Anthropic is Making Claude Dumb. I Found The HIDDEN Fixes

本站添加: 2026-05-06

[00:00:00]If you've been using Claude recently, you've probably noticed that it's become dumber. And if you've thought that, you're not alone because there are thousands of people online that have also started to notice this. You can see, is it just me or has Claude gotten significantly worse over the past 30 days? Is it just me or is the new Claude Zoomer like it's lazy and neurotic and self-righteous? When you ask it to do stuff, it gets mad at you. Then I dug even deeper and I've seen reports noting that there's been a 67% drop in performance. And Anthropic themselves have even acknowledged that Claude code was nerfed. They said that by changing Claude Code's default reasoning from high to medium, they actually reduced user outputs. And even though these changes have taken place over the past couple of weeks, there are still new reports coming out that Claude is acting stupid. So the question is, how do you fix it? Well, over the past few weeks, I've completely changed my behavior around how I use Claude. And I've found there are some simple things that you can do which make Claude instantly smarter. I've been using Claude for over a year in my social media to help me grow to over a million followers in my company to help me scale a 30 person agency and software business. And the truth is the way that you need to use Claude now is materially different from how you had to use it a year ago. So in today's video I'm going to run through the eight fixes that I personally use to keep Claude sharp. And I can promise you if you do all of these things, you will extract the maximum value out of your Claude plan and you'll stop running into some of the issues that people are reporting online. Oh, and a quick thing before we dive in. I've actually built a system prompt which solves half the problems that we're going to cover in today's video. So, if you want it, click the link in the description below, sign up for our newsletter, and join the community where you'll be able to find the prompt. You can then drop this prompt into any Claude project to speed up what I'm about to show you in this video. But there are also some other tips and tricks today that I'm going to show you that are going to move the needle beyond that prompt. So, let's get into it. Firstly, the foundation. Before we get into the more complex stuff, you need to get the basics right before anything else. Now, this might sound dumb, but even I'm guilty of this sometimes. Make sure you're on the right model. Clearly, Opus 4.7 is the more advanced model. If you're doing difficult tasks, make sure that you're on Opus 4.7. And I've seen it sometimes, even on Claude Code, if you're set to Opus 4.7, it'll actually default back to Sonnet. and you'll be thinking, why is it making so many mistakes before you realize that it's actually on Sonnet?

[00:02:17]So, make sure you're on Opus 4.7, also with the 1 million context window if you can. The other thing that's really important, specifically when you're on Claude Code, is to make sure you're on high, extra high, or max effort. Max effort is going to get you the closest result to what other amazing coding models like GBT 5.5 Codeex can now produce. Frankly, low and medium just do not produce the result that I want, and they make too many errors. So, unless you're doing something simple like a habit tracking dashboard, you probably want to be on high, extra high, or max.

[00:02:44]It will take longer. It will use more tokens. But there is a clear trade-off.

[00:02:48]And in my opinion, you've got to stick to these three. Ideally, extra high or max. All right, moving on to fix number two. It's actually understanding which Claude feature does the thing that you want. I see a lot of people trying to stuff context into their prompts instead of using co-work and attach that to your memory system. And this is a big reason as to why Claude dump. If the chat doesn't have the context, it simply can't help you. So, I think it's really important that you get an understanding of what each part of Claude does. The first thing is the Claude chat. This is the typical interface that you see like this when you're interacting with Claude day-to-day. It's great for questions.

[00:03:18]It's great for one-offs. It's great for simple things. But remember, the memory inside the chat is still a black box.

