--- title: 'Claude Code Best Practices: Tips from Anthropic' source: 'https://youtube.com/watch?v=sfsBxdDKkUg' video_id: 'sfsBxdDKkUg' date: 2026-07-14 duration_sec: 0 --- # Claude Code Best Practices: Tips from Anthropic > Source: [Claude Code Best Practices: Tips from Anthropic](https://youtube.com/watch?v=sfsBxdDKkUg) ## Summary The video is a live stream where the host and chat watch an Anthropic talk on Claude Code best practices. The talk covers how Claude Code works as a terminal-based agent, its use cases, and tips for getting the most out of it. The host and viewers discuss their own experiences, including using sub-agents, MCP servers, and managing context windows. ### Key Points - **Introduction to Claude Code** [00:00] — Claude Code is described as a terminal-based AI coding agent that uses tools like bash, glob, and grep to explore codebases agentically rather than relying on indexing or RAG. - **Best Practices: CLAUDE.md Files** [10:00] — Using CLAUDE.md files in the project directory provides persistent instructions and context for Claude across sessions. These files can include project overviews, style guides, and test commands. - **Permission Management and Auto-Accept** [15:00] — Claude Code has a permission system for write/execute actions. Users can configure auto-accept for safe commands (e.g., npm run test) to speed up workflows. - **Context Window Management** [20:00] — When the context window fills up, users can run /clear to start fresh or /compact to summarize the session and continue. Compact mode creates a summary for the next session. - **Planning and To-Do Lists** [25:00] — Using plan mode (Shift+Tab) lets Claude search and propose a plan before making changes. To-do lists help track progress and allow users to intervene if the agent goes off track. - **Advanced Techniques: Multiple Agents and Escape** [30:00] — Running multiple Claude Code instances in parallel can boost productivity. Pressing Escape once stops execution while preserving state; pressing Escape twice allows jumping back to previous conversation states. - **New Features: Model Switching and Thinking** [35:00] — Users can switch models with /model (e.g., to Opus) and trigger extended thinking with 'think hard' or 'ultra think' to improve reasoning on complex tasks. - **Sub-Agent Logging and Communication** [40:00] — The host demonstrates having sub-agents write logs to markdown files (e.g., planner_log.md) to share state and actions with the primary agent, enabling better coordination. ### Conclusion Claude Code is a powerful agentic coding tool that excels at terminal-based tasks, and its effectiveness can be significantly enhanced through proper use of CLAUDE.md files, context management, planning, and sub-agent orchestration. ## Transcript Hello. Looks okay. Hey Edgar, nice to see you. I saw your uh I answered you on u LinkedIn. Welcome. Welcome here Spaso. Nice to see you. So, I saw Antropic put out like a bunch of videos yesterday. I haven't watched any of them yet. Hello, BotFather. Nice to see you. Yeah, I wanted to watch this as well. So, uh, I haven't seen it yet, but I don't know if this is like up to date with the latest changes like with the sub anient edition and everything, but I thought it could be interesting just to watch it together. Uh, and yeah, see what you guys think. Thought it could be interesting. Uh, I'm going to give it a few minutes before we start though. Uh, see the new stealth model on open router. Do you mean the is it does it have a name or is it like that alpha horizon alpha one? Is it this one? Horizon alpha. Edgar. Yeah. Yeah. I read something about it on um I don't know where I read but um someone thought it could be like the new opensource model from OpenAI. The new Yeah, open source model. I don't know what they call it. Someone thought about that. Talked about that, but I don't know. Haven't tried it. So, I don't know what the code clo best best practices video is going to be about. probably just going to be talking about how the Antropic uh me staff members uses cloud code. So, be interesting to hear what kind of tips they have. Uh I suspect it's going to be heavily software dev focused, but a lot of you guys mostly deal with code anyway. So it should be relevant. Okay. So you gave it a try. Took like 10 minutes to build a 3D chess game. As far as one shot, it's pretty good. Okay. Interesting. So, what uh did you just use it on uh like um like uh cursor or something or just in the chat playground? Does open router have a playground? I'm not quite sure. Haven't tried it. Maybe this weekend. I'll have a look at it. I was hoping for GPT5 maybe like for this weekend, but it doesn't look like it. They usually announce their live streams, don't they? Maybe next week. I'm still I was just testing uh I was just trying a different look for my app. I'll show you later. But, uh, I think we're going to start watching this video, but I'm just going to give it like a few more minutes. Uh, I also, if you were here yesterday, we worked on uh my new simple app and I kind of fixed some issues I had. Hello, Anthony. Nice to see you here. We're going to watch some new videos from Antropic is the plan. Yeah, I see what you mean. Uh, Botfather. So, they kind of maybe they are waiting for Gemini. They have a bit of like a beef between Open AI and Google on these releases. So hello Mickbab. You think GT5 is going to be like the do you mean 5th of August or 7th? I haven't really read any like rumors about exactly any dates. But yeah, if you're here yesterday, u a lot of you guys tested out the app and I think I fixed most of the issues. So, I might try it again later, but not now. Going to start with this And I kind of ran my I have this stream agent that does some research for me on AI topics just to see what is happening kind of on Reddit. Hit X and yeah, hacker news just for some interesting topics. MCP growing adoption. Google Cloud artifacts enhanced with full API access. I haven't tried the new claw uh artifacts on the API yet. Anthropic cloud code rate limits controversy. This is the number one topic. Okay. So the new I tried out like a new I asked kind of my sub agent to create a new design for the app that was kind of based on uh Windows 95. So we got this. I don't think I like it too much, but it did a pretty good job though. But isn't this background? Isn't that Windows XP? So on the sign in page. Oh, we have an issue there. Okay. Yeah, I just tested this. Uh, but it looks a bit janky to be honest. So, I think I'm going to change up the text here and just go for this instead. So, this was the post. Open AAI's new open source models were briefly uploaded to hugging face. Can't really read this. Antropic dropped 17 videos to watch. Yeah, that's what we're going to do today. Look at some videos. Has anyone uh watched this video yet? If you see, have you seen it? It's pretty popular. Almost 50,000 views in like 15 hours. Hello Christian. We're going to watch some uh Claude code best practices video is the plan. So, I think we're just going to start it. I'm kind of excited to see it. Please enable English subs when starts watching. I can do that. No problem. So, I'm going to play a bit. If you have your sound on, just tell me if it's too low or if it's too loud. It's kind of hard to tell. So, I'm going to play just a few seconds and let's hear how the sound is. Let's get started. Welcome everyone to Cloud Code best practices. In this talk, I'm going to talk about kind of what CL >> how did was it too loud, too low. Okay. So, I'm much louder. I can turn it up a bit. This was pretty low. Okay. Yeah, it kind of Let's try this one. Hello, by the way, Stor and Jake. Let's try this one. Let's try now. >> Then we'll peer under the hood a little bit to kind of understand how cloud code works. And then knowing that because it's useful to kind of know how your tools work, we're going to talk about good use cases for cloud code and also best practices we've figured out both internally and from our users uh for getting the most out of this tool. Uh but before I get started, I'd like to introduce myself real quick and talk about how I ended up on the stage. So my name's Cal. >> Yeah, this is pretty low. Yeah, I'm trying, but I don't have any more to go on here. I guess the recording is very low. It's on max here. I don't have any other way to turn it up now. I think then I need like an extension or something. Let's try this. Some videos are uploaded with a very low. Um setting Let's see how this works. Yeah, I can turn down the mic a bit. That's fine. Something like this. Something like this. Maybe it looks pretty good. Yeah, let's try this. and I joined Enthropic about a year and a half ago uh to help start up a team we call applied AI and it's the applied AI's kind of mission our team's mission is to help our customer okay let's try it and um if you're having issues just let me know we can try to adjust it okay so yeah let's watch this uh claude code best practices So this is probably going to be an anthropic talk on some kind of conference. I guess maybe it's an internal one. It's going to be Cal. Let's get started. Welcome everyone to Cloud Code best practices. In this talk, I'm going to talk about kind of what cloud code is at a high level. Then we'll peer under the hood a little bit to kind of