Transcription
I'm gonna let you in on a secret. Nobody gets 10x results from default vanilla chat GPT, vanilla claw, vanilla Gemini. It just isn't how it works.
But most of us have slept on the big levers that these model makers have released to help us to customize these models to get the most out of them. This video is all about those levers, what you missed, and how you can customize your AI to get the most from it. It's how AI averages you out and how you can stop it.
Lever number one is memory. So Chad GBT has a way they handle memory. Claude has a way they handle memory. We'll get into it.
Instructions is lever number two. Again, Chad GPT, Claude, Gemini, they all have their versions of instructions. We're going to hop into that.
Style controls. Chat GPT has like eight different personalities. Claude has different style summaries. We're going to get into that.
And then apps and tools. What apps and tools do these models support? Look, together these are big big changes, but most of the time we hear about them in clickbait articles and we're told do one thing specifically. I don't want to look at doing one thing specifically because that doesn't change the averaging function.
You are being averaged out into a median AI user and I'm interested in you understanding the levers so you can customize your AI into something that truly allows you to be transformative. So what does being averaged even mean here? The simplest way to understand this is imagine a restaurant that wants to create one dish to satisfy the widest possible range of customers. Let's say it's pizza. They don't want to delight anyone in particular. They just want to avoid disappointing too many people. Papa John's, Pizza Hut, you get the idea.
The chef studies what most diners order. They analyze which flavors get consistent approval across different demographics and they just optimize for the middle. What do you get? Well, it's edible. It's competent. It's technically fine. You can make the cheese look nice on an ad, but not your preference, right? It's not spicy enough if you like heat. It's not subtle enough if you like delicate. It's not adventurous enough. It's too adventurous if you're feeling mild.
This is exactly what AI does with answers. It's like the Pizza Hut approach. It's not trying to give you the best response for your situation. It's trying to give the best response for everybody who might ask a similar question. It's the statistical middle. It's the median.
When you ask for restaurant recommendations, you're going to get restaurants that would satisfy the most people who ask for restaurant recommendations. When you ask for career advice, you're getting advice that applies to the broadest set of people in roughly your situation. When you ask for code, you get code that follows the conventions most developers would expect. This is why your output always feels just a little bit off. It's not wrong. You can't point to an error necessarily. It's just not yours. The recommendations hit tourist spots instead of the places you'd actually like to hit. The advice applies generally, but not to your constraints.
And most people experience this and they think to themselves, well, the AI is just okay, right? It's probably the AI's issue. They don't realize there's a mechanical reason and they don't realize that it's fixable.
So, how do models learn to be average in the first place? This is not speculation, by the way. We we know this. Modern AI assistants go through something called reinforcement learning from human feedback. And here's how it works. The model generates multiple responses to the exact same prompt. Human raiders compare them and pick which one they prefer. The model learns to produce outputs that the raiders would choose. You catch the keyword, right? Raiders, not you. A pool of people who rate outputs and judge which seems better. The raiders are not experts in your field. They're not familiar with your constraints. They don't know your preferences about where you want to go in Paris when you travel there. They're looking at two responses in picking whichever one seems most helpful, most clear, and most appropriate. Hint, it's probably the one with the Eiffel Tower.
The model's optimization target is thus not give the specific user what they need. Give Nate what he needs. It's produced something a typical human would rate pretty highly. So when thousands of raiders evaluate millions of outputs, the model learns to hit the middle of the preference distribution. It learns the answers that would satisfy most people. It learns the median. And this is not a secret. Anthropic publishes papers describing this. So does open AAI. Nobody's hiding it.
And there's an irony here because the training process that makes these models so helpful in general is exactly what makes them mediocre for you and me specifically. The same mechanism that prevents the AI from being weird or offensive or unhelpful also prevents it from being calibrated to your particular needs. The implication is significant. Every time you use default settings, you're getting an answer optimized for a hypothetical typical person. The training literally encodes what would most people want here as the target. And you're not most people, you're you.
For the last couple of years, prompting was the only way to escape the average lifestyle. You would frontload your context into your question. You would specify your constraints and your preferences, and you would steer the model to adjust. And every conversation had to start from scratch. That has now changed. There are now at least four distinct ways to steer AI away from the median. Four levers beyond the prompt itself. Most people are using none of them or only one. And here's what you need to know about each one.
Lever number one is memory. Memory is the AI retaining information about you across conversations. So instead of starting fresh every time, it remembers your context. It remembers your job. It remembers your projects, your preferences, etc. The promise is very powerful, right? The AI knows you and builds on that. The reality is platform specific.
Chad GPT's memory works in multiple layers. There are saved memories that are facts you explicitly ask it to remember. And then there's something broader like a sense of chat history where chat GPT references your entire conversation history to understand your preferences. This can be very general. When chat GPT pulls from past conversations, now it does pull clickable citations that let you know exactly which chat it's pulling from. And while this makes the system transparent, I can tell in context that it's still not a very good memory implementation, it misses stuff I would consider obvious.