[00:03:23]And even if you have your memory and your project set up, which you definitely should, by the way, it's still not going to be as effective as the next thing, which is using Claude Co-work. co-work. If you go into the cloud application and click the middle will actually allow you to work inside a project or a specific folder. And I've done videos on this. I'll link the most recent one in the description, you can actually set up your custom memory. So you can control the exact memory that you give Claude. So through working in a project folder, you can give documents, you can give memory without Claude having to read a massive prompt and then trying to give you an answer, which leads to a lot of hallucinations and it's one of the big reasons why Claude is dumb in its responses. So, make sure that you're fully aware of what you're using Claude for. Are you simply asking a one-off question or doing something simple, or do you need something with context where you would benefit from co-work plus a memory system? Now, clearly, if you're going to use the chat, you need to be utilizing Claude projects. You can do this both within co-works and within the chat. But because you'd likely have a folder with a local memory system in co-work, this is going to be more applicable to the chat. So if you click on projects and then I create a new project for example, I can actually give it instructions which helps dictate the amount of information that it knows and thus its responses. So if I click on AI edge for example, this has all the context of me, what I'm doing, what my goals are for AI edge and I have instructions for how I wanted to respond and how I wanted to update memory over time. And then every time I ask it a question, I will actually type it in a new chat because then it has the context that I've given it in the memory and instructions. And by opening a new chat, I'm not stuffing the context window by constantly responding in one chat. One simple thing that a lot of people screw up is they just use one big long chat and then they keep using it. You actually want to start a new chat every time and rely on your memory system to give Claude context that actually by default will make Claude a lot smarter.

[00:05:06]And then lastly, of course, you have Claude code. This is whenever you want to build something. So, if you want to build a dashboard, a habit tracker, a Python script, a more advanced workflow that relies on hosting, or even if you want to connect an external memory system, which I'd be doing in my personal OS build, which by the way is coming very soon. It's really, really cool. Basically, what I've done is I've used Claw Design to create the ultimate habit tracker, to-do list, calendar, finance system, allin-one with an external memory system. So, it remembers everything that I tell it and becomes smarter over time. The data I'm showing you is dummy data, so it's not real data. But this is the real dashboard that I've built and that's going to be coming very soon. So, make sure you subscribe. But Claude code is the only way that you can actually do that or another coding tool like Codeex. You can't build something like that within the chat because you need to link an external memory system for it to log all of that data. So, anything more advanced, I use Claude code. And I think you have a clear understanding now of what the other three can do. But make sure that you're very aware of what you're using Claude for. If you're trying to get something out of the chat that is clearly meant for Claude code, then it's going to seem dumb. So just understanding how Claude actually works and the different features is a major part of unlocking Claude's capabilities.

[00:06:08]All right, fix number three out of eight is using the system prompt that I'm going to give you in the description below. This is extremely high leverage.

[00:06:16]What you want to do is download the prompt. It's available for our Instagram community members. It's in the Google Drive under this video. Once you download it, go into Claude, go into instructions, and add it. What this is going to do is it's going to tell Claude how to behave. It's going to tell it to think deeply. It's going to tell it to respond if it doesn't know something.

[00:06:34]Because a lot of the time, what makes Claw dumb is that it'll try and dance around the fact that it's not 100% certain on something and it'll give you the wrong answer. If you explicitly tell it in its instructions to tell you when it doesn't know something, then it's going to give you a better response. So, I've actually constructed the prompt as a plug-and-play system that you can put into all of your projects across the chat and across co-work. So, it only tells you when it knows something and it uses its maximum effort to actually think about a problem. And importantly, especially if you're going to use the chat, it automatically updates memory every single time you give it new context because there's a line in the instruction that says, "Every single time I tell you a piece of novel information that you deem to be valuable, make sure to update the memory file." If you don't have this instruction, then Claude just doesn't take it on its own account to update the memory in the correct way. So, you'll find that the memory it stores about you is just super broad. It will have your age, some of your preferences, but then it will also have random stuff like you might be looking at buying a new camera and cord will remember that, oh, you film on a Sony A7, but it won't remember that you're going through a major company product pivot. Like, it just remembers the wrong stuff. That's why you need that line in the instruction.

[00:07:38]Once again, the full prompts available below, but the instruction setup is a really key part of making Claude less dumb. All right, now moving on to part three of the video. This is how to actually prompt Claude to get the best answers. This is also one of the biggest things in the entire video. Most people prompt Claude like they're searching Google. They'll just send a sentence and then they'll hope for the best. What you really need to do is tell Claude how to respond. So firstly, if it doesn't have the specific context through your memory system, you need to give it the context.

