Transcription
Most people use Claude the same way. They open a chat, start asking questions, and just keep going in the same conversation. And the longer they go, the worse the output tends to get.
Now, here's what's actually happening when you do that. Every time you send Claude a message, it rereads your entire conversation from the very beginning. So, message one might cost like 500 tokens, but message 30 costs 15,000 tokens because it's rereading everything before it. Your cost isn't adding up. It's compounding. And the more that builds up, the more stuff Claude is trying to keep track of all at once. Your early instructions, your knowledge files, your latest messages. And at some point, there's just too much in there, and Claude starts dropping things.
This is why I want to introduce you to a concept called context management. Managing how much Claude is holding in its brain, if you will, at any given time. It's the main difference between people who get consistently great output from Claude and people who feel like it's a coin flip every time. So, in this video, I'm going to show you what context window actually is, why it fills up so fast, and then give you seven strategies to manage it. So, Claude stay sharp the entire conversation for you.
So, let me zoom out for a second and explain what's actually going on here when it comes to context window. The way I think about context window is like short-term memory. It's essentially how much of your conversation plus any files or documents that you've shared with Claude that it can actually see and work with at one time. Claude's latest models, Opus 4.6 and Sonnet 4.6 at the time of recording this video have a 1 million token context window. And a token is basically a small chunk of text. A word is roughly one tokenish, sometimes a little more. So a million tokens is roughly 750,000 words. That sounds like a lot, right? But when you factor in conversation history, any files you've uploaded, project instructions, and then MCP servers, Claude skills, your cloud MD files, all of things we'll talk about today, all that fills up super super fast in your context window.
And here's the real kicker. Essentially, what happens is Claude pays the most attention to the very beginning of your conversation and the very end. Everything in the middle gets sort of less attention. So, as your context fills up, you're literally paying more to get lower quality results. And that is why context management matters so much.
Now, quick but super important point before we get into the strategies. If you want to properly manage your context, claude code is going to be your best bet. And I run it inside VS Code. And the reason is simple. in regular claude chat or even in cloud co-work you don't really have any insight into how much context that you've actually used when you are working on a specific project or a task like here in co-work insight again I can go into settings usage but this is super inefficient to have to do this and again I don't have any insight into a per task for example amount of context being used so that's another reason why I like to use cloud code inside of VS code I can see exactly how much context is being used per chat and I know when I need to take action.
All right, so with that, let's get into the nine strategies.
So strategy number one is start new chats for new topic. Don't just keep going in the same conversation. So just as an example, let's just say I'm in this chat here in Cloud Code inside of VS Code again and I'm done with the specific task that I'm working on in this chat. So rather than switch to a new task within the same chat, I just open up a new uh cloud code chat window here and I have a brand new context window because again every single message in a long chat is exponentially more expensive than that same message would be in a fresh chat.
And I want to clarify something here that trips a lot of people up. Context window limits are per chat, not daily. Every conversation that you have has its own context window. So, if one chat starts giving you garbage answers, it's not because you've quote unquote used up Claude for the day, it's because the context window for that specific conversation has filled up. Now, this is different from your usage limits. Usage limits are essentially the overall amount that you can use Claude within a certain time period across all your chat.
Now, to get the most out of your AI, you need to manage your knowledge in order to manage your context. AI's leveled the playing field, right? Same model, same tools, same capabilities. So, what's your edge? It's not the model. It's not the prompt. It's you, your knowledge, your context, your judgment, and experience building your business. But right now, all that's scattered across dozens of apps and invisible to your AI. Well, that's the problem that Recall 2.0 solves. I've used Recall for over a year now to save content that I can't afford to lose. Where Recall 1.0 is my personal knowledge base, 2.0 turns that saved knowledge into the fuel behind my AI. Now, the headline feature is agentic chat. So I can come over here in the lefth hand column, click on chat and you choose what you want to talk to. You can chat with your save knowledge, the internet or both. You can pick your model, GPT, claw, gemini, etc. No lock in. You can also chat with specific context. So let's just say I want to do some research on sleep both within my recall knowledge base and from the web. So here is the condensed research from my recall knowledge base right through here. Then I can go down here. It's found new studies. So I asked to do new web research as of 2026. And then also give me an action plan. You get answers that no other AI can give you because no other AI has access to what you know. And this is huge. Recall now has MCP and API access. You can connect external tools like Cloud Code to it. So I invite you to try Recall for free today or you can use code recall rick for 25% off an annual subscription or the monthly max plan valid until June 1st 2026.
Okay. Strategy number two. Use the context command and the status line command. You can't manage what you can't measure, right? And Cloud Code gives you two tools for this. The first one is the context command. So forward/context. You just type forward/context into cloud code and it'll break down exactly where your tokens are going. And here's what I like to do. I'll run this in a brand new chat before I've even typed in anything just to see what the baseline looks like. So let's do that right now. I'm going to type in forward/context. And what it's going to do, it's going to show you how much of the context window is being used up without me even typing anything yet. And as you can see here, I'm using almost 20,000 tokens before I've even typed anything. And this is all obviously super super helpful because now it's telling me where my tokens are going before I've even typed anything into the chat box.