Chat GPT also has projectonly memory. When you create a project, you can isolate the memory from general chat GPT use, and what you discuss in that project stays in that project. One recent change worth noting is that temporary chats now retain your memory, style, and personalization settings. They used to be very stripped down, and now they're less so.
So, what's your key tactic with chat GPT? Tell chat GPT to remember specific preferences that you care about. Remember that I prefer one-s sentence answers to factual questions is a great example. Remember that my audience always has people that think that they can build their own local models. That's another example. I'm sure there's some of you out there. The automatic system captures a lot, but intentional memory is very reliable if you're starting to cultivate it with that mindset.
How does Claude work differently? Well, it has two components. Claude can search past conversations, sort of like a rag style retrieval, and it can also generate a memory summary that synthesizes key facts across your chat history. And that summary will update periodically. The distinguishing feature is that Claude's memory is project scoped at default. So every project has a very separate memory space and your startup discussions don't bleed into your vacation planning. The isolation is very intentional. Claude keeps contexts very focused because it needs clean context to work. This gets it the way they build their agents.
Claude also supports memory import export. You can bring in memories from chat GPT or push them out to cloud memory in another account. The interoperability is limited. There's not a one-click import, but technically the capability is there. My recommendation with claude is to use your projects very deliberately. If you're working on something with a very distinct context like client engagement, just create a project for it. The project gets its own memory, its own instructions, and it works really well.
Gemini has personal intelligence that connects to your Google apps. It has, you know, Gmail, photos, YouTube, etc. The pitch is that you can ask about tire options on the car and Gemini finds your car model from the Gmail receipt and gets the tire sizes right. Settings. Personalization lets you connect or disconnect specific Google apps. So you can kind of tune how much personalization it has. The key tactic with the Google ecosystem is just to decide how much data you're willing to give Google. If you want to connect them all, you get immediate personalization. The trade-off is a privacy surface area and you're going to have to make that call.
So that's lever one. That's memory. What about lever two? Instructions. Those are persistent context about who you are and how you want your AI to behave. Severely underused by most people.
Chad GPT has several instruction layers. It has custom instructions which are multiple text fields where you can talk about what it should know about you, how you would like chat GPT to respond. It has project specific workspaces that come with their own instructions. And it even has custom GPTs. The key tactic here is that your biggest leverage is in being specific. Be concise is not super effective at steering the model. Instead, say, "For factual questions, please answer in a sentence. For analysis requests, I really need you to walk through the reasoning step by step." When you are clear about what you're looking for, you are helping the model to understand under what circumstances you want that behavioral response.
Now, Claude splits instructions across multiple places, right? There's profile preferences, there's project instructions, there's styles. The key tactic with Claude is that Claude's style feature is really underused. If you have a distinctive writing voice, if you upload samples of your best work, Claude can generate a style profile from them. And every response, Claude will then be thinking about how to match your tone, how to match your sentence structure, etc. This is much more powerful than trying to describe your style in words. And even if Claude doesn't get all the way there, it gets you most of the way there on first drafts.
Claude markdown files deserve their own note. So for developers using claude code, the instruction layer that actually matters is a claude markdown file. Boris Churnney, who created Claude Code, described his team's practice. Whenever Claude does something wrong, they add a rule to claude. Markdown so it doesn't happen again. The file is checked into Git. The whole team contributes. Essentially, the file contains project architecture, coding standards, and common commands that everybody on the team can see and update all the time. Treat this as a living document. Every time Claude does something you don't want, just add a note. The first version is going to feel sparse, but within a month, it's it's going to be all filled out.
So, that's instructions. Lever number three is apps and tools. And by the way, if you're wondering, I can't remember all this. That's fine. It's all going to be in the Substack. Tools are capabilities the AI can use. Searching the web, running code, creating files, reading documents, etc. If web search is enabled, the AI looks things up. If it's disabled, it works from training knowledge. So, look, most people have default enablement and they don't think about it. And that's the issue. And I want you to understand that there are a lot of different ways to configure your apps and tools that will profoundly shape your experience. And we should start with model context protocol because that underlies so much of the rest of this. The MCP standard explains how most AI systems today connect to external tools. Think of it as like USBC for AI. It's a universal interface that lets any AI connect to any tool through the exact same protocol. Enthropic created, but everyone's jumped on board. There are over 10,000 MCP servers out there, and lots more on the way.
So, how do people use these connectors? ChatGpt will call them apps, and you can connect to Gmail and Calendar, etc. And once connected, ChatGpt will automatically reference them where relevant. What I have found in practice is that the where relevant is very ambiguous. You don't have to select them manually, but you may have to remind chat GPT. It has the capability. It also doesn't have a super deep search capability.
On claude, you have a much wider range of MCP servers, but the connectivity isn't always reliable. It is, for example, quite tricky to connect to Stripe, but very easy if you want to connect to Figma. And that changes all the time as people mature those MCP server implementations. So claude is one of those things where you have to think intentionally what are my tool sets and then look regularly and say are there MCP connectors into claude where I can use them. Now Asana was just added.