understand how cloud code works. And then knowing that because it's useful to kind of know how your tools work, we're going to talk about good use cases for cloud code and also best practices we've figured out both internally and from our users uh for getting the most out of this tool. Uh but before I get started, I'd like to introduce myself real quick and talk about how I ended up on the stage. So, my name's Cal and I joined Anthropic about a year and a half ago uh to help start up a team we call applied AI and it's the applied AI's kind of mission our so Rammer said actual content starts from 10. Yeah, if he's just going to talk about himself, we might as well skip ahead a bit. Or is it is talking a bit here I guess. Uh but I guess just for the video sake I'm just going to play the full thing just uh doesn't matter too much. >> Team's mission is to help our customers and partners build great products and features on top of cloud. So what that really means is I spend a lot of my day prompting Claude to get the absolute best outputs out of these models. That said, I also love to code and I'm definitely one of those coders that like starts a lot of projects, has some crazy idea, and then just never finishes them. So I have this graveyard of just like code that I started, never really finished. Um, but I'm always spinning new things up. And late last year, I was in Slack and I was hearing about this new tool that a few people are using. They were saying it was really cool. And so on a Friday night, I downloaded the tool that would become Claude Code and I threw it at this kind of new note-taking app that I wanted to build. And like that whole week, I remember when Claude Code came out, there was this waiting list. So I I really wanted to do a video on it, but I remember I couldn't get access because I was a bit late and it kind of filled up the waiting list right away. So, uh, I think I had wait to wait like a week or something before I got access to it. And I think I was super impressed like in the first the first time I tried it. I was kind of impressed already back then. Hello, Christian Rosa. Nice to meet you. We are watching the video Claude code best practices again. Just kind of totally changed the way that I code and think about software engineering. I was carrying around my laptop with me all weekend. I was super addicted to just watching Claude Code work and I would press enter and I'd switch over to my browser and refresh and I watched this huge powerful application come together in front of my eyes. And I got way farther into this thing than I ever would have on my own. And it just blew my mind. And while I was doing this, I was a little worried. I was like, you know, I, you know, I kind of know how these things work. I'm like, man, I'm using a lot of tokens. I hope I don't get in trouble or anyone like notices. because I'm not really contributing to anthropic code. >> Yeah, I guess that internally at anthropic they can just spend I think I heard some kind of podcast uh on cloud code they can just they have unlimited token access of course. Hello Alpha. So Mickbab, you switched into cloud code in April. So you switched from cursor or I did that like a few months ago. Um but what I didn't know is that the cloud code team had built this internal like leaderboard tracking how much all the anthropic employees were using this and over the weekend I had shot to the top. And so through that, I got to meet Boris and Cat and some of the early cloud code team and I was able to start talking to them and say, "Hey, I love this tool. I also know a lot about prompting. Can I help you all out?" And so through that, I got involved and now I'm one of the core contributors on the team. And I do a lot of I work a lot on the prompting, the system prompts, how the tools work, the tool descriptions and tool results, as well as I work on how we evaluate this tool. So when we think about changing the prompts, how do we make how do we know we made things better or the same and we didn't totally ruin cloud code. So with that said, let's kind of dive in. So here's my current mental model of cloud code and how I describe it to people when people ask me. Cloud code is like that co-worker that does everything on the terminal. It's the sort of person that just never touches the guey. They're a whiz. I think of when I was a junior engineer. I had this mentor and I would walk over his desk and I would say, "Hey Tony, can you help me with this bug?" And he would whip it over his terminal and he'd be like doing all these crazy bash commands and changing things around in Vim. And I'd always walk away thinking, "Wow, that was crazy. I should learn how to do that." Um, I never did, but having Claude Code on your computer. >> Does any of you guys use like Vim? It seems very complicated to me. I think Claude Code has like Vim emotion. Hello Sean. So this video wasn't good. Was putting me to sleep last night. Yeah, I'm not quite sure. Uh I haven't seen it yet, but um yeah, if you It was bad. I guess it's going to be painful to watch it again. is kind of like having Tony next to you all the time. So, how does Claude code kind of work under the hood? At Enthropic, we try to always do what we call the simple thing that works. And what that means for Cloud Code is it's what we would consider a very pure agent. And anthropic when we talk about agents what we really mean is some instructions some powerful tools and you let the model just run in a loop until it decides it's done. So early not a fan of uh I guess Neoim. So Edgar says Vim is hard to start. You have to be deep into the editor stuff. If you if you have figured it out you're mad fast but it's really steep. Yeah, there's this big streamer on YouTube, the Primogen. I think he uses whim and it looks crazy. Mix says for the revolution for me the re revolution was when I was able to use cloud code in VS Studio. Yeah. So you kind of you kind of link it up to your ID. What I found, I remember when I tried uh cloud code kind of for the first time, uh I kind of gotten used to cursor that kind of indexes your um cursor kind of indexes your um your your projects. Uh but with cloud code, it always used these bash commands to get uh context from your projects. I I think I remember I found that very interesting. thought it was a clever way to do it and it was very efficient searching up uh and kind of finding uh the relevant parts of your project to bring it into context. And that's really what Clyde Code is. So it's tools, powerful tools, and the tools that you know someone that was really good at a terminal would be able to use. tools to create and edit files, to use the terminal, and then you can also do things like pull in other things with MCP. Now, on top of that, there's how Claude understands the codebase. And if you're going to build a coding agent or a coding tool a year ago, you'd probably have ideas like, well, okay, I'm going to get this user message about something about this codebase, and I'll need to figure out which files are relevant. So maybe I'll like index the whole codebase and embed it and do this fancy like kind of rag retrieval thing. That is not how cloud code works. We don't do any sort of indexing. Instead cla kind of explores and understands the codebase how you if you were new to a team and new to a codebase would explore a codebase and that is through a gentic search. >> Yeah, that's what I was talking about. So Sean says just use a local JavaScript worker runs in the background lightweight and will index everything into functions indexes graph. It doesn't need an agent with node or bun. So you use sorry I have to read that again locally worker runs in the background lightweight and will index everything into functions indexes graph. Okay so you mean kind of for keeping the context indexing and semantic search apparently slightly more efficient. Yeah, I guess it could be more efficient. Maybe it's faster. I'm not quite sure. It's a pretty complicated subject to optimize. I guess how you should Let's say you have a code base like 5 million lines of code. What is the best way to to understand a project? Gets kind of complicated, I guess, after a while. is the same sort of search tools you or I would use things like glob and gp and find and it can work its way through a codebase and understand what's going on and when we talk about a genic search that really means the model can go do some searches and then it can look at the resulting maybe if you in cursor it would be really annoying if you had to kind of wait for the the prep every time you need to make a change, right? So maybe in cursor it's more it's more natural to have thing indexed and uh I guess cloud code is meant to be sort of more agentic than cursor. So uh it's supposed to run more in the background right? then time is shouldn't be like a big factor. I guess code to rag won't work well but code AI analysis to analysis into rag with reference to the file is better. Okay, that's pretty interesting. Yeah, I guess there's a lot of solutions to actually doing this. Say, maybe I need to figure out a few more things. I'm going to go do some more searching and then come back. And then on top of these primitives, on top of this agent, we have a few things. We have a very nice light UI layer where you get to watch Cloud