Paste This Into Claude, Never Get A Generic Response Again

Today we're talking about how to stop getting generic responses from Claude. And what most people think when they hear that is it's just about making Claude sound like them. But that is just the tip of the iceberg. There are five layers of context and voice is just the most obvious one. And once you set up all five of these layers, you'll stop having to reexlain yourself and start getting higher quality outputs on the first try. So in this video, we're walking through what each layer is, how to set it up, and why it matters. And at the end, I'll cover the mistake that 99% of people make that quietly destroys whatever context they've provided. So layer number one is voice context. What makes AI sound like you? This is what everyone pictures when they hear generic responses. The complaint is always something like AI sounds nothing like me. But it doesn't have to be like that. So how do we fix that and make it so it writes just like you. First you need to understand why you get generic responses so you can fix it. And all five layers of the context iceberg build on this foundational understanding. When you prompt AI with no context, it defaults to the statistical middle, the average of its training data. So by definition, AI will give you an average or a generic response. But by providing context, we're able to tell AI where on the spectrum we want the response to come from. My favorite mental model for this is thinking about giving someone directions. If you just tell them the city you're in, they'll get directionally close to you, but there is a near 0% chance they'll go to the exact location that you want them to. But if you tell them the street to go to, they'll get closer. And if you give them the exact address, they'll get even closer. It's really the same with context. We need to provide it in the right direction so it starts to triangulate towards the right output we want. On screen, you'll see how this works. As I provide more context, the chances of it providing a response in your target style go up significantly. Okay, fundamentally, we get it. But how do we provide the right context in the right format to get a response that sounds just like you? Well, a lot like the address example, we start general and then we get more and more precise. So, first, your general writing style. This is your vibe and it's pretty simple. Grab all of your writing and upload it to Claude and use this prompt to create a writing style file. This is foundational, but it's not where a lot of the juice is. You need to think about the different writing domains because you don't write the same way depending on who you're speaking to. For example, your emails don't sound like an internal team update. So to handle those different scenarios, create claude skills for each writing domain. So you can create one called draft emails and then another called internal updates. And then for those skills pull directly from anthropic docs. We want to make sure that the description of those skills say what the skill does and when to use it. This is how Claude will automatically determine when to use a specific skill without you having to explicitly call it. From there, once you create that, use them, test them out, and see what you like and what you don't like. So, what you might realize is, let's say draft email, it's not specific enough because even when you're writing emails, depending on who you're responding to, it might be an entirely different style. So, at that point, you may want to make a skill called draft customer support email or a draft prospect follow-up email skill. Then once you've partitioned these domains and they'll get more and more specific as you use them, you need to establish hard rules. The simplest way to picture this is the opener of an email. If you literally never use the word deer name, then AI should never use that in its opener. To fix this, you would want to add explicit gotchas in the /draft email skill that says don't use deer as an opener. The only acceptable openers are hey-name or name, this is where we want to be prescriptive so it knows exactly what to do so you can stop repeating yourself. So, that's how you get Claude to write in your writing style. But, like I mentioned, this is just the tip of the iceberg. The next four layers are what really interests me. Layer two is your knowledge context. This is what Claude can reason from. Layer one made Claude sound like you, but it didn't actually touch the information that Claude knows. Who cares if it sounds like you if the analysis is garbage or you can't understand it at all. So, by default, Claude is trained on all of the data out there. And it's functional for people who are either a rocket science or a fifth grader and everyone in between. And a lot like layer 1, it merely biases towards the statistical average. So there are three ways we can enhance the depth and the specificity of Claude's responses. The first is that the best place to start is to ingest curated information directly into your system. You can use this prompt which will create a directory of content within your system. So this will make a /nowledge folder, a / raw, and a /wiki. And then on screen I have a graphic of what these folders do and why we do this. But one step further that we want to do is we want to add context as to why that specific piece of information is valuable. If you're adding a book about sales, give some context as to what you like about it. This will help the system index on that source for that specific