Now, the second tool is the status line command. You essentially set this up once and it will sit at the bottom of your screen showing your context usage. As you're working on something within a cloud code chat window here in VS Code, you will see the little context window sort of chart which will give you a percentage of how much context window that you have that you've used so far. Also over in the terminal, if you're using cloud code within the terminal, this is where I can type in status line and you can see right here where it says setup cloud code status line UI. It gives me the number of tokens for the context window and then also the 2% here is this context. it will this will change and tell me exactly what percentage of the context window I've used for this particular chat.
Okay. Strategy number three. I recommend compacting manually yourself when your context window gets to be between somewhere 50 to 60%ish. Compacting is essentially when Claude summarizes and compresses the older parts of your chat to make room for you in the context window. Think of it like someone condensing a whole bunch of meeting notes into bullet points for you. you're keeping the key decisions, but you're losing some of the detail. So, when you see that 50 to 60% of the context window being used in that one chat, that's when I like to compact. I just run the compact command, which I'll show you in just a second.
Now, before I do that, I like to always save my progress first. And what I mean by that is you're essentially telling Claw to track its progress up to that point. All the key decisions that have been made, what's been happening in the chat. And I actually have a keyboard shortcut that I set up for this that just says save. So if I type in save, you can see here it says save all and update progress files and plan files accordingly. Ensure updates have also been pushed to my git. So I'll go ahead and click go. And after saving the progress and updating it to my git repo for this particular project here, you can see that the little context window chart has shown up here. And it's telling me that I've used 59, is that right? 69% of my context window. So 31% is remaining. Okay, I've used up too much for my liking here. So now I want to do is I'm gonna type in uh compact and click go. And you can see where it's going to say, all right, it's going to start compacting now. Okay, and it's just done compacting and we have freed up 107,000 tokens. So now just for the heck of it, let's just type in for/context and we can see where we're at. So we have we've only used 31% of the token context window. We freed up 107,000 tokens out of a 200,000 token context window size. So it freed up a ton of space in my context window here. So again, after a few compacts, save your progress, get a summary of what you've been working on, and then start a fresh chat. And then you can just pull that summary into the new chat that you just saved from the previous chat. This is going to give you a fresh context window with all the important context carried forward. And you're going to get better output because you're starting with that fresh context window.
All right, for strategy number four, it's what I like to call the fiveinut cash rule. And this one catches a lot of people offguard, and it did me as well. I actually wasn't aware of this until putting this video together. So, when you're actively chatting with Claude, it's caching your conversation, so it doesn't have to reprocess everything from scratch each time. This is called prompt caching. That cash expires after about 5 minutes of inactivity. So, if you step away from your computer to go grab coffee or check your phone or whatever and you come back after 5 minutes, your next message is going to reprocess the entire context at full cost, if you will. And that's why some people feel like their usage randomly spikes. So, here's what to do instead. If you know you're going to take a break, cool, you should be save your progress, like we just talked about, compact like we just talked about, before you walk away. That way, there's less to reprocess, if you will, in the chat for Claude when you come back.
Okay, for strategy number five, it's the allimportant claude.md file. This is your project instructions file. This is kind of like your system instructions. It's essentially how you want Claude to act and what you want it to do. And you'll have a claude.md file for your overarching Claude account, but then also at the project level. So for me, the one for my overarching account is here on my computer locally. So if I click on it, it opens up here. These are my account level instructions. And you can see here, it's not very long. And we're going to talk about that here in just a second. So that's the overarching claw.md file. But then again, you're going to also going to want a claw.md file for each project that you're working on. So for example, in my newsletter project, if I come down here to click on claw.md, this is my claw.md file for that specific project. And here's the thing, Claude's going to read these at the start of every single chat. So if your claw.md file is a thousand lines long, every time that you say hi to it, the whole thing is going to get reread and counted as tokens towards your context window. So rule number one when it comes to your cloud.md file, keep it lean. Keep it under about 200 lines. And honestly, the shorter the better. Boris Churnney, who's the anthropic engineer who created claude code, shared that his claw.md file, his overarching one, is roughly a 100 lines.
So what do you include in your claw.md file? This is where you're putting your most important rules, like your text stack, for example, your coding conventions if you're coding. And here's a key tip for you for your claw.md file. Treat it like an index. And what I mean by that is essentially it becomes like an if this, then that. You're telling Claude where to find information rather than putting all the information in that cla.md file itself. It's like if this, then go here for this reference document or use this tool. So within my newsletterclaw.md file for example I have a section reference files load as needed click here before improvising. If a file exists for your task use it and I give the file name and then also what the file is about.