Code work. you see all the texts fly by and we have this nice permission system that allows the agent to work and allows and kind of forces the human to butt in when the agent is doing something dangerous. And then on top of that, we also care a lot about security in this tool. And so because quad code is just such a lightweight kind of layer on top of the model and the fact that our model is available not just behind anthropic APIs but also with our cloud providers AWS and GCP it's very easy and native to point cloud code at one of these other services if you feel more comfortable consuming cloud that way. Additionally, like Sean said, uh it makes sense to build a graph of u syntactical objects together. Uh you get semantics and syntax in direct access. Yeah, like I said, I don't have like any deep knowledge about this. What is the best way to keep context of any projects? And I think there's a lot of different solutions you could implement. What is the best one? I don't know. I don't even know exactly how Cursor does it. I guess they have like custom models and a lot of custom solutions. I guess they have now a lot of people ask me, "Hey Kel, what can I use Cloud Code for? Like what is it good at? Where is it interesting?" And the reality is it's kind of great at everything. So let's start with discovery. Often times in your career you will be dropped into a new codebase. Whether that means you're switching teams, you're switching companies, I don't know, you're starting to work on some sort of open source project. And probably when you're first getting started and getting familiar, you're not very productive because you're just trying to figure out where things are in the codebase, what patterns kind of the team is using, things like that. And Claude code can kind of help supercharge that onboarding process. You can ask claude, hey, where is this feature implemented? Or since it's great at the terminal, you can say, hey, look at this file and look at the git history and just kind of tell me a story about how this code has changed over the past couple weeks. One thing you can use cloud code for and I think this is underrated is in >> but isn't let's say you have this semantic search vector store rag system. Doesn't that get very complicated? Uh and the chances of missing relevant context increases the bigger the size of the codebases. It just seems like a very hard thing to scale with success. But I don't know I don't know much about it but I just remember the small rag uh projects I have done. There wasn't always I got the relevant context, but that was of course very low level or like very simple terms, but uh it just seems like it's going to get a big issue if you scale it, right? But again, what is kind of the alternative? Instead of just diving in and starting to work, you can use cloud code as a thought partner. So oftentimes when I'm working with Claude and I want to implement a feature or we're going to change something up, I'll open up Claude and I'll say, "Hey Claude, you know, I'm thinking about implementing this feature. Can you just kind of like search around and kind of figure out how we would do it and maybe report back with like two or three different options, don't start working, don't start writing any files, writing any files yet." And Claude will go off and use those agentic search capabilities and come back with a few ideas. and then I could work with Claude to kind of validate things and then we can jump into the project. >> Okay. So, you work on something uh professional space and I do this per project. Yeah, I guess it gets really complicated at like big scales. But, uh what do you think about Edgar? Could let's say we have a 10 million context window. Does that help or is that just means that uh every request is just going to use so many tokens? Oh, hello Rex. Welcome as a member. Really appreciate it. Thanks for the support. Appreciate it. So, what are you working on, Rex? Of course, cloud code is great at building and writing code. And I would say this in on two different fronts. One, it can do the zero to one sort of stuff. You drop it in an empty directory and you say, "Hey, build me an app. Build me a game." That demos where very well. It's very fun to do. It's very gratifying. Of course in reality what really matters is is cloud code good working in existing code bases and this is primarily what we focus on. Um and the cloud code team we have in our codebase abnormally high I would say unit test coverage. Okay. Okay. So, Botfather says, I think the whole point of indexing is that you get chunks and probability of relevance and with search you go you don't get that and you can easily not match the word uh by how it's written in the code. Good point, I guess. Hello V uh V Rick. Uh hello Sean, nice to see you back. you only need enough context window to keep the entire codebase uh and uh changes that have been made. Yeah, that's true. But I guess if you fill up that context window, you have to run that through the API request every single time, right? I guess it's getting cached. Maybe that's true. uh as long as an as long as instruction following isn't improving uh substantially, it depends. For searching a single bug, uh you could do this. If you need uh to reference different parts uh to draw conclusion, it falls off fast. Okay, I see what you mean. And that's because cloud code makes it so easy and just straightforward to add unit tests. So we have great code coverage. And then the other thing we have in cloud code in our own codebase is we have great commits and PR messages because when we finish working we'll just say hey cloud write the commit for me write the PR message for me. We also see great opportunities to use cloud code in kind of the deployment like deployments and in other parts of the life cycle. And this is a few other people have talked about this but this is using the cloud code SDK. So using it headlessly, using it programmatically, being able to sprinkle in a coding agent agent anywhere. And so that's things like sprinkling it into CI/CD to use it in GitHub, for instance, to help people um programmatically. Yeah, that's kind of basic, isn't it? Uh Sean says, uh I tried out Serena MCP for semantic uh code search. It was a bit flanky. Okay, I might check that out. Hello, Dream Slow. Uh, so you're using CL code in a very large code base with 30,000 plus files for an online shop and it's not performing well. Okay, that's interesting to know. So, you say augment code is better with their context engine. So, they probably index your full code base then I guess. Uh so Sean says so locally and without an LLM you can use a simple file system scripts to index every file on your machine based on semantics keywords functions concepts and tokenized uh it makes the LLM locate fast. Oh that's pretty cool. So when you say semantics, do you use vector search or semantics is only like keyword based or is it like a hybrid? God, I wish I could just paste the files here so you could see it. Yeah, semantic search is not a panacea. Definitely there are better ways to search. Yeah, I see what you mean. Dreamflow. So, augment code, we had the tech friend in the chat here the other day. He is working with augment code now and I sent him an email. We might do something on the channel with augment code. Semantic is searching by the code meaning. Uh, yeah. Okay. So, yeah, I kind of miss I kind of some of the English kind of gets lost with this. So semantic is that let me see now semantic that is uh is that relevance? It's kind of hard for me to keep track of all these terms. And then finally it's great kind of with support and scale. It can help you debug errors faster. One thing that we saw when we started giving cloud code to customers and talking to them about it, we didn't totally predict this was a lot of customers or potential customers said, "Hey, we've been me we've been kind of putting off this like large codebase migration. People that are on old versions of Java trying to get it to a new one or a team that's on PHP and they're trying to get to React or Angular. We've talked to multiple teams like this and having a tool like cloud code makes projects like that a little more digestible when you go to your team and you say hey we're going to spend a month you know refactoring or rewriting large parts of the codebase. >> Yeah. So I think uh Ramsh was a bit was correct. The first 10 minutes here was a bit yeah not so interesting. Uh semantic is less uh what it is but more what it means. Yeah. So relevance, right? Same as an LLM searches for concept within its knowledge. Hello, Tetuin. Is there nothing called structured context like creating instruction that guides the AI on how to search and use their tokens on not to do greedy searches all the time? I'm not sure. I haven't heard anything about that. Maybe uh Sean or uh Sean or Sean knows semantic search is basically using the embedding API. Yeah, I know about the embeddings. I I created some simple rag systems with embedding uh models before. Hello Nares, nice to see you back again. As mentioned in the video that Chris is watching, Claude code does structured search of the codebase. Yeah, I just remember that was kind of difference when I came from cursor and I switched over to cloud code like tested it out for the first time. Uh