Vollständiges Transkript anzeigen (4.596 Wörter)
If you've ever wanted to move from ChatGPT to Claude, or the other way around, what comes with you? And what do you lose? Oftentimes, your custom GPTs, your projects, your memory, your instructions, all of that stays behind. And that's one of the big reasons people never switch between the two products. I built everything inside of one app for a year before I figured this out. Now I use both and moving a task between them takes a couple of minutes. The trick isn't either app. It's actually a folder on your computer that both of them can read and write to. So, let me show you how it's done.
So, the question here is why this video and why now. As you may be aware, only a week ago, both Claude and ChatGPT released new models. So in Claude, it was Claude Fable 5.1 and in ChatGPT, it was GPT-6 Astra, both of which are very good at different types of tasks. Oftentimes when this happens, when a new model is released, a lot of people debate which one's best, if they should switch to another one, is it suitable for the tasks they care about? All these types of conversations. And honestly, I think it's a bit of a waste of time. It's exciting and interesting to read about, but when you think about AI and you want it to be useful for your business and your team, you don't necessarily need to get caught up with what everybody else is talking about. You just need to set up your infrastructure in a way where it doesn't matter which model is best, you can switch to it easily because you set up your process in a way that you can switch easily. Now, a big caveat to start here is the only way you can set this up is if you're using ChatGPT Work or Claude Cowork. These are two different desktop agents that enable this ability to easily switch between different products. So, it's the first thing you need to do is get access to the desktop agent. Now, once you've gotten access to a desktop agent, the key thing here that sits at the core of all of this is a folder on your computer. And in the end, after you've set everything up, you can easily point ChatGPT Work or Claude Cowork at that folder and swap them out for a given task without having to worry about being stuck in a given product, because oftentimes, if you don't set this up, you're stuck with a given product because it has all of your memories, your instructions, your context, all that stuff that's really hard to port over to another product.
And it's important to know that there's a lot of nuance when you're evaluating these different models. Because when you hear people talk about GPT-6 Astra and Claude Fable 5.1, the pricing per token is the same, so it's $10 in and $50 out. So when you see different providers, they often price their models per token, so that's basically part of a word. But more recently, both OpenAI and Anthropic, the companies behind ChatGPT and Claude, they've actually started talking more about per-task cost, not per-token cost. And the reason is is that even if the tokens are the same, GPT-6 Astra time and time again is usually between eight and 9x cheaper than Claude Fable 5.1. So if that matters, great, now we have an understanding that this model can be cheaper in certain circumstances. But that doesn't necessarily mean you should always use this model, GPT-6 Astra, because Claude Fable 5.1 is very good at a variety of other tasks that GPT isn't.
Quick pause, if you're enjoying this, you're going to enjoy two other things. First off, Blow's 30-day AI insights series, completely free. You'll get 30 insights in your inbox, how you can apply AI to business and your work. The second thing is if you'd like to work with me, Blow's series of offerings to see if there's a good fit between the two of us. Now, let's go back into the video.
And right now that's the case. But in a month when GPT-7 or Fable 6 comes out, the what they're good at will vary, so it is important to keep up with which model is good at certain tasks. But the thing we want to make sure we set up in the beginning is a foundation that we can switch between providers and not be stuck with any given one. So right now, most of you, when you try to switch from ChatGPT to Claude, or vice versa, some things will come and some things won't. So the things that'll be stuck in that product are going to be custom GPTs, if you're going from ChatGPT to Claude, connected accounts, that applies to both providers, and the wrapper around the model. So oftentimes, people don't talk about this, but it's important. So in Claude Cowork and ChatGPT Work, there is a wrapper, or a scaffolding, around the model. And that scaffolding tends to be instructions, tools, memory, all things that you can't see in the background that allows this model to work on really long horizon or hard tasks. So those are some of the things that you can't port over to the other providers. Some things that will come over, but are a little harder to move over if you've not set them up correctly, are projects. So both in GPT and Claude, there are GPT projects and Claude projects. All a project is are some instructions and associated context files. So you can, you can copy and paste these out of one provider and paste them into another. I'd actually recommend not doing that and/or using projects in general. So if you have a desktop agent or a desktop app like Claude Cowork or ChatGPT Work, you should base all of your projects on a folder in your computer because that's what the equivalent is on your computer. A project is simply a folder that has instructions, which we'll talk about in a second, as well as context files in that subfolder that you can reference. And this is something that you can do in any machine and you can also share these. So if you create a folder and you share it and sync it locally to OneDrive, Google Drive, SharePoint, you can share this with other people as well, so you're not secluded to just having one project in your computer. So this is partially movable from ChatGPT to Claude and vice versa, but there's a lot of work of copying and pasting. But if you set it up like I said previously, you won't have to worry about that. And the final thing is that both ChatGPT and Claude in their settings allow you to import some things from the other provider. ChatGPT is much better at this, so they actually port over way more from Claude than Claude does from ChatGPT. But I'll show you actually how they're both set up, but it's very straightforward, you just go to settings, import, etc, etc. And for ChatGPT, it moves over a lot of things, like instructions, memory, chats, etc, and Claude, it's usually only memory, but note that there are settings you can import back and forth between the providers.
