Transcription
Everyone you know is about to ask you about Claude, and here's how you can actually help them. You already know the backstory. Anthropic told the Pentagon no. The White House retaliated, and the public responded by making Claude the number one app in America. Millions of people who had never heard of Anthropic a couple of weeks ago have just downloaded this new app.
And here's the problem. Almost all of them are going to treat Claude as a drop-in replacement for ChatGPT. Same problems, same expectations, same workflow. And that's not how AI works. AI models are not interchangeable brands like Coke and Pepsi. They're built differently, trained differently, and optimized for different things. Switching from ChatGPT to Claude with the exact same habits, it's like switching from Excel to Photoshop and wondering why the spreadsheet features are missing. Look, they're both software. Yes, they're LLMs, but at this point, they've diverged so much that you really can't call them the same tool.
People who open up Claude and type their usual ChatGPT prompts and get back kind of unremarkable answers, are not going to understand what they're missing, and they're probably going to walk away in frustration when they realize things that they've taken for granted in ChatGPT, like image generation, just aren't there in Claude.
This is the guide for the conversation you're going to have when your friends say, "Hey, what is this Claude thing?" It's not a feature tour. It's a practical explanation of what Claude does differently than ChatGPT. How you can use those differences, and what changes about your work when you do. And it's all grounded in what we can all verify. It's not marketing claims from either company.
So, what's actually different? The differences are not cosmetic. Claude and ChatGPT were built with very different training approaches and different philosophies. And those approaches produce measurably different behavior. ChatGPT's default behavior tends toward being more agreeable, more expansive, and at least in some personalities, more warm. If you ask it a question, you often get a thorough answer plus context you didn't request, plus an offer to elaborate. Now, OpenAI has worked hard to rein in the worst excesses of this pattern, but the general orientation to be helpful, to be thorough, to keep the conversation going, that remains the default. And for hundreds of millions of AI users, this is what they think AI is.
Claude was built using an approach called constitutional AI, where the model is trained against explicit principles: be helpful, be honest, avoid harm, rather than purely optimizing for what feels like a good response. The practical effect is that Claude is more likely to flag a problem than to smooth it over. It's more likely to ask what you're really trying to achieve here than to rush to produce something plausible. It tends toward conciseness rather than padding.
Now, this doesn't mean one tool is better than the other across the board. It does mean they have very, very different strengths, and using Claude really well requires understanding how those strengths are different and how to activate them. This is what that looks like in practice.
First principle: Claude is more likely to tell you your plan has a hole in it. So, ChatGPT has a documented tendency towards sycophancy, right? It's got a documented history of telling you what you want to hear rather than what you need to hear. OpenAI's own researchers have acknowledged this, most visibly when a GPT-4o update in April of last year made the problem so extreme they had to roll it back within a few days. Since then, OpenAI has put very serious work into fixing it, including refining training techniques, building new eval metrics, and publicly committing to steering models away from uncritical agreement. And so, the current ChatGPT is meaningfully less sycophantic than the version that triggered the rollback. But that underlying tendency hasn't fully disappeared because it's rooted in the training approach. OpenAI's models are trained heavily on user feedback: thumbs up, thumbs down, which inherently rewards responses that feel satisfying to a human in the moment. Anthropic's constitutional AI trains Claude against explicit principles like honesty, which creates a super different default posture. The practical difference is that Claude is somewhat more likely to flag a concern, to question your framing, to tell you something you didn't ask to hear. It's not necessarily dramatically more likely, but it happens enough that you're going to notice it over a couple of days of real use.
And this really matters for your work because the most expensive AI mistakes are not factual errors these days. They're plans that should never have been executed, the ones that went unchallenged and produced AI slop like a hiring plan with a timeline that assumes engineers ramp in three months when the real number is six, or a pricing strategy that ignores a competitive response. You get the idea. Claude is more likely to flag these kinds of issues than ChatGPT is right now. Not always, not infallibly, but the difference between slightly more likely to push back or somewhat more likely to push back and less likely, that really compounds when you're talking with AI frequently. And what it means is your plans start to get stress-tested more often. You make fewer expensive mistakes, but you also have to be okay with having your ego pushed back on some. You can't just expect that Claude is going to agree with you.