looking at it using GP and stuff to finding context in my projects was a bit different. But since my project was very small, uh it worked pretty well. So basically his so far is just talked about what claude code is. Uh but for me claude code uh is for me it's not only about writing software. I use it for a bunch of other things too. I think if you're a bit open-minded, uh you can be very creative with a system like Clo. I know a lot of people here use it for like they connect it to their notes app, their obsidian, their notion. So with MCPS and stuff, there's basically you can be very creative and you can kind of link everything together if you just find the right tools. MCP servers, we have hooks now, and we have the sub agents. So, at least for me, that doesn't like do code professionally. I can use it for a lot of different things uh in my business. That's why I've been kind of looking at the sub agents more lately. Okay, so let's see what it goes into now. Romesh, we are for we are past the 10 minute part. And then on top of that, and this kind of matters across all these, is once again remember Claude is great at the terminal. And that means it's going to be great at all those different CLI tools, things like Git, Docker, Big Query, things like that. I never have to worry about, oh, I'm going to get myself, how do I get myself out of this sticky rebase? I'll just fire up Cloud Code and tell it the situation and be like, hey, can you fix this for me? It's incredible. Now, let's talk about best practices. And the first one is not going to be a surprise, but the first one is use claw.md files. So, remember that cloud code, like I said, is an agent and it has some tools, has some lightweight instructions in the prompt, but it doesn't really have memory. And so the main way we share state across kind of sessions or across our team when we fire up cloud code in the same codebase over and over again is this cloud.MD file. >> Yeah. I don't know. Do you use claw MD files in all your projects? Do you create like a custom one for every single project? I have kind of played around with that. uh for smaller projects I don't use it too much but for my video platform project I use it and for my productivity productivity project I use cloud MD I use it more like a memory so I kind of save some time uh when it comes to adding and updating MCP server ers and stuff like that. So Mick says augment code is very good second choice after cloud code. Yeah, I hope we can hope I can get some access to it and try it soon. So when we start cla what happens is if there's this claw.md file in the working directory, it's just plopped into context. It's plopped into the prompt and basically what it says is, "Hey Claude, by the way, these are important instructions the developer left for you. Be sure to pay close attention to this. And there's various places you can put the cloudmd file. You can put it in a project and check it in so all your teammates share it. You could put one in your home directory if there's things you just want claude to always know about regardless of what you're working on. I don't think I would have like a global cloud MD. I guess if you have some things could be useful in the global cloud. I guess but I have not used that. So dreams lol you use a specific code style like uh use arrow functions instead of normal functions. Okay. So dreams uh do you set that to like a global cloud MD? Hello Ben. Uh I think one for each projects will be better. Yeah, that's kind of what I do. So, Stor says open code also work uh pretty well for an as an open CLI alternative. Yeah, I've seen people use that too. Can you see it or can you plug in like any model into open code? Can you use like open router? Yeah, you just you set them up for the projects. Uh dreams. Yeah, that's what I've been doing too. Okay. So you can do open router in open code. I guess the performance is like dependent on how good the model is on uh tool use then and yeah some models are not so good at using tools and stuff. And the things you put in here are things like, "Hey, by the way, maybe this is how you run the unit tests." Or just so you know, to make kind of your searching and life easier, here's like just like an overview of kind of how this project is laid out, where the tests live, what different modules are, things like that. Or here's our style guide, all sorts of things like that to just make Claude's life a bit easier. And you can build these things up over time. The other thing you can do, which is important, is permission management. When you're running Claude Code, there's all sorts of different kind of permission things flying by. Kind of out of the box, what happens when you start our tool is for read actions. If Claude is searching or reading, we just let it go. But once it starts writing or running. >> Yeah, I guess we have the if you go to claw code here and on kind of my uh I have this uh this part. So this is the settings local JSON. You can see this is just getting added to the allow permissions. I rarely do any manual adjustment to this file, but um I think every time you just say always allow this, it just get added here into your local settings. Uh I sometimes add some permissions allow all time here. Uh, but I haven't really explored much with the dangerous permissions flag. You can place the info about your machines from installed apps. Uh, in Okay. Yeah, I never thought of that. Uh, but I guess you can also add open models to claw code. I think you can. Yeah, I haven't really explored that too much. I think it's going to optimize for the claw entropic models, but uh it maybe it should be possible, right? Bash commands or doing things that could change change stuff on your machine potentially. That's when we kick in this UI and it says something like yes, yes, always allow this or no, um, I want to do something else. And using that permission management and being smart about it can help you work faster. So there's something called auto accept mode where if you're working with cloud code and you press shift tab, claude will just start working. There's things you can do like you can configure claude in the settings where specific commands like on bash like if you just are like tired of saying yes run mpm run test you can just always approve that. So fiddling with your permission management is a great way to kind of speed up your workflow integration setup. So one thing that is going to help you get the most out of cloud code is remember that it's great at the terminal and if there's applications that you use which have kind of a way to access them through CLI and GitHub is a great example of that. They have a powerful tool called GH you can basically give more work to cloud code and you can do that either by just installing more CLI tools or you can attach more MCP servers. Um I >> yeah I always go for MCP servers. That's kind of my thing. >> I would say just through experience that if you're using something like um a CLI tool that's well known and well documented and you're trying to choose between the CLI tool and just installing it on your machine and grabbing an MCP server, I would recommend using the CLI tool. Um, and then also if you internally have your own tools at Anthropic, we have something called coup that does a whole bunch of stuff for us. You can also tell cloud about that and you pro that's the sort of thing you'd put in cloud to MD and then context management. So remember that claude is an agent and when it's an a what what it does it's calls these tools and the context builds up and up over time and at least for anthropic our models have a context window of 200,000 tokens and you can max this thing out. So you >> yeah we kind of talked about this yesterday the kind of how to manage kind of clearing out the context compacting the context and you can disable uh the autoco compact and there's some performance uh degregation when the context window is almost full and when you always when you kind of see the context is approaching the limit, you should always do like a clear or start a new session. I try to do that, but uh I'm not so good at it, but I haven't really I haven't really any measurement or any tests to do how much performance you lose when you fill up the window. I'm not quite sure. But uh I try to be aware of it. Uh so Stor says it can help drive cost and also help manage masking hiding data. Uh you could for example use sub agents for certain tasks that leverage leverage local models. Yeah, that's pretty interesting. But uh I don't know if you saw yesterday we looked at kind of selecting models for for the sub engines. Now I guess that's a better step in the right direction. So you could pick the haiku for a simple task compact according to what your next task is. Yeah, we looked at that yesterday. There was some guy just doing kind of yeah he was doing kind of this uh slash compact but uh he kind of take uh from from the context we have now just compact what we need for the next phase of the project. So just leave out that is things that is not relevant for the next operation we're going to do or the next phase or the next sprint I guess. So that is also something interest that could be interesting to try out. You kind of have two options when you're in a long session with Claude and you're working