Now I've referenced this folder, the importance of having a folder on your computer per task, so that's key. You want to make sure that each folder in your computer is per task and not like a generalized folder. If it's a folder that's covering an entire topic, like a client or a project, that's usually not great. It needs to be divided up by a certain task. And this is the crux of everything. If you set this up effectively, this folder or multiple folders, you'll be able to port between different products of ChatGPT and Claude. So there's a few things we're going to walk through here. So the first thing inside these folders is going to be instructions. So you're going to actually have two files here. So a Claude.md and a agents.md, and you can simply have the AI create this for you. And these are in the instructions that the AI looks at every single time it opens this folder and does this task inside this folder. It's important to make sure that these files are minimal in size, since the AI is looking at it every single time you're interacting with this folder. I'd recommend keeping it less than probably 100 lines. And if you're not aware, .md just means markdown. That's all that's referring to and it's a type of file format, similar to a .docx file or a .xlsx for Excel or something like that. So that's the first thing, this is going to be mandatory for all folders. You're always going to have this. The next thing are skills. So these are optional, but they're extremely useful between different products because skills themselves are open standards. So I can create a skill in Claude, I can download the skill and upload it in ChatGPT, and it'll work as just as well because it's an open standard. I can port them back and forth between providers. So in this case, if we have a series of skills related to this folder, we want the AI to make sure they're accessible to ChatGPT Work and Claude Cowork, and you can simply ask the AI to do that for you. So they, "Hey, make sure this skill is available in both products." It'll then save this skill in both locations and make sure it's available for both. Our next one here is again, completely optional, which is a lessons.md file. Again, this is markdown, just a file format. And the name here doesn't really matter. This could be preferences, this could be lessons, this could be corrections, memory. You can name it whatever you want. But the premise of this is for the AI to externalize its memory and lessons learned in regards to this task. So anytime that I give the AI corrections or preferences for this task, it should take those and append them and add them to this file. And in the instructions, say that I should always reference this lessons-learn.md file every time I interact with the user to make sure I understand their preferences that I've learned over time. Now, the key with this one again is to make sure that it's minimal in size. So every lesson that it learns, it should be just one line with a date and making sure that one line is extremely information-dense. Because if the AI is referencing this every single time, it needs to be short. Over time, if you have lessons in here that are repeated over, over, and over again, you can simply just bake that lesson into a skill or instruction and then remove it from here, because you've already baked it into the instructions or skills. But that's only for the repeated ones. Now, our next two are subfolders inside this parent folder. So, depending on the task, oftentimes, you'll want the AI to process an input file, give you an output, then archive that input. So this could be the AI writing reports or processing something that you do on a daily or weekly basis, usually something where it's recurring. This is the common structure I build for people that I work with, where we have an inputs folder, an outputs folder, and an archive folder. So, what the AI is going to do is it's always going to be looking at the inputs folder inside of this parent folder. You, the user, will drop in files here. When you drop in files here, the AI will look at them, it'll process them, give you the output, so that could be updating an Excel sheet, updating a dashboard, or a report, etc. Once it's finished with that, and you've approved that the output looks good, the AI will then move the inputs information from the inputs folder into the archive folder, so then the inputs is always clean and ready to be processed. So this is a way for the AI to clean up its own work as it works through the task at hand. And the other folder are simply your standards. So this could be examples for a given process. If there's a certain way you want an AI to write, you could have reference examples for summaries, and it can look at those to see exactly your tone, your structure, etc. And that's where you'd put these examples. And again, for this, you want to make sure the AI in the instructions references explicitly that it needs to look either at these subfolders for the inputs, outputs, and archives, or look at this examples folder for a given task for inspiration, so I can match your specific tone. Now, we've walked through all of these. Now, why is this so important? Because what we've done here, at a high level, is we've given the AI instructions, we've given it skills and memory, as well as some architecture or process around that specific task. All of this is externalized from the product itself. It's not inside of Claude Cowork, it's not inside of ChatGPT Work. It's on your computer, which means you can switch back and forth between these tools using different models, and it can still do the task at hand because it's externalized from the product and it's sitting in this folder. And that's really the key to this entire video. Now, the next thing here is importing things from one product to the other. Like I mentioned, both providers have settings that you can use, and I'll actually show you how these are