reason. On the left, you can see a response for when I ask, "How can I improve my landing page?" It gives me a generic response. And then on the right, after ingesting curated data, it gives me a more thorough response from experts that I really trust. The second thing to do here is consider limiting the context Claude has access to. This may seem counterintuitive, but this is a massive unlock for businesses. So, let's say you have an internal resource about how to complete specific tasks at your company. One of the biggest challenges of using AI is it will pull from all of its training data, not just what you shared. So, to limit that, I create a skill called /kbans answer. And what it does is it only looks at my knowledge base to provide an answer. This is a way of cutting off all the other resources. So, it just focuses on the information I provided. And so, here's how you can recreate the skill yourself. And you can see it in action here when I ask about my company's newsletter writing process. It's a clear step-by-step process and it references my exact team members by name. The third thing we can do here to personalize the responses we're getting from Claude is be explicit with the depth at which Claude responds to you. So you can use a phrase like explain this concept to me like I'm 15, which will simplify concept but give you enough details to start to really understand it. Or you could say provide a level two analysis which will surface non-obvious analysis of the situation. Here's a before and after of the same question. How does the social media algorithm work? On the left, it gives me a pretty simple response. And on the right, I ask it to make a level two analysis, and it gives a more thoughtful answer. Before we get to layer three, which will cover how you can stop having Claude agree with everything you say. There's a new product that will help with everything that I'm covering in this video. Which brings us to today's video sponsor, GenSpark, and their brand new hardware product, Second Brain Note. When they sent me this, I was pretty blown away. It's an audio recording device that attaches to the back of your phone [music] and when you press this button, it starts recording and then it automatically ingests into Second Brain, GenSpark's AI platform. Something I say all the time. In the AI world, your data is your remote. And the Second Brain Note is able to seamlessly capture one of the most critical data sets, what you're actually saying out loud. And it does this in three ways. The first is that it's really slim. So, here you can see it compared to a credit card and can seamlessly fit on the back of my phone using the case it comes with. Funny story with this. I actually just raced an Iron Man and this thing sat in my bike's container while I left myself notes the whole ride. It was my way to pass the time on a 7-hour bike ride. The second is the battery will last for over 35 hours. So, you really don't need to worry about charging it. And the third is that Second Brain software automatically syncs with your recordings. So, that means the longer that you use this, the more meetings it records and the smarter it gets about your specific work and it becomes your real differentiator. And I can connect it directly to Gmail, Slack, and Notion and more. Here you can see me in the second brain chat where I ask, "What did I discuss with my client on Wednesday?" And it gives me an overview of that in-person conversation, which is a good start. But when you pair that with GenSpark Super Agent, you can create some fire workflows, like sending my team an overview of that same meeting directly in Slack. It honestly feels like magic. And I love that GenSpark is getting into the hardware game in a way that seamlessly fits in a piece of hardware that I already carry on myself, which is my phone. So, this device is a limited first release, and I work with GenSpark to get my audience early access and a 10% discount. It's $179 total and comes with a charger and the MagSafe wallet that I was showing in this video. It's a pretty limited run, so click the first link below to get your hands on it today. Layer three, collaborative context. We want to make Claude a thought partner, not a mockingb bird. At this point, we've established how you can make AI write and sound like you while providing non-obvious answers and analysis. But the next problem is something that I know everyone watching this has experienced. Claude or any tool out there really just agrees with everything you say and this is the yes man phenomenon. I found it so interesting to understand why this actually I'll get to covering how to solve this problem. But first, why does this actually happen? AI initially got trained by having humans grade the outputs of the responses and people tend to give higher ratings to answers that agree with them and sound reassuring even when a push back would have been more correct. So agreeing is just AI's easiest way to get a thumbs up. So, it defaults to agreeing with you because that's how it was trained. And that's also why AI will often cave and flip its answer the moment you give it any push back. Let's say in the real world, ask any artist or athlete who has fallen off. It usually traces back to a circle of people around them that would just say yes. It's just all noise, no signal. So, this layer is about building the collaborator and the thought partner you need. The simplest way to do this is update your Claude MD file to have the following text. This will constantly remind Claude to not just agree with everything