and you're going back and forth. You'll see in the bottom right you'll start to get this little warning that'll say, "Hey, you're starting to fill up the context window." And kind of depending on what's going on, you have two options. You can run slashclear and just start over and that clears everything out except for for instance cloud.mmd or you can run slashcompact and what'll happen is basically it's like a user message is inserted and it just says something like hey I need to go summarize everything we've been up to. I'm going to give this to another developer and they're going to pick up where I left off. And then that summary is what kind of seeds the next session. You can go from there. We spend a lot of time tuning this kind of compact functionality so that as you max out the context window and then run compact, you can start back over and keep going. Hello, Nick. Uh, bro, I'm loving the Cloud Code Max. Thank you. Oh, that's good to hear. Uh, have you been uh did you go for the 5x plan or the 20x? So one thing when you do like this slashcompact I don't think there's any way to look at maybe you can look at what the the summary of the context window is. I haven't really checked that out if you can do that. Uh yeah the 5x. Okay. Yeah I'm on the 5x2. The $100 plan. Uh, I really like it so far too. And I haven't really had for my uses, I haven't had any like issues with limits so far, efficient workflows. What can you do with cloud code and how do you get the most out? So, using planning and to-dos. I talked a little bit about this before, but one of the best things you can do is when you open up cloud code, instead of saying, "Hey, I need you to fix this bug," you can say, "Hey, I have this bug. Can you search around, figure out what's causing it, and just like tell me a plan how we're going to fix it? >> I don't know if a lot of you uses this, but I try to do at least sometimes I try to do shift tab tab for the plan mode. I tried to use this and I've been pretty happy with plan mode, but I don't know if it's super necessary, but uh I try to use it if I'm going to start like a new project in cloud code or you can bring in like a plan from Gemini or something like that. And this can save you a lot of time because you can verify, you can read Claude's plan and you can verify what it's going to do. And then the other thing that we have is we have this to-do list feature. So often when Claude's working on a big task, it'll create a to-do list. And if you're kind of paying attention, you can kind of watch this to-do list. And if you see anything kind of weirder in there or something that doesn't make sense, that's when you can press escape and say, "Hey Claude, let's change the to-do list. I think you're on the wrong path." Yeah, the to-do list is something I like. I I don't know if I have any to-do list here now, but uh uh do I I don't think so. Yeah, I guess we can see it here. So, I like the to-do list. I'm a fan of that. So, Mick, you had the $100 uh and you did not hit the set limits. Yeah, the Opus is a different story. Uh I yesterday I hit the opus limit after like just Yeah, like you said maybe 15 20 minutes. Plan mode just stops cloud code from starting to starting to add or change the code. Yeah, it's like a damn That's a good description. Yeah. So when you run plan mode, it's not going to change anything uh on your files right? It's just going to try to yeah write the plan and get ready to start working. Do you feel any difference between Opus and Sonnet? To be honest, uh you don't? Since I don't have so much experience with Opus since I don't use it too much, I can't really say. Uh if someone has like a uh I had this annoying issue here the other day and I switched to Opus. I think I switched to Opus. Uh after a while we did solve it, but I don't know if if it was because of the model. I'm not quite sure. Sonet was being uh obtuse today. Yeah, I guess other people that use it more maybe have a different experience than me. I try to use Opus for plan mode. That is kind of my workflow. Smart vibe coding. So, it's very tempting and it's very powerful to just let Claude work and press enter and see what happens at the end. I think there's a few things that can help make this better. And there's I think a talk later today about just this for 30 minutes, but doing things like having test-driven development, having Claude make small changes, run the tests, make sure they pass, always having Claude do things like check the TypeScript and the linting, and then commit regularly so that if it's kind of going off the rails, you can always fall back. Yeah, I think the commit regular is very important when you're using like an agentic uh CLI system or any agentic coding. I try to use that all the time, right? And he talked about the testdriven development workflow. Uh I tried that last week and it was pretty cool. Uh very slow. It ran for like 6 hours to build like a pretty simple app. Uh but uh it ran for yeah I think it was about 5 to 6 hours but uh it was pretty cool to watch. I use opus in the planning phase. Yeah, me too. I can say that it's hard to notice the difference between opus and set when you switch from set. Okay, but when you switch from opus to set you start seeing it making more uh mistakes. Okay, that's some good insights. spotfather. So remember with Sonet if you need to tell it think hard, think hardest or ultra think. I haven't tried that yet. Maybe I'll try it afterwards. Test driven is cool for backend stuff. For front end it's painful. Okay, so you got your $200 plan and pushed it really hard around 3K simply counting tokens with CC usage. Yeah, that you've been pushing it really hard. You extracted some good value then. Pretty cool. And try again. You can use screenshots to guide and debug. So Claude is built on top of our models which are multimodal. You can always just grab a screenshot, paste it in or if you have a file some >> Yeah, I lo use a lot of screenshots. I think it's like a nice um addition. So I try to use like screenshots and uh like I've written uh problem statement too if it's possible to screenshot. Of course I like to do that. I don't know if it helps but um if it's possible I try it's so easy to do a screenshot and just add some textual explanation of your problem too. Okay, so you can do screenshots on Windows. That's a bit annoying. Backend TTD is really nice. Yeah, I have to think. I I tried to I need to explore the testdriven development uh workflow a bit more. So maybe I should skip it for the front end part. That's pretty That's probably a good idea for the next project. I'm going to try that somewhere that's an image. You can just say, "Hey, Claude, look at this mock.png and then build the website for me or whatever." And then advanced techniques. So, as you're getting used to using Claude, what are some things you can think about uh to kind of push things to the next level? And one of the things we see both internally and with customers is when you've started to use this tool for a while, it's going to be very tempting to use multiple clouds at once. And so I know people at Anthropic and a few customers that run four clouds at the same time. There's various ways to do this. You can have it in T-Mo or just different tabs. I don't have a lot of experience with running multiple tabs. I have done it a few times. But do you do multiple tabs on the same project? I have done mult I do multiple tabs on different projects. So here I have my the app I'm working on. And here I have my more productive setup. So I can do my um yeah I can create my videos here. Uh I can do my stream agent find trending topics. So, this is more of my productive setup. And here in the other tab, I kind of have the app I'm working on. So, and Ben, it's more objective doing back end. So, that will make sense. It's better. Yeah, I see what you mean. I do multiple for different projects quite often. Definitely. Yeah, that's what I've been doing. Screenshots on Windows cloud code. Uh I'm using screenshots on Windows cloud code via VS Code. Okay. But one thing you need to save the screenshot. Yeah, you you have to drop it. You can't copy it. I saw a nice workflow using it uh with cursor as well. Cool. Yeah, front end would benefit from a really good computer used for automated testing. Yeah, a lot of people recommend me I should use the Playright MCP to test out apps. Uh I think maybe I tried it like months ago, but I because you can just go to the local host right with playright. Uh I don't know. Uh this way you can just uh continue working instead of just waiting for the answer. Okay, Sean says no. Okay, so use the puppeteer MCP for automated front end testing, but it's very slow. Yeah, that's kind of what I was thinking. I guess if you had like a station that was just doing that, but if you do it on kind of your Can you do it headless? I guess you're already doing it headless, I guess. But I haven't tried it too much. All sorts of crazy things. So I would challenge you to try getting multiple clouds running at once and kind of be orchestrating all these things. It's quite fun. I can only do two, but I know people that do four. Use escape. So escape is your best friend. While Claude is working, you can kind of keep an eye on what it's up to and you can press escape to stop it and interject and