Now here's where it gets really interesting. Beyond just keeping it lean think about the claw.md file as what I'll call a systems constitution. It's a place that contains your decisions, your architecture rules and pointers to where everything lives. So, every time you make a decision in a conversation like we're using Stripe for payments or client onboarding always goes through these five steps. That decision should get captured either in a memory file or in your cloud.md file because if it's not captured, you're going to have to reexplain it every time. And reexplaining is respending tokens and thus using up your context window. And in fact, a great thing to do when you're done a task or having cloud code do something for you at the end, ask it, were there any decisions that were made or anything that we did in this particular task that should now be included in the cloud.md file and it'll tell you. And by the way, you can also add rules to the cloudMD file to have Claude manage its own token usage, which is really, really cool. Here's an example in my newsletter cloud.md file for context management. So I say start of a session read this project log.md file to pick up where last session left off. That way it doesn't have to read the entire history. And so this is now turning your cloud.md into a document that actively saves you tokens on every single interaction.
For strategy number six, I recommend using plan mode in claude code. And this one's huge. It prevents the single biggest source of wasted tokens, which is Claude going down a wrong path. So here's what happens without plan mode. You give Claude a task and it just immediately starts building, right? It just goes and then 5 minutes later you notice that it went completely in the wrong direction. Well, all those tokens it spend going down the wrong path, those are wasted tokens against your context window. Well, using plan mode in Cloud Code fixes this. And you can toggle it on. And you can see down here where it says edit automatically. Well, I can toggle this over to plan mode. So, let's just say I want to have Claude Code clean up my Obsidian folder structure, which I actually do want it to do. So, what I'm going to do here is I'm going to flip it over to plan mode. You can see that. And so, I'm going to speak into this and say, "I still really think my Obsidian folder structure is overly complex and it's hard for me to find things. Please recommend how we can clean it up. So, if I didn't have plan mode on, it would automatically start doing things." But you can see here, it's taking into account my question and going through the process of how it should answer me. Okay. Now you can see that it just said now I have a thorough picture. Let me design a simplified structure. Okay. So it's just wrapped up sort of the first part of the plan and it has some clarifying questions. So I can just go through and answer these questions and I just keep going back and forth with cloud code here in plan mode until the plan that it comes up with based on my feedback is the one that I actually want it to execute.
And here's the tip here. Remember we talked about the cloudmd file in the previous strategy. This is one of those rules that's worth adding into your claw.md file. I have a rule that says, "Do not make any changes until you have 95% confidence in what you need to build. Ask me follow-up questions until you reach that confidence level." So, it'll do that every time I'm using plan mode. And that one rule has saved me a huge amount of wasted tokens. And then one more thing here, I just alluded to it a second ago. Use Opus 4.6 for your planning, which is going to give you the deepest reasoning, the deepest level of thinking, if you will. And then once you're happy with the plan, you can switch over to either sonnet or haik coup to actually execute that plan. So you're getting opus level planning depth but sonnet level token costs for execution. That's going to save you on tokens and it's going to be cheaper for you.
Okay. Strategy number seven. Be mindful of MCP servers and cloud skills. every MCP server that you have connected to Cloud Code loads all of its tool definitions into your context on every single message. And one server alone can be around 18,000 tokens per message. So again, I'm already seeing here 4.4% uh of my context window is being taken up by MCP tools. And you can see I'm scrolling through here. I'm using Google Calendar, Slack, uh notion, Superbase, etc. So the fix here is pretty simple. only keep the MCP servers connected that you actually use regularly. And claude skills work much the same way. It's outside the scope of what a cloud skill is for this video. I've talked about it here on the channel before. I'll link to it right here. So, as you can see here, these are the skills that I am currently using. And each one of these is taking tokens just to scan through the skills. And again, that's happening every session, even when the skill is completely irrelevant to what I'm working on. And I've heard from people who cut from they have over 40 skills and they cut it down to about 12 to 15 and they saw immediate improvements in the quality of the response that they're getting from Claude. Not just token savings, but better output. So here's the rule of thumb. Only keep the skills that you actually invoke more than say 20% of your sessions. If you're not using it regularly, I would remove it.
So there you have it. Seven strategies for managing your context window so that Claude actually stays sharp for you. Now, honestly, there are way more advanced strategies out there. But honestly, what we just covered is going to handle the vast majority of this for you. And it's not complicated, right? The real work is just building these habits into how you use Claude every single day. Again, if you're an online entrepreneur and you want to be leveraging AI to be taking care of the 80% of your busy work every day so that you can be focusing on more of your zone of genius, only the things that you can do. I want to invite you to join my AI playbook community where you get access to my your zone of genius app which is your full AI team that will take care of the 80% of your day-to-day task for you. I'll leave a link in the description below to the community. Again, thank you so much for watching as always. Appreciate it. I'll see you in the next video.