utilized. So, we'll start with ChatGPT, since it's better than Claude right now, of taking stuff from Claude. So, if you're inside of ChatGPT here, you can go to your profile here, go to your profile, you'll go to settings, under settings you want to go to import, and in here, in import, you'll have the option to import from Claude Code, from Cursor, and usually there's an option for Claude Cowork as well. All you have to do is push import. Once you've done that, you select which ones you want to import over, and you select import, and then it imports those back for you. After you've imported these over, you can see some of the stuff that's imported from my Claude Code and Cowork over time. You can then set this up to sync, so if you turn this dial on here, it's automatically always going to sync anything you add to Claude that's not in ChatGPT. So, that's going to be your chats, your instructions, your history, all that stuff. In here, you can see where it says the specific content that's being synced. If you go to customize, you can see exactly what it's pulling over. So, it's pulling over chats, instructions, settings, skills, plugins, MCP servers, and commands. I if all that's enabled, when you do import, it's going to pull all that over for you from Claude into ChatGPT. So, that's the ChatGPT setup, how we're moving from one place to the next. For Claude, this is much different. So, in here, you're going to go to settings, when you click on here, you'll go to settings. After you're in settings, you'll then go to memory, and under memory, you'll have an option here that says start import. When you select this, all Claude is going to do is give you a prompt. So, this here is a prompt that you're going to copy. Once you've copied this prompt, you're going to go into ChatGPT, you're going to paste it into ChatGPT, and it's going to then give you back some some memory files, you're going to take that memory text, you're going to copy and paste that memory text into this this section here. Once you've pasted in here, you add it to memory, and then Claude's going to save that into its memory. So, you can see Claude is solely focused on memory, when ChatGPT is focused on chat, instructions, all types of things. So, those are the two different ways we can actually utilize what the product providers are giving us to port one thing from one place to the other.
So now, let's assume that you've gone through the process and you set up your folders and you're ready to port back and forth. The question we want to ask ourselves is when should we use different models for different tasks? Well, for me, there's usually two different questions I ask myself depending on the situation. So, let's say that I'm looking at creating a board report continuously with one of the models, and I want to test this with both the newest models from each provider. Well, I'm going to ask myself first, which does this better first off? So, which is better at doing this? After I've understood the quality of the outputs for this task, I then want to ask myself, which one burns fewer tokens for that task? So, is it 10x the price for Claude to do this that ChatGPT would do the same? If so, maybe I can fix and improve the prompt and context in ChatGPT to do that task if cost matters to me. But in addition to cost, it also could be time. So, maybe Claude takes 10 minutes to do the task, but GPT takes two. If that's the case, and that's also going to factor into how I choose which model is going to do this specific task for me. But again, this varies by the task. If the task is really important, and even if Claude is more expensive and more time-consuming, but it does way better, I'm likely still going to go with that if it's a high-stakes task or situation. One of my favorite things to do, especially for high-stakes situations where there's financial, legal, reputational damage that could occur for this task, I like to actually pin the AIs against each other. And what we can do is we can have one AI do the work and the other AI review the work. And the way that we would do this is we have a simple prompt here, assuming that I've done the work, let's say inside of Claude. I'm going to paste this prompt into Claude and all it's going to do is Claude is going to give me out basically code that I can copy and paste into ChatGPT for it to review the work, because ChatGPT has access to all the files and all the information in that folder. And all Claude is going to do is tell ChatGPT what the original goal was, who the audience was suited for for this output, as well as any rules that it followed along the process. So, this summary output, we're going to get into a code block, which basically means it's going to be text that we can easily copy and paste back and forth between the tools. We're then going to open up ChatGPT, in that same folder that has all the context and all the files that the AIs have been working with, and we're going to paste in this prompt with the output from Claude. So, we're simply telling the AI, now I want you to review the work that another one did. Here's the summary, which is the code block we pasted in, and I want you to read all the files that are inside this folder that it worked with, and then give me back what the other AI may have gotten wrong, what it could have missed, or what you would potentially change to improve this. And the reason we're using ChatGPT to review Claude or vice versa is that if we asked the AI that did the work to review its own work, it's likely biased towards that work. So, oftentimes when you write an email for somebody and you review the email that you've drafted before you send it, when you read it, you're likely reading what you intended to say, but maybe missed some of the elements to that. But if you have fresh eyes from a different model, it's going to dramatically improve the review process. So, we've talked a little bit about choosing models