say, "Hey, I think you're going on the wrong path or I want you to do something else." Knowing when the right time to press escape is versus just letting Claude figure it out is >> Yeah. When you escape, hit escape. I was wondering like, do you kind of keep the state of where the agent was and you can just continue from that state? I'm not quite sure about that, but I think that's the way it works, right? Okay, botfather. So, it's headless. Uh, the latency with puppet MCP is probably caused by the LLM calls. Yeah, it has to be that headless will kill you if you need to do real stuff that checks for bots though. Okay. iPad kids will be having eight clothes running at the same time. So, Stor says he has a really good experience using testdriven development and unit test for JavaScript projects. Uh, I've been thinking about the workflow where the agent reads the Chrome logs uh and then can do debugging as well. Yeah, I saw there was some um we had a quick look at like a Chrome MCP server that has access to the console logs. Sitter escape once, interrupt execution, keep states, ex escape twice, uh select previous state to return to. Okay, so it keeps the states and you can also pick the Yeah, I haven't really played around with that, so that's good to know, Sean. Thanks for that. Yeah, Stor, if you try it, let me know how well it works. The Chrome MCP key to getting the most out of the tool. And there's a hidden feature. Not too many people know about it, but if you press escape twice, you can actually jump back in your conversation. You can go back. >> He just said it. That was pretty funny. >> And you can kind of reset tool expansion and MCP. So, this is taking it to the next level. If you feel like with bash and with the tools that cloud has that it still can't do something, this is when you should start looking at MCP servers and then headless automation. I think this is the thing we're most excited about, but also we are still trying to wrap our heads around internally, which is how can we use claude programmatically. >> Yeah, that's pretty interesting. For me, this is kind of now it's starting to just get a bit interesting, right? Uh, I think if you if you're going to build something very productive and like very useful, you have to take advantage of the tool expansion and the MCPS at least. Yeah, you have to do that, right? That is kind of what I think is the best value with cloud code. Of course, you have kind of the native models, but uh being able to kind of integrate all these tools and MCP servers together, like we have the hooks, we have the MCP, and now we have the sub agents, too. Uh now it's really starting to get interesting right? Puppeteer MCP can operate the Chrome browser and has access to console logs. Yeah. Okay. Headless Chromium. In a puppeteer config file, you can configure headless true or false. Doesn't it have to be headless? At least if you're on the same computer, right? I guess if you had like a machine on the side, you could kind of keep it uh not headless, but uh it's very annoying. >> We have that in GitHub actions. We want to figure out other creative places we can start using it. I would challenge you all to do the same. So, with that said, uh I'm going to jump over to my computer because there's one other best practice, which is it's always good to stay on top of everything that's new. So, we're shipping super fast. I'm going to throw I'm just going to go over >> before he starts here. Let's try to run my So if you go to my productive setup here on cloud code uh we can do exit and we can do cloud uh cloud. I'm not 100% sure if all the API keys are up to date here. So it might be an issue. Uh but I want to try to run my trend server agent here. or let me go to agents. So if I do like chaining to agents now I can do first you can run the trend surfer agent and after that you can chain in the video producer agent to create a short video uh mostly based on exposts. Let's just run this in the background here. So I want to first find a topic. Then I want to kind of chain in my producer uh agent. So we are starting up the trend server agent here and we're going to use the the findings from this to create a short video. But we have some issues with the Grock API today. I think my API key has expired. But I'm just going to let it run in the background. As long as the software allows pro programmatic uh use headless uh should not be needed. It will be cool to look at multiple agents operating multiple browser instances at once. That's pretty interesting. Hello Cesar. Does anyone know why I can't see uh tool agents? Do you mean um Ces or do you mean in uh if you go to slash agents or do you mean here you can kind of see the agents that is in your project and if you do slashMCP you can kind of see your connected MCP servers. slash agents. Yeah, slash agents. And you you of course have to create one and then you should see them down here right afterwards. Or you can go to your um your project and you can go on cloud and you should have them under agents here. Should be pretty straightforward to find them. A few things that are new as of today. Um, one thing is when you're in cloud now and you fire it up, you can do slashmodel. You can see what model you're running on. I'm on default, which happens to be sonnet. We can jump over to opus. You can do the same thing in /config. Switch it here. So, that's new. Make sure you're running the model that works for you. There's another thing that's new about these models which is you can say something like um can you figure out what's in this project and for a long time for a while we've had this like think okay so yeah here we have this uh keyword think hard or was it think harder and was it ultra think haven't tried that yet >> hard or extended thinking now this is great but with our past models the we won't let our model think between tool calls and that's probably when the thinking matters most. So starting with cloud 4 they can now our models now think between tool calls and we can watch this happen. So we have claude in this project. There's a few different files in here and I'm just going to tell it to think hard and figure out what's in this project and we can watch Claude start to work. And so the way you know you triggered thinking is you'll see kind of this lighter gray text. >> Let me try that. So if I uh if I uh explain this project so let's see what model we are on first. So if I do slashmodel so I want to be on set and I do explain this project. So this is just my simple meetup app to me and what I do ultra think. So let's see if we trigger kind of the thinking part here. Yeah, we do. So now we kind of trigger this thinking. I think I've tried it before. But sometimes it's just going to do this autonomously. I think think uh yes. So you have uh triggers that work uh on its thoughts. It will pick up clues such as lack of context, bad prompting, vague instructions, good use of sub agents, uh, reading its own thoughts. Yeah, maybe I have to explore a bit more with this keyword or this phrase. Haven't really used it too much to be honest. >> And then it'll call some file, it'll call some tools, it'll read some stuff, and then we see some more thinking. And this is awesome. Um, so I encourage you when you're working on tasks and solving bugs, throw a think hard in there. And then >> or ultra think like I did. >> And the other thing, and you know what, we'll just throw it up real quick, is I have this in VS Code, but of course this is in Jet Brains as well, but we have these new great integrations with VS Code and and Jetrains. Um, we can do things like Claude's going to know what file I'm in. What file am I in? That is not what I meant to say, but Claude's going to figure it out. And you can do things like this. Haven't really played around with this feature. So, this is the connect to ID feature, right? So, we can connect it to Cloud Code. How hard is it to do? Yeah, I have some issues with my APIs here. So, I might fix that later. But, uh, let's clear this. I just want to see how easy it is. So, if I Let's open up a new window here. Let's just um I think I have a CD uh what do you call it? CL computer cloud uh computer. Has anyone an experience experience with connecting to the to cursor for example? I haven't tried it. So how do you do it? Is it ID installed ex installed extension to cursor. If you install a claw extension in VS Code, uh you get a diff at list. Okay. How do I do this again? It's not here. I'm not so used to uh on the Mac here. You just open the terminal uh cursor and run cloud. Yeah, it says installed extension to cursor. Okay to connect. Okay, I see. But is cursor is does has the kind of the same extension setup as VS code? Can you do extensions in cursor? I don't think I've done that. At least not on Mac. I've done it on Windows. Uh yeah, I might try it later. There's no rush. Okay. So, it's the same extension. Yeah, I might take a look at that afterwards, but I'm not quite sure if it's the same as VS Code. Okay. I'm not quite sure if maybe it's pretty interesting. So these are the sort of things I would encourage left panel. But that's just a shot, right? Yeah, I haven't uh haven't really haven't really ah here it is. Okay. CL code this one. like this. Okay. And that's so that's it. Okay. Good. So I could uh So now