Gemini is shorter on tools than it should be and it's one of the big weaknesses of the Gemini ecosystem. While personal intelligent will connects personal intelligence will connect to apps. Gemini itself is not big on tool use and that is one of the reasons so many builders prefer chat GPT or claude increasingly claude. So think about if you're using this lever your tools are really steering the inputs. They're not just features that you add. If you want the AI to work with your real files, think about where they live and connect them. If you want verified code, think about how you enable code execution. Turning tools on and off changes the character of responses. A model may lean more on web search than you want if you enable internet. These tools are not always good or bad. It's about you being intentional about what you want.
And lever number four, style and tone control. So style controls let you adjust how AI communicates. Chad GPT has eight different personalities ranging from friendly to candid to nerdy all the way to cynical. On top of presets, they also have granular characteristics around warmth, enthusiasm, headers, and emojis because apparently people complain about emojis. And so you can pick a personality and then dial it the way you want. Your key tactic there is to describe the default personality and then to be very clear in your instructions and in your settings so that there is no conflict. If there is ambiguity or conflict between your instructions. If you say be verbose in your instructions and concise in your personality, you're just going to burn tokens and make chat GPT sweat. So don't do that. Think about what you really want.
Meanwhile, Claude offers three built-in presets: formal, concise, and explanatory. The custom style feature is quite sophisticated, and it allows you to sort of upload what you want, which is what I've talked about. But fundamentally, if you don't want to create a custom style, you should be picking a style that reflects how you actually behave. Like, if you are actually a very casual Claude user, don't select formal. Go with something like explanatory where you can have longer conversations. Think about your actual usage, not your aspirational usage.
Across all four levers, I've observed a really common failure mode. Being too vague to really steer the model. Like I said, be concise doesn't move you. Be direct doesn't move you. The instructions that work need to be specific enough to change the shape of the output. Compare the difference between be more helpful and when I'm stuck on a problem, please ask me diagnostic questions rather than immediately giving solutions. I learn better by being guided than by being told. Wow, that is so much better. You're going to get so much better responses. Compare, I'm a professional, terrible, with I've been doing product for 15 years. Please skip fundamentals and go straight to nuance. The specific versions tell the AI where you are at and help it to help you.
So it's not just delivering that averaged out median answer that always feels off. And this is where we start to separate people who get real value from AI and people who find it perpetually mediocre because every interaction is generating information about what you need. And if you set your levers correctly, it starts to compound. Because every time you think that's not quite right, think of it as discovering a steering input, not just something you can get frustrated about and say, "Well, AI didn't get it again." Because most people will correct in their head, get frustrated with AI and move on.
The people getting 10x results, they do something different. They capture the corrections and when they notice a pattern, they encode it back into the AI and add it to their instructions. They tell memory to retain it. They update their style settings. Boris Churnney runs five claude instances in parallel and another five to 10 on cloud.ai and ships roughly a 100 PRs a week. His workflow is not magic. It's just the discipline to look at every mistake that Claude makes and update a rule in claude.mmarkdown. You don't need to be an engineer to do this, right? You can keep a notes file. You can find yourself making the same correction twice and write it down. You can review your instructions manually every month. This is actually not that hard and the gap will widen over time as you start to invest in getting the levers right.
Now, I want to be honest here. Steering fixes the personalization problem. It does not fix everything. When the model hallucinates, that's not an averaging problem. No amount of personal context fixes that. There's also a ceiling in creative work. When AI generates pros or images, its training data pulls toward the center of the distribution. You can steer against this, but you're still fighting gravity.
Steering always takes effort. You're figuring out your position. You're encoding it. You're maintaining it. It's costing you time. And if you use AI only occasionally, to be honest, it's probably not worth it. But if you use your AI multiple times a week for similar types of work, the math changes because a few hours of investment every now and then buys you permanently better output. And the compounding effect in saving you time is real. and it gets better the more you use it. Know which kind of user you are.
If this feels like a lot, you can really start very simple. You can pick one task where you use your AI regularly and you just the output doesn't feel right. And over the next few sessions, notice the adjustments you're making and write them down. And that's it. And then go in and find the custom instruction setting for your preferred AI and stick those in and notice the difference and iterate. That's as simple as it gets.
I want you to know that the median isn't mandatory. The AI you're using is trained on everybody else's feedback. It learned to please everybody a little, which means it learned to please no one in particular. Default output really is median output. It's optimized for very typical users with typical needs. And you are not typical. I am not t typical. Your constraints are specific to you. Your goals are specific to you. And the farther you are from the average, the more default settings will fail you.
Please don't forget your levers. You can you can go beyond prompting. You can do memory, instructions, tools, and style. And by the way, prompting is still useful. I didn't talk about it in this video, but it's still helpful for steering in conversation. Most people are going to ignore these levers or do only one. If you are starting to get averaged and you're tired of it, you can adjust more than one lever and you can very quickly start to compound toward a more personalized AI that actually fits you. So, the choice is yours. You can stay at the median or you can steer the ship and get the AI you want.