you say. On the left, I bring in a Claude video idea that it tells me is a brilliant idea. But on the right, after updating the Claude MD, it doesn't necessarily agree with this idea. So, that's the simplest way, but the second is getting a second opinion from another AI model. My favorite way to do this is install the Codeex plugin directly into Clawude Code so that you can redirect any analysis to that model to evaluate it. This helps remove any of the AI model confirmation bias where AI is more likely to approve an output if that specific model made it. You can also update your skills to automatically do this by using this prompt. This is something that I do just to have a second opinion in the room. The third way to do this is create a thought partner, which is about creating your own ask the board skill. So, I have a video on my channel which has over 100,000 views where people loved this specific skill. What it does is it's creating a group of AI agents that are your adviserss that will poke holes and critically analyze whatever you're working on. To do this, first clone experts who you want on your board. For example, if you want me on your board, you could take all of my YouTube transcripts and create an Awesome Archesy AI clone. If you do that, let me actually know in the comments if you did it, and I'm curious how my clone responds. But anyway, then you run this prompt to create a board based on those people that you created that will critically review your output. I do this for either key decisions or evaluating a final product to see if it's good. Now, if that sounds like a lot of work, you can use build partner.ai, which I created for this exact reason. So, you can download the plugin and then just ask expert advice and that'll analyze your problem, your concept, whatever you're working on, pulling from direct industry experts. Now, this layer, the system is truly becoming yours and responses are no longer generic. But there is one problem underneath all of that that most people never even notice. Layer number four, strategic context. become goal oriented, not task oriented. AI is designed to complete a task as quick as possible, but sometimes that might not be the best strategy for what you're working on. When I was a CEO of my last startup, when closing artists like Ed Sheeran, Tate McCrae, and Chancellor Rapper, one word came up constantly. Strategic. Music is a small vibes driven business. So, one wrong move and an artist, a manager, or label may never work with you again. It's it's really that cutthroat. And as a result, you couldn't just try and complete the task at hand. say following up with Ed Sheeran's management because that wasn't always the right strategic move given the bigger picture. So I call this goal-based execution. How do we think about tasks with the broader goal in mind? To do this, first give the goal that's above the task, not just the task. So AI only sees the task you hand it. So if you say follow up with Ed's management, it is an isolated task. So it'll optimize for completing that specific task. But that task actually serves a bigger goal. Get Ed to sign with us, which serves an even bigger goal. build a company that artists can trust. When an AI has visibility into your broader task, it gives it the opportunity to catch these issues before they actually present themselves. This is the context that changes a generic AI response to a response that's personalized based on your broader goals. So, within your CloudMD file, you can have a line that says, "Before starting any new project or building any automation, make sure there is a clear linkage to my broader goals and ensure any safeguards are put in place so that individual tasks don't conflict with these goals." And to establish what those goals are, you have to provide them. So say, "Enter interview me to land on my 1-year, 5year, and 10ear goals to help inform anything I'm working on." And the time gaps you can just change depending on if you want a shorter or a longer time horizon. The second thing that you want to do is you want to gate the one-way doors. I want to shout out Jeff Bezos for this concept. Every action and decision you make is either a one-way or a two-way door. A two-way door is something you could change if you have to. For example, what time your weekly meeting with a team member is. These are decisions that are very easily changed. It's a two-way door. You can walk back through it. A one-way door is something you can't reverse. For example, emailing Ed Sheeran's management team. Once you send that, the email is sent. That is a one-way door. Now, that's on a task level, but you can think about it on a broader goal level as well. For example, quitting your job. That is a one-way door. So, a way I like to enforce thinking about things in one and two-way doors is by creating a specific plan automation skill. I've spoken about this on my channel, how important it is to thoroughly plan and think through anything before you build. And by creating this skill, you can have it biased towards two-way doors. And if there are any one-way doors, add a human verification element. Here's a prompt that you can use to create this skill within your system. This mental framework of one and two-way doors is so important to actually build quicker as well. So, if you ever find yourself stuck on a decision, ask yourself, is this a one or two-way door? If it is a two-way door, make the decision quickly to your best ability. If it is a one-way door, sit down and think about it. Now, before we get to the last layer, which