the same subscription. I click over to code and ask for what I want. I'm going to build a Kanban board that pulls in my meeting notes from Granola and extracts action items into a to-do list automatically. I described what I want in a couple sentences. I'm purposefully keeping this prompt simple. Now, I switch to plan mode. When I do that, it will ask me a few questions. This is its version of the context interview. I just answer all of those questions. If I wanted it to go deeper, I can have it ask me even more. Then it will lay out a plan. If it looks good, I approve it. Then it goes off and starts building for me. It will build every part of this, then go through and test the site itself. If something doesn't work, it will debug and fix it. Eventually, it will give me back the completed build. There it is. I've got a fully functioning app. It pulled in some of my meetings from Granola, then pulls it into this nice interactive, really aesthetic to-do list, all built from a plain language description, no code written by me. And this is just a starting point. I showed that with Granola just to show that you can connect to the apps you already use. That could be your calendar, your email, your project management tools. In the Claude Code tutorial I did recently, I built a similar app to this, but added in a Sauna integration. After I pulled those in, I could drag and drop people's names to assign tasks and push that directly to my team's boards. What you build can actually plug into your existing workflow. I also built a game, a website, a Chrome extension, and an app that utilized AI vision and analysis in that tutorial. And I used only a little more than what I just showed you. Claude has Claude Code. OpenAI has Codex which is also really powerful and pretty easy to get started with. For example, I built the exact same app in one shot there as well. And you actually get a lot more usage with Codex right now. You'll hit rate limits faster in Claude. You can even use Codex on the free plan of ChatGPT right now. And Google just released updates to anti-gravity that makes it actually competitive with Claude Code and Codex. But out of all those, Claude Code is my favorite. And I'm talking about using these for building personal use tools or internal tools for your team, not apps you're building to sell or make public. And this was of course not a full Claude Code tutorial. I wanted to show that you can just open it up and send a prompt and get something working back. There's much more to learn if you want to get all the power of it. I do have a full tutorial on it, but my recommendation is to open it up and send a prompt first before watching that. You can build something just like this or to solve all sorts of other problems just as easily as that was. Think of something you'd find useful. Doesn't have to be perfect. Write the prompt, switch to plan mode, and send it. Getting your first build done matters more than getting the perfect idea. Try it out first, then go deeper with that tutorial.

Now, I want to go a little deeper on choosing which ecosystem might be best for you. Again, each can get the job done for just about anything, but each have particular strengths. And I will mention here, beyond going deep, you absolutely will get the maximum out of the ecosystem on the paid plan. You won't get the most out of any of them on the free plan. So, I do recommend upgrading once you pick one. This can always change, but as of right now, if you want the broadest all-in-one AI app, ChatGPT does also include incredible image generation. It has a really good voice mode with live vision, a few other interesting features that Claude doesn't have. Those aren't that important to me. Like I do all my image generation on a separate platform and I don't really use voice mode. But that breadth is where ChatGPT shines. If you do anything that involves writing, Claude is the winner. I use it for critiques and intros and hooks all the time. So it is my personal primary ecosystem. As far as Gemini, it is great if you're already really deep in the Google ecosystem. The others can connect to the basic tools like Gmail, Calendar, and Drive. But if you want the Gemini Assistant on the side panel while you're within all the Google apps, that is available with the paid Gemini plan, and it's pretty nice. It also has the best integration with YouTube. I can just drop in a link or multiple links and it will analyze the full videos quickly and seamlessly. That's for the chat interfaces. The building tools that I covered in the last section, it's a little harder to say which is the leader and it's changing all the time. Claude Code is the most recognized and established. It's the one I use the most, but it is the most expensive for building. You'll run through your usage limits much faster there. Codex is similar in capability, but with much higher usage. Then anti-gravity just barely had a ton of brand new updates. So, I don't have enough experience using those yet, but it does seem like it will be a very strong competitor. Now, you might be thinking, what if you pick the wrong one? Your choice isn't permanent. You can switch later. But if you first stick with one for a couple months and actually build your workflow around it and go deep, it makes switching a lot easier later. You know what to look for and it's easy to pick up a new platform since so many of the features have a parallel. Whichever ecosystem you choose, once you've gotten this far, you're ahead of 99% of people using AI. But there are a handful of specialized tools outside these ecosystems worth knowing about.