Principle two: When you're using Claude, you want to describe your situation, not your desired output. So, in ChatGPT, people will often write a prompt like a command: "Write a cover letter." "Give me five ideas." Now, Claude will respond to this just fine, but it responds to situations noticeably better. And the difference in output quality is worth calling out here. We know this is true. One of the reasons why this is more likely to be the case for you is because Claude was trained via constitutional AI to reason about whether a request is well-framed, while ChatGPT was trained via RLHF, or reinforcement learning with human feedback, to satisfy the request as stated. That difference predicts some behavior here. A model trained to evaluate framing will do more with a well-framed input. Multiple independent comparison reviews from Access Intelligence, from Type.ai, from Fluent Support, and others note that Claude tends to ask more clarifying questions and engages more deeply with context than ChatGPT.
So, why does this matter for you at work? Well, Claude is going to have a little bit more trouble guessing beyond what you've told it. If you give it a thin situation, you're going to get thin thinking. If you give it a really rich context layer, you're going to get strategic reasoning that changes how you approach the problem. Now, yes, more context is helpful for any model. The difference is what each model does with that richness. ChatGPT tends to use additional context to produce a more detailed version of exactly what you asked for. Claude tends to use it to think about how you framed the task. And Claude may come back with something that is more than what you asked for, in some cases less than what you asked for, in some cases exactly what you asked for. But either way, Claude is responding and addressing the frame and the context you gave it. It feels more like a thinking partner. And so, before you tell Claude what to make, it makes sense to spend a couple of sentences on what you're dealing with. Claude will appreciate it, and so will you.
Principle three: Give Claude your work, not a blank canvas. This is very counterintuitive for people who think AI is for generating content. Claude is better at editing and refining existing work. You can get Claude to generate work from nothing, but it's a little bit more concise, and you have to be very specific in your ask if you're asking it to generate work. In a blind test conducted in February of this year with over a hundred voters per round across eight prompts, Claude won four of eight rounds, while ChatGPT won one. This was from Access Intelligence's independent comparison of different models. And what they found was that users consistently rated Claude's outputs as more natural and publishable without a ton of editing. So, Claude scored 85% on the structural coherence of text versus ChatGPT's 78%. And this is across a big piece of text, like a 2,000-word analysis. Type.ai's analysis documented that ChatGPT tends to fall into a very distinctive AI voice, while Claude's outputs read a little bit more like human writing. And multiple reviewers like Fluent Support or Logic Web or Medium's I Technically can independently reach the same conclusion: Claude tends to write more naturally, and ChatGPT sounds more generic.
All of this talk about text and writing matters a lot because we spend so much time writing during the workday. So, if you are structurally editing, not just grammar fixes, but someone is telling you like, "The third paragraph undermines the first," or "You buried your strongest point," Claude does really good work at that level. ChatGPT tends to polish at the individual sentence level. So, if you run the same document through both with the instruction, "What is the weakest argument here and how could you fix it?" then you have a chance to compare the outputs, and what you tend to find is that Claude has that eye for prose structure in a way that ChatGPT is a little bit weaker. The other thing that I'm very aware of is that I've talked about the idea that Claude is more concise, and so it can be challenging to produce longer pieces of text, but I've also talked about the idea that Claude is good at writing prose that is more in a human voice. These things are in tension, and you have to work with the prompt right now to get Claude to write a piece of prose at a length you're comfortable with in a style that matches your voice. And the style piece is going to come easier, especially if you give Claude an example, much easier than with ChatGPT. But the length is something that you may struggle with. And when you're used to getting very wordy responses from ChatGPT, that can feel like a surprise.