[00:08:04]My hack for this is simply using voice messages. So I'll send a voice message to Claude. I'll say, "Hey, I have this problem. This is what it is. I need you to give me the correct answer, but first ask me 10 clarifying questions so you understand the situation." Then I'll answer those 10 questions on voice prompts to save time. where you can use whisper on your computer and the desktop app even has a microphone capability now. And then I'll say, "Hey, give me five follow-up questions based on what I just told you." And when I notice the quality of the questions start to deteriorate, that's when I know that Claude has all the necessary context.

[00:08:34]And only then will I enter in my prompt.

[00:08:36]And in the prompt, I will say, "Okay, now your role is to act like an executive CEO or a YouTube script writer or whatever the prompt is. The audiences for X, Y, and Z. I want you to give me the full guide and I want you to format it like this and if you can actually provide it examples of the output that you want. I understand this is annoying.

[00:08:53]So for certain tasks you don't need to go out and fetch examples but if it was a YouTube script why not go into notebook LM get the transcript of the video and say look I thought this video was great. I want to structure it in a similar manner. So that is what makes a good prompt versus a bad prompt which is just write a YouTube script about AI. I promise you, you'll be bulletproof if you have a your memory system for broader context, and b you leverage clarifying questions more via voice prompting every single time you have a major question. Now, clearly, if I just want some advice on what's the best application for auto DMs on Instagram, then it doesn't matter so much. But if I want to make a big business strategic decision, then clearly it's going to need the context in order to give me an answer. And a lot of the time, people will just go with the answer that Claude gives it. Whereas if you just gave a little bit more context, it actually would arrive at a completely different answer. And that's the scary thing about AI. Like I have to voice prompt to get it to ask me questions now or I just have trust issues, especially with bigger decisions that it won't give me the right feedback. Now obviously as a CEO, as a strategist, as a controller of these AIs, you still don't just want to be listening to whatever Claude tells you. But clearly if you can leverage its feedback, get it to be critical, and that's something that I also have in the system prompt down below, then you can actually have a different lens on things which can help you arrive at different conclusions. I still think Claude is an amazing brainstorming partner. Like I get so many great business ideas just through the act of reflecting with Claude. It's almost like having a psychologist. I don't treat the outputs as final, but I do treat it as a brainstorming partner where I'm fully cognizant of the fact that it needs the right context for it to give me the best possible responses. Oh, by the way, if you do all of this stuff, this just came up on my feed when I was researching.

[00:10:31]And you still run into rate limit issues, I have a video that I just put out about that. We'll also link it below. And that video is all about how you cannot run into your clawed rate limits. So, yes, you can implement all of this stuff. Yes, some of it will help with rate limits, but you also need to be efficient with how you use the application. I did a full dedicated video on that cuz I am getting lots of questions about that. All right, moving on to fix number five. This one is underrated and it's actually in Claude's official documents. They're literally telling you this is the best way to prompt engineer. They suggest using XML structuring. XML is actually how Claude reads responses. So XML is essentially formatting specific parts of a prompt like this. So, instructions, context, examples, input, etc. If you include XML tags into your prompt, you're going to automatically get a better response from Claude. So, I'm literally asking Claude itself now. I'm saying, I'm filming a video, give an example of how I can use XML tags in a prompt, and it will literally code you or in this case, me right now. Basically, XML is this, my voice, direct, no fluff, email, output format, etc. So, every element of the prompt and you can actually get it to reverse engineer this. You could say something like, "Hey, I'm about to ask you about business strategy. Give me a list of XML tags that will help you answer the question better." But just using these tags really helps Claude and how it understands. So, you can actually save processing power and make it smarter because it doesn't have to interpret what you're saying. If you don't do this, it has to interpret a slab of text that you send it and convert it into XML anyway. So, if you can include XML, then do it. Obviously, on voice prompting, it's harder. This only really applies if you're writing, but this uh can actually really help Claude out. All right, fix number six of eight. This is another one that's really important, actually using Claude tools.