is the mistake that 99% of people make, I want to welcome everyone new to the channel. What is going on, guys? But for the rest of you that have been watching all the time, this is our anti-slap agreement. I make these videos entirely for humans and optimize it for you, not for AI scraping robots. And so, all that I ask part of this agreement is to subscribe to this channel so I can keep making videos like this. Also, every video I give away a Claw Max subscription. This video's winner is Jesse loves you for building an AI agent to process payments for a medical business. So, to enter the next giveaway, comment below with a recent problem that you might have run into. I'd love to help as many of you, so I'll respond to those comments below. So, layer number five is context maintenance. Refactor the bloat. This is a mistake that 99% of people make. If you're watching this channel, you likely use AI all day every day. And once you start following all four of these steps, your system will continuously have more and more contextual information. But that doesn't always mean that it just keeps getting better. In engineering, there's a concept of refactoring. And what that means is that over time, your codebase may get too complex that you have to step back and reorganize everything so that it's easier to manage. And in the AI world, everyone is go go, but rarely do they refactor the context in their system. So what happens in practice is a month ago, you could have told your system, be concise, and then last week you said another part, make sure to thoroughly explain this. And neither is wrong, but this can be the reason that your system produces inconsistent results. And the longer this gets neglected, the worse it gets. So, how do we fix this contextual rot? Well, think about your system in two different buckets, always on and on demand. Always on is the stuff that loads into every single session. For example, this is your claude MD file. The biggest mistake is that people think of this file as a container to store every specific rule. but instead these always on systems you need to think of it as a guide which will lead AI to the answer that lives in the other files. So to fix this problem use this prompt which will do two things. First it will make sure that your claude MD has proper mappings to the information and skills in your system. Again it's like a guide a table of contents to navigate within your project. Then if you already have them it'll review your existing ones to make sure that they aren't outdated. So that's always on. Now what about on demand bucket? This is all of the contextual information and skills in your system that are waiting to be referenced by Claude, right? It's on demand. It's go do this, use this. And what happens is that over time, you'll get an overlap between these on demand files. And this will lead to getting inconsistent responses. To avoid that, first, let's make sure that there aren't any skills that really overlap. And if there are neighboring skills, make sure that the skill descriptions themselves clearly guide Claude when to use them. So, let's say you have a draft email skill and then a draft customer support email skill. You would want to update the draft email description to explicitly say, "Don't use this for customer support emails." So, unless you're doing these periodic changes, you would have never thought of it in the first place. This is so important because you wouldn't have known to say this when you first made the skill. So, it's literally impossible for your system to stay up todate, especially if you're constantly improving it. The next is you have to think about your contextual information and how you structure it. Earlier in this video, we discussed bringing in information that's specific to your domain. Let's say you have a ton of sales context you added. You might have had it all sitting in a single sales folder, but the more you add, that might not be enough anymore. You may need [music] to partition it further. So you may need a folder called cold calls or lead nurturing, but what worked on day one might not be enough for day 100. So to audit this ondemand bucket in your project, run this prompt, which will do both of the things that I just covered. The reality is that generic responses don't just have to do with writing in your own voice. As we covered in this video, that's just the first layer. But layer two is knowledge context. This is what Claude reasons from. This is curated sources, training data, depth on demand. Layer three is about collaborative context. This is making Claude your thought partner, not just a mockingb bird. You add pushback rules, structure your system for second opinions, and you establish a board of adviserss. The fourth layer is strategic context. Make Claude have goal oriented perspective, not task oriented. Establish one and two-way door principles in your system. And then layer five is that maintenance context. [music] 99% of people make this mistake because they never think to do it. All these layers pile up context. So you periodically have to refactor the always on and the ondemand buckets to [music] keep your system clean. Now, if you want to dive more into setting up your system so it's optimized to build faster and get better responses, go check out this video where I walk through how to set up a self-improving system so that it automatically gets better without you. Taking what we learned in this video and pairing it with what I cover over there will take your setup to the next level. I'll see you over there. Peace.