Principle four: It is okay to ask Claude to show its reasoning. Claude has a capability called extended thinking. The model allocates additional processing to work through complicated problems step by step before answering, and then shows you the chain of reasoning it followed. So, on genuinely hard problems like contract analysis or debugging intermittent failures, extended thinking is going to produce meaningfully better output. In fact, Anthropic reports up to a 54% improvement on hard reasoning tasks. Part of how this works is that Claude is showing its reasoning as it works through the problem and burning extra tokens, and then reading that reasoning to continue solving the problem. And this is a really important distinction because technically speaking, Claude's models are not inference compute models. And so, where OpenAI burns inference tokens to come back with an answer, and sometimes, like on the pro versions of OpenAI, takes a really long time, like 20, 30 minutes on a task to come back, Claude tends to respond more quickly but uses all of that verbiage to stay on track and solve the problem. This really matters because you can see the chain of thought in Claude's writing and change or arrest it over time. In co-work, you can actually send a message to Claude and change how the agent will respond before the agent finishes the task. In the regular Claude chat, you can't do that yet, but you can see the response. And if you don't like where the chain of thought is going, because you can open it up and look at it as it produces, you can just hit stop and send a new message and clarify. This really changes how you think about approaching hard problems because you have to ask yourself, "Is this problem something that I am going to need to intervene on?" And if I need to intervene on it, have I set myself up to focus on the running text that will be coming through that chat window as Claude is thinking about it, and given myself the focus to catch when that goes off the rails and say, "No, no, no, no, not that." Because anyone who uses Claude a lot will tell you that they often are doing that unconsciously. They're so used to working with Claude and the way it produces text to solve problems that they're just kind of keeping an eye on it. They're like, "Ah, no. I stopped it and then I changed it." ChatGPT users work differently. They're used to just hitting go and then waiting for the response. So, there's a little bit more steering along the way you get with Claude.
Principle number five: You're building a workspace, not a chat box. Now, both Claude and ChatGPT have projects for work. Both let you upload documents, both set persistent instructions, and both organize conversations around domains. So, that concept is super familiar. The way most people use projects, though, is incorrect. They treat them like filing cabinets. You stick in docs. You send a vague instruction like, "Help me with marketing." And then you get conversations that are barely different from not using the project. Here's how to use projects correctly. Your project's custom instructions should be operating rules for every conversation in that workspace. Not, "Help me with marketing," but, "I'm a product marketing manager at a B2B SaaS company in cybersecurity. My team sells to CISOs and IT directors at mid-market companies which have 500 to 2,000 employees. Our biggest differentiator is ease of deployment, and my VP prefers data-backed arguments and dislikes jargon. All content should align with the positioning doc which I've uploaded and the brand voice guide also uploaded." Now, every single conversation in the project is going to inherit that context. You don't re-explain your role, your audience, etc. You just say, "I need a one-pager for the Gartner meeting." And Claude already knows what that means.
Now, why does Claude specifically reward this? Claude tends to follow complex system-level instructions very, very consistently across conversations without a lot of drift. So, when you set detailed operating rules like that in a Claude project, they tend to stick. And this connects back to the training approach. A model trained to follow principles rather than optimized for user satisfaction tends to be more disciplined about following the principles you set. Now, how do we know this is true? Well, Pixelate's 500-task comparison measured instruction compliance directly. Claude hit 94% exact compliance versus ChatGPT's 87%. And I realize that that's ironic because I spent time earlier in this video talking about how ChatGPT is optimized to give you an exact response to what you asked for. That's true, but I want to think about the work context with you. Often times at work, we are not super clearly specifying what we want done because we're not perfect, and we're used to giving humans somewhat vague assignments. And so, if you're optimizing for human-pleasing initial responses, you may ironically not be optimizing for task compliance because our work tasks sometimes get specified over a couple of turns of conversation. "Oh, that's not what I wanted." How many times have you told ChatGPT that? "Ah, that's not what I wanted." Or, "Yeah, that sounds good." And then you look at it. Or even more tempting, you say, "That looks good," and maybe you've clicked that like button, but you look at it again after your director rips it to shreds and you say, "Oh, it wasn't that good." I think that one of the most interesting questions in AI today is the difference between what a naive human perceives as likable and good in the moment, which rewards ChatGPT's reward function, and what a director or a leader with very high taste and deep domain expertise sees as high quality. Now, to give ChatGPT credit here, they are actually investing very heavily in solving that gap, and that is why they've published the GPT Eval, which measures how models do on solving real-world tasks as graded by experts. Still, if you're looking at how work changes today, think real carefully about how you frame project instructions, because if you frame project instructions with a little bit more context for Claude, you're going to get very good quality task adherence inside those projects with Claude. And I think that that's a nuance that often gets missed because people see projects and they think they're the exact same thing.