for a given task, but when do we actually switch models for a task that's already happening? So, if I have Claude doing something for me for a while, when do I know it's time to actually switch from Claude to ChatGPT? Well, there's three scenarios this really falls into. One is you've hit your usage limit. So, every single week, we all get a certain allotted amount of usage for a given product. Either you're paying via API credits or you're getting usage for a subscription. Let's say that you hit the limit, but you need to keep doing that work and you need to keep working on that task. Well, in that situation, you can simply switch from Claude to ChatGPT or vice versa if you've hit the limit in one of the other. The other one is that you have a given model that keeps failing on a given task. So, you've tried it in a model and it keeps failing in the same way over and over and over. And you've already tried to improve the prompt and the contexts, but it keeps still failing. In that case, you could just point a new model at it from a different provider and see how it performs. And then the final option here, the third one, is when a new model comes out. So, let's say that GPT-7 comes out or Claude 6 comes out. I want to try that out on this task and see how it performs. So, I'm going to point that at this specific folder and I'm going to try it out and see how it goes. Those are the three scenarios in which I would actually switch during an activity that I've already chosen a model for previously. Now, one important thing I'd call out here, when it comes to testing, is making sure that when you test these different providers against each other, you test in an effective way. Oftentimes, when I see people test, they make tons of mistakes. One of the biggest mistakes is when they test it, they share their previous output for this task with the AI as well as the inputs, so the AI cheats. It grabs the output you gave it, gives it to you and says, "Here here I'm done." And you automatically assume that it did a good job, but really it cheated. So, to mitigate that cheating, we're going to have two different test folders. So, we're going to have the Claude test folder, and we'll have a ChatGPT test folder. These two different models are going to be looking at these separate folders for the same task. In that folder, you're only going to have the inputs, not the previously done output. But we do want to have the output, we just want to have it separated so the AI can't see it. And the reason we want to have this separated is because we want to compare the AI's performance against ours. Ideally, this output is something you've done manually, if you want the AI to automate or augment that process for you. So, assuming we just have the inputs in here and the output is secluded, we then can thoroughly test the AIs against each other to see how they perform. And the reason we have two test folders is we don't want the AIs colliding with each other as they're working through the process. So, again, we have the inputs in the folder, the output separated. Once they test, we then compare the outputs from each of the models, as well as compare that against our human output that we did here, and see whichever one is the best, once we've decided, we then note down the winner and use that model for that task going forward. Now let's do a quick recap of the most important things we talked about here. So, the first thing, is that the core foundation element of being able to switch from one product to the other, is having a folder that's locally on your computer that can be synced to the cloud, but that folder is the core component. Because these different AIs can look at that folder to do the same task, because inside that folder, we've structured in a way that's easy for them to do so. When structuring it, inside the folder, there's two mandatory files that need to always be in there. That's going to be a cloud.md file and an agents.md file. Both of these are relevant for ChatGPT and Claude. So, cloud.md is for Claude, chatgpt is for agents.md. Now, one of the benefits of having two AIs that can access these folders and you can swap between them easily, is that for high stakes tasks, you can actually pin them against each other. You can have one AI do the work and the other one review its work. Our fourth and final point for the recap, is when you're trying to decide when to actively switch from one product to the other, there's usually three scenarios that we fall into. The first one is that you've hit your usage limits. So, say that you're working on an activity, and you've hit a cap for the week, and you got to keep working on that so you can finish it. So, you can switch from one product to the other, Claude to ChatGPT, or vice versa. The second one that we've ran into, is failures. So, maybe you have Claude doing a task, but it keeps failing in a certain way, and you've already tried to improve the prompt in the context. If so, it's actually time to try to ChatGPT on that task and see how it does, and see if it can out perform Claude on that because it keeps failing. And then the final one is new models. So, when a new model gets released, you want to test it against the old model to see how it performs from a different provider. So, if Fable 6 comes out and you want to compare that against GPT-6, you can point it at that folder and see how it goes. And that's it. So as a reminder, two quick things. First off, blows a 30-day AI insights series, completely free. You'll get 30 insights in your inbox, how you can apply AI to business and your work. The second thing is if you'd like to work with me, Blow's series of offerings to see if there's a good fit between the two of us. And if you've enjoyed this video, the YouTube gods know that you're likely going to love this video. So it see if they're right, go ahead and give it a click. And I'll see you next time, Internet.
Video
Video