First, like I've covered, get good at one ecosystem and you'll see how much you can already do. Oftentimes, you'll see AI tools that have a lot of hype and then realize you could do that within Claude with what you already know. Or if you can't do it natively, you can build it yourself in Claude Code. You only know that if you've gone deep. But there will be some limitations where a specialized tool can be genuinely helpful. I wanted to save the section for later because it's the part of AI that can feel really overwhelming and hold people back when really it shouldn't be where you start. Go deep into those primary ecosystems, then dabble here when you hit a limitation. There definitely are valuable other tools. I test a ton of these, but I'd say I only have around five that I use really regularly, and a couple others that are every once in a while. You really don't need a huge stack, but I'll cover some I use quickly here that I do find a lot of value in. First off is an important one. If you need anything creative that involves image and video generation, I use Higsfield for this. It has every leading model all in one place. I can run the same prompt through ChatGPT image 2 and nano banana pro at the same time to see which one looks better. And if a new model comes out tomorrow, that'll be in here too. Then I can bring them into the video tab and select the best model there. Currently that's Kance, but could be any of these others at a given time. And sometimes each model has their own strengths. So instead of paying for each platform and jumping between them, you can just do it in here. If you just need a one-off image every once in a while, just stick with ChatGPT or Gemini. They generate amazing images right there. But if you do it a lot, it's way easier in here. I can run four ChatGPT generations at once or 20 if I wanted instead of waiting one by one. Another I use all the time is Granola for meeting notes. That's not something that's built into those primary ecosystems yet, although I wouldn't be surprised if it is someday. I like Granola because it doesn't need to have a bot join the meeting. That's really nice. It has all sorts of other features, too. And I also connect mine to Claude, so it can pull the meeting notes into there. Whisper Flow is one I use all the time for speech-to-text dictation. It will correct misspeaks or you can go back and say that's not what I meant. I meant this and it will just figure it all out. It's way easier than traditional dictation because I mess up speaking all the time and I want to go back to fix it. Whisper Flow just does that all on its own. Now for deeper extremely customized agentic workflows and automation, N8 is amazing. It is definitely more technical to learn. You can do it as a non-technical person, but it is a heavier lift than anything in this video by far, including Claude Code. I do have a super in-depth tutorial on it if you want to learn. Notebook LM is part of the Google ecosystem, but makes more sense to include in this section. I love the tool. It's the best way to learn and organize information. You ground it in your own sources on any given topic across all sorts of formats. Then, it uses those to give you answers and cites all those sources by default that you can click through to see where it came from. That helps reduce hallucinations. Basically, it has some of those tips I recommended built in by default, but then you can transform that information into all sorts of different formats that help a lot with visualizing and retention. Things like infographics to see how things connect, generate a podcast for passive listening, or generate a quick quiz or flashcards. It's an amazing platform and you can get a lot on the free plan. And if you did choose Gemini as your tool, it does have a very seamless integration there since the notebooks you create show up in both places. And if you need music, Sunno is really fun. It's actually insane how good the music is that comes out of it.

Most of the other tools I used to use are either built into these main ecosystems natively, like Lovable for example. I used to use that for vibe coding, but I see no reason to pay for it when I can do it all in Claude Code or Codex or even just in artifacts and canvas all with the existing subscriptions I already pay for. And for plenty of the others, I've just built my own version in Claude Code, more customized with no extra subscription.

The biggest thing that separates people who get results from people who don't is implementation. It's rarely that you need another tutorial. Watch one, then go implement it immediately while it's fresh or even while you're watching. So, here's exactly what to do right now. Pick the main area you want help in. For most people, that's work or business. Start with a context interview. Answer every one of those questions and package that up for a project on that topic. Next, actually write down everything you do in a day, not just broad tasks. Break that down into every subtask you can. Drop those tasks in and ask what areas can be streamlined or automated. Then, iterate and refine. Also ask which of them you could build a tool for using Claude Code or Codex or just an artifact. Build that, iterate, refine, use it. Just implement and test one of these tips each day. That's all it takes to get ahead of 99% of people who are using AI.