Principle number six: Claude can work on your computer. This is actually just a capability that ChatGPT doesn't have. In January of 2026, Anthropic launched Co-work, a desktop agent for macOS, and Windows support is getting added, and it's available right now to Claude Mac subscribers. Co-work doesn't chat about your files. It actually opens them and reads them and edits them and organizes them and executes multi-step tasks autonomously on your actual computer. And so, you can tell Co-work, "Go through the invoices in my downloads folder, extract vendor name, amount, and date, and create a summary spreadsheet, and please flag anything over X dollars," and it will just do it. For security, Claude Co-work only operates with folder-level permissions. So, it accesses what you authorize, and it shows you what it's doing in real time, so you can always stop it. This is where that chain of thought becomes really helpful. This reframes the AI category. ChatGPT is still positioned as a conversation partner, and Claude, with Co-work, is framed as a conversation partner plus a worker that handles file management and data wrangling that eats hours out of everyone's week. We all work with files all the time. Now, Co-work does too.
Principle seven: Know what you're giving up. So, if you're teaching someone about Claude, you've got to be honest about what ChatGPT does better, or there's no point. And so, I want to lay out what you're leaving behind so there are no surprises. Claude doesn't generate images. That's often the most disappointing thing. It also doesn't generate video through Sora. It doesn't really do real-time voice conversation. And you're giving up mathematical reasoning if that's important to you, and some emphasis on scientific knowledge as well. You're also giving up a degree of web research breadth that you get in ChatGPT if you're a heavy search-for-this user that you just have a hard time with in Claude. You're also losing some of the global persistent memory where ChatGPT still has an advantage, and you're losing the custom GPTs marketplace. You're also losing the app store where there are many, many companies building apps that appear in ChatGPT now. And so, I don't want to pretend that you're not losing things. ChatGPT offers things that Claude doesn't offer. And the best way to position someone who's new to Claude versus those gaps is to say, "Look, if you want to use ChatGPT for those things, you still can. But you can also start to learn a new tool that may give you a perspective on work or useful tools for work that you didn't know you had before." And that's really the best gift that you can give someone who is new to a new kind of AI because Claude really is a much, much less adopted tool than ChatGPT overall until it went to number one in the app store. And so, people are shaping their perceptions of AI through ChatGPT. This is really the first time for many people that Claude is even on the horizon. And so, I would just say, if you're talking with someone like that, or maybe if that's you, give Claude a shot without assuming it's just ChatGPT in a trench coat. Listen to some of these differences that I've described and take them seriously. Use them as a chance to, at best, level up your fluency and get super comfortable using multiple AI tools, which is one of the breakthrough skills of 2026. At worst, use them to understand why you like ChatGPT better after all, but now you have lots of reasons and can name the specifics. Or use them as a way, at worst, to understand how AI is different under the hood and get a sense of the complexity that makes this space so interesting. It's not complexity for the sake of being complicated. It's just meaningful differences that shape how you work. If you have a tool that's more likely to push back and say, "Hey, I'm not sure the way you frame this budget is optimal," you might get into the habit, as I have done, of inviting the tool to push back, and that shapes your behavior over time. Whereas, if you're in the habit of using a tool that agrees with you all the time, you are less likely to invite the tool to push back. So, these subtle differences, we're shaping the way we humans, in our brains, think about and model AI. They really add up and matter. And that's why I made this video, because I believe changing our habits is hard. And I want to make it as easy as possible for you, whether you're new to Claude or whether you know someone in your life who's new to Claude, to figure out that difference and communicate it effectively. I hope this has been helpful. Good luck with Claude, and we'll see how long it stays at number one in the app.