I'm supposed to be able to do something like like this. It says one line selected. Hello Benjamin. Nice to see you. Uh, I'm good. What about you? Print line. Okay. So, it is working. I just going to print a line. I meant the line selected. Okay. So, you can see we have selected this line. Yeah, it looks pretty good. That was pretty easy. Pretty easy to setup. Yeah. Oh, that works. Can I integrate code uh in Intelligj uh idea? Which plugin? Hello, Masha. Uh can I integrate cl code into IntelliJ IDEA? Uh with which plugin? I'm not quite sure. Maybe Sean or some other guys here know. Like I said, I just I haven't really integrated cloud code yet into so many ideas. Uh okay. Okay. Pretty interesting, I guess. We'll see if I going to use it more. Hello, Joseph. There's a similar plug-in for IntelliJ as for VS Code. Okay, we can check it out. Junior Yeah, maybe you can find it. Uh, Musha, maybe there's some information on cloud codes uh in documentation. But yeah, let's finish this video. Uh, 5 minutes left. So, the native jet branch aentic AI for Intelligj is called Juny. Okay, I didn't know that. encourage you to stay on top of. We have a public uh kind of GitHub project called Claude Code under Enthropic. You can post issues there, but we also post our change log there. And so I check this once a week and make sure that I'm on top of all the new stuff we're shipping because even I can't keep up with it. So with that said, we have like four minutes left. I'm happy to answer questions about anything cloud code related. We have it here. >> Yeah. So they do make a lot of updates on cloud code. I just check it on uh cloud for example cloud code uh github they have this uh change log here. So they made some changes today I think fixed shell environment setup for users fix connection stability stability issues. So the change log is pretty good here. Hello cont. What's on the agenda for today? Uh Benjamin, we have watched this claude code best practices video that came out yesterday. Nothing really new here, uh to be honest. There's also a cloud code plugin for Intelligj, but I don't see updates for it anymore. Okay, that's a shame. So, we're just going to watch this video here. Just a few minutes left. >> I can live demo some stuff if you're interested. Um, let's do a few. >> Thanks. Real quick. This might be obvious, but multiple cloud MD files in a project. I presume that's possible and it just figures it out or no? >> So, there's a few options of course like in the same directory. You couldn't here and one in a subdirectory. And I think we changed this so that all the subdirectory ones aren't read in because like Anthropic we have a monor repo and people would open it at the top and blow up their context with all the cloud MDs. So we encourage claude when it's searching around and it discovers cloudMD files in um child directories that are relevant to be sure to read them. But by default it just reads the cloudmd file in the current working directory when you fire it up. And then also you can set one in like your home directory. Um there are things you can do though. We have this new thing like in your cloud MD you can start referencing other files. So you could for instance um do something like this with an at sign um if you have other cloud MD files that you just kind of know you always want to read. >> Okay, that's pretty interesting. I didn't know. Let's say we did let's say we had our agent here uh let's create uh if we do like a claude MD could we reference my agents here. Uh, let's do the email outreach agent. At And if I do include MD, if I do read cloud MD, will it tell me the properties of the agent? This appears to be a reference to an email outreach agent who aligned with projects. Uh okay. Haven't played tried that too much. I'm not quite sure. >> Dan, um, to do something like that. >> Hi. Okay. I, um, have not had luck getting Claude to respect my Claude MD. Like there's one thing particular. Yes. >> Where I'll ask it to refactor something and then it will leave inline comments explaining the like the what of it is and it's like like something that's extremely obvious. >> And so I'll tell it like go and remove any inline comments that describe the what of what's happening and then it will remove it and then immediately do it again and like the same pass. So do you have any strategies for dealing with that? Yeah, I heard a lot of people had issues with the cloud MD file and uh cloud code not following the instructions you gave it in the cloud MD file. So I see a lot of people complaining about that. So there's kind of two things that fix that. So that was actually kind of a model problem. There's nothing in the prompt. We have actually a lot in the prompt for 37 that said whoa, do not leave comments. And despite that, the model just loves to leave comments. Um, so it doesn't surprise me that your cloud MD didn't help much either. >> Yeah, that's probably all the training data, right? So it's overriding the system instructions, I guess. >> We already did a lot I did a lot of work to try to tamp it down from what happens out of the box. So we mostly fixed that in Cloud 4. Now there might be some new weird behavior quirks, but the other thing we made better in Cloud 4 is it's just better at following instructions. Um, and we've gotten a lot of feedback from early testers that, uh, all of a sudden, whoa, my cloud MD is being followed way more closely. Um, and it might be a good chance to go look in your CloudMD and decide, do I still need this stuff? Maybe I can take some of it out. Maybe I need to add a few new things. So, moving over to the new models might be a good time to take another look at what's in there and see what you need and what maybe can go. >> Uh, for the record, I'm trying to think of something that you might not have thought of. Yes, we're doing multi- aent execution and parallelization. Can you make it so that for four agents, say agents two and three use the context from agent one, maybe agent four uses the context from agent two at a certain point. >> Yeah. >> Um yeah, etc. >> That's interesting. We're trying to So, kind of like I said at the beginning, we're trying to do the simple thing that works, which is just one agent. >> Yeah, that's going to get really complex like Yeah. that's great at coding and does everything. Um, I think we want to figure that out. Probably what's going to happen is if you wanted to do that, you would ask all your agents to probably like write to a shared markdown file or something like that so they can all kind of like check in and communicate. That's what I wanted to try afterwards. I want to try to create a sub agent that uh every time it's done like a task before it uh exits the task. I wanted to write like a log and write the log to a file. So I want to be able to see what it's going to pass on to the primary agent. Um, sometimes like I'll be working with cloud.md or claude and I'll just say like, "Hey, I need you to write some stuff in like ticket.md for another developer and then I'll fire up another cloud code." And I'll be like, "Hey, read ticket.md like another developer left this note for you. Like this is what you're going to work on." So, I would think about trying to write that state to a file and then just kind of like count on the model's ability to just like read files and make sense them. Um, is probably the best you can do today. And maybe we'll figure out clever ways to expose that uh in the product as something more native. Cool. All right. And with that said, I have some rare clawed code stickers that I found in my backpack. So, come find me. >> Okay. So, I guess that was the video. Yeah, maybe it wasn't too uh I feel uh people here and in the chat are kind of advanced users already. Uh so it wasn't maybe too uh many new things here to be honest. And I guess this talk is from what does it say on the Mac there? May 2022 or May 2nd 22nd of May. So some things has changed since then. Uh so yeah maybe not so much new information here for maybe people watching today. But um uh I think we had some good conversations around this and some good ideas from the chat. So uh I guess it was fine but maybe not exactly what I was hoping for. What do you think? Yeah, it wasn't that much new stuff here. Maybe you can use a memory MCP for the subbations to communicate with each other. Yeah, there has to be some way. But I don't know if is it beneficial. So what I wanted to try was just something simple. So I think I want to go to the I think I have like a very simple way to do this. But I can't really kind of remember the what did I call this directory? Uh yeah, I think it's this one. Agent test. Okay. Agent test. So what if I instruct the agents to we have the planner agent here, right? So if I go here and I look at the planner agent. So, I just like to copy this path here. And I just like to do this uh read this. And I want this agent uh before it uh completes its task or its tools, it should write a log of everything it has done into a file called uh planner.md log planner log.md. So you could just get claude code the primary agent here to update the agent. You don't have to do it um in the you can either do it manually here or you can just have open changes in cursor. Do you want to make this edit? Yes. So use the right tool to create a planner MD