[00:12:08]So, for example, if you're researching, don't just type into the chat, oh, research X, Y, and Z. Actually click on research mode because if Claude gives you a dumb response and it's not on research mode, then it's actually not using its full capability. Yes, it'll take longer, but if you're researching, make sure this is ticked on. If you're web searching, make sure this is ticked on. Sometimes you actually might have this ticked off. you'll ask it a question relating to a recent event and it won't have the most necessary information because it doesn't have web search enabled. So that's very important especially if you're trying to reference data. And obviously you can go to the next level. I've done a video on this with using Claude skills when you want a specific thing done in a specific manner which obviously makes Claude way smarter because you're not having to constantly tell it what to do. It already knows.

[00:12:49]And obviously adding necessary files, photos, etc. will also help with context. But just clicking this little plus button here, the connectors button is a major one because people don't use the tools for their prompts on the platform. A lot of this stuff might seem basic, but people, you know, download the app and they just start chatting and they actually don't think about this stuff. So hopefully this video shines a light on some of the basic things which can make Claude a lot smarter. All right, now we've spoken about the basic stuff. Let's talk about the power moves.

[00:13:14]These are the operator level tactics that you can use which compound into much better results. So the first thing is, and I discussed this before briefly, make sure you actually start new chats and rely on an external memory system because by having too much context, you actually confuse Claude. So often it hallucinates and it will cherry pick some of the data and not respond to other parts of the data cuz it simply has too much to look at and it'll also burn through your tokens. So make sure instead of trying to repair a blow to chat, just start a new chat. It honestly makes Claude so much smarter. Like using new chats instead of old chats is one of the highest leverage hacks that you can use. Another big one is data pruning. So these are three rules I live by when I use Claude. Firstly, in projects, make sure you're pruning files. So if you haven't referenced a file in a few weeks, remove it because it's doing work every single time that it responds to you by having to read these files. So make sure the information is relevant and you don't let it bloat. Same with memory. Make sure it doesn't get too bloated. If memor is outdated, go through manually either in your local folder or in the project memory system and actually fix this. This is what's known as context rot. Giving AI a lot of information sounds great, but it can become counterintuitive. So, to summarize the three rules, firstly, make sure you are pruning your project context. Secondly, make sure you keep your memory entries sharp and you prune over time. And thirdly, make sure to start a new chat if a chat is going sideways rather than trying to compress the chat. Like, you'll notice if Claude autocompresses the messages, the chat's probably too long and you should probably start a new one anyway. And then my last power move, and this one is a big one. Get Claude or another LLM to review its own work. So once it gives you an output, ask it to fact check.

[00:14:50]Sometimes I'll get some help with researching for a tweet. It'll give me a bunch of references and it'll even use research mode. And then I'll say something like, "Is all of this real?

[00:14:59]Can you fact check?" And then it will fact check itself and be like, "Oh, actually two weren't real or two, you know, I grabbed the wrong headline."

[00:15:04]Like you'll be shocked or maybe you're not shocked at how much Claude will actually give you the wrong information off the first bit of information. So especially for something important, get it to fact check itself. Like I'll even have it help me with, you know, retention on hooks for videos and then I'll ask it to critically evaluate the hook and then the second try will be much better. And to make this even more powerful, this is something I do on a daily basis. Actually stack two LLM. So I'll have chat GBT 5.5 and Claude open in front of me and I'll use both. I think 5.5 right now is actually smarter.

[00:15:33]So for the strategy part or the research part, I'll use 5.5. For the writing part where I feel like I have more control and more creativity, I'll use Claude.

[00:15:40]And I'll just model stack both together.

[00:15:42]If you're just on Claude though, you can literally just ask it to review its own work. It does most of the job. So, these are my eight ways to make Claude go from dumb to smart. Remember, the prompt that you can copy and paste into your instructions is down below. I think that alone is a gamecher, but hopefully everything I've said in today's video will help you get more out of your Claude subscription. Thank you very much for your time today. I will see you in the next video. As always, peace