log before providing your final response. But this is going to work though. So I'm going to do this important colon. Use the right tool to create a planner. What kind of file was that? Planner log.md. Does it have the right tool? Uh yeah. Just want to see what's going to be. Yeah, I know. Sean, it's lazy, right? You're getting very lazy. But writing up this structure here, it's going to take forever if I'm going to do it myself. So, I just want to see what happens now if I run the do I like this integration into clo cursor? I'm not sure. So, I'm going to do let's plan a simple HTML app. We should just have like a simple uh Bitcoin price tracker that keeps track of the price in real time. We also need a tracker of the changes of the last 24 hours. And to do this, we are going to use the and I'm just going to do at aent sprint planner. So I'm just going to see now what we're going to get in our log file here. if we even get like a log file. I'm not quite sure because this agent now sub agent it's kind of work. We can't really get any insight into what it's doing. We see the tool uses I guess and we see the we see what kind of tools it use but we don't really see anything else. Initializing Huh? That was strange. Um, let's do an exit. Let me try again. I haven't really find a good way to understand what the agent is doing here. We can't the sub agent we can't really do any debugging if you kind of if you don't know what it's doing, right? Doesn't matter too much. Maybe. So, if you were here yesterday, uh we worked on my new app and before I ended the stream, a lot of you did test it and I found some valuable issues. Uh but I have corrected them now, I think. So, maybe we'll try it out. You have to capture what it's doing. What it's doing. Uh if you want to see it, it's making API calls. Yeah. But Sean, do you know like I want the agent before it kind of completes its uh task here. And it uses his own reasoning which is something else. So you can see. Yeah. But I kind of want Okay. So now it wrote So this is what I wanted. I want wrote the a I want the agent to write hooks. Yeah. Hooks is a way to do this. But now I told the agent to write like a log. So now I got the planning log here, right? So this is I hopefully this is what it passed on to the the primary agent. So here is the planning log. So this is what I wanted. So I can see what tool it used. Deliverables created. But is this good enough? I'm not quite sure. I just thought this was the easiest way to do it. I'm not sure though. Created a comprehensive sprint plan for an HTMLbased Bitcoin price tracker. So, let me grab this part here. Okay. Yeah, that was fine. Kind of want to do the same with the other agent just to compare it. So if you look at uh this is my agent for execution my sub agent. Let's create a logging system for the instruction executor agent. So before it completes uh the task, it should write a log of uh what action has been used uh to the file execution log MD. Just want to see what happens. I guess you could always but the thing is if if we run the let's say the sprint planner agent here and I ask kind of the primary agent what this agents did execute. It doesn't really I don't think it has the context unless this was passed along from this sub agent here. But I thought if we get this this sub agent to write the log before it ends, maybe we get more information. Is it useful? Maybe. But this isn't exactly maybe a log. It's more of like a plan, I guess. So Sean, what do you use uh hooks for? I haven't been the biggest user of hooks yet, but uh I did like a small deep dive and I think I kind of understand them now. Okay. So, let's see. We have the log here and what is going to be the plan. We have a logging protocol. Hello, Matthew. Gemini Deep Tink released. It is I have a drop plan. Let me check. Do you know where I find it? So, I have access to Flash and Pro. Okay, so it's rolling out now. I've been looking forward to Gemini Deep Think. I want to try that model. That's pretty cool. I'm going to watch out for it. Thanks. So the instructor execution sub agent will now log all actions before completing the task including timestamps, instructions, actions taken, outcomes, and any issues encountered. Let's put that to the test. Okay. So now we're going to use the to execute the sprint pan we created. So delegate the task to the sub agent. So what I want to see now if uh after this sub agent the instructor executor agent is before it completes this task it's going to update this uh a log file here what it's did what it did just want to monitor this and see how well it works. Did they announce something on uh Gemini? I don't know. Is it this? It's not this. Google, where did you see it, Matthew? Google deep mine. Oh, deep mine. Of course. Why didn't I think of that? Here it is. Let's watch the video. I'm just going to see how this goes. Just going to wait for the log here to see uh But I'm not quite sure if this is very practical. So someone mentioned yesterday that they think using kind of the this kind of principles for agents now is the best way to do it or agentic coding kind of the the kiss principle. Keep it simple stupid. So I wanted this weekend to look at kind of a workflow that is uh based on this simplicity because after the stream yesterday I had a more look at Claude flow uh Claude flow GitHub is it and it looks really complex. It's not this one. Where is it? Code flow. Claw flow. Here it is. It just seemed so very complex this. I know a lot of people said it was pretty good, but uh haven't tried it, so I can't really comment, but uh Okay, so here's our app. Oh, we have a dark team, too. That's pretty neat. Okay. So, I guess the app turned out fine. We have the tracker. We have the 24-hour change. 24-hour low. Is that correct? We have an update and a refresh. We have Okay, that's fine. But what I wanted to see if it's going to write some kind of logging. Okay, so it did update the log here before it ended. Okay. So created these files, did some API research and planning uh price fetching functionality. It's something. I don't know if it's perfect, but uh it's something at least it's more information than you get here. Or I guess Yeah, I'm not sure. Have to think a bit about it. Uh but let's let's watch them this one for researchers, scientist and academic tackling hard problems. Gemini 2.5 deep think is here. It doesn't just answer. It brainstorms using parallel thinking and reinforcement learning techniques. Uh we put it into the hands of mathematicians who explored what it can do. Okay. So let's listen in on this. I don't know how the sound level is going to be. I'm going to put it down a bit. Let me see. With my most recent experience with uh Gemini Deep think the answer was spectacular. This is a mathematical with my most recent experience with uh Gemini Deepth think the answer was spectacular. This is a mathematical conjecture that was made by some people some years ago. They didn't manage to prove it back then. They checked many cases and then they just left it as a conjecture. I asked the statement of the conjecture to Gemini, deep think, and it seems like it proved it right away with a completely different method. When I was thinking about solving that question, I was thinking about maybe three different things, three different ideas, but it seems that deep think was thinking about 20 or 100 many different possibilities and then pursuing them. The deep mode that the users uh will get access to is basically very similar to the deep mode that we use to achieve the gold medal standard at the IMO. So the achievement of IMO is um uh basically a a major milestone in artificial intelligence towards this uh building human level intelligence. Yeah, looks interesting. I'm not quite sure if this is a model for me personally, but uh I'm going to pay attention to see what it's being used for. But if it's going to be like um a very long uh iteration and loops of thinking or like uh reasoning, kind of hard to say what you're going to call it. Um but I want to try it out maybe for some planning, tackling hard problems. It's probably going to be expensive, I guess, since it's doing this parallel thinking. It's probably going to use a lot of uh compute, I guess. And did they say um when it's rolling out? Ready to try it? Uh it's rolling out on ultra subscribers. I have an Ultra subscriber. Um subscription. So, let's see if we get access. Not yet, though. Okay. So, This was pretty interesting, but I don't think I don't think um I'm going to explore this more today. Other things I was looking at was um just wanted to see if there were other any other Sorry about that. Any other interesting videos they dropped yesterday? Since these are from May, it's maybe not so relevant at least at least for I don't think so. But one thing I'm personally quite uh I thought was pretty cool was um of course I'm very biased here but uh yesterday there was this release. So Open AI is going to build um a data center called Stargate Norway here in Norway of course. I thought it was pretty interesting because even though in Norway we are not a member of the EU, we haven't really seen a lot of investment into the EU from kind of the biggest uh biggest um AI companies yet. I guess we have mistrol but uh haven't seen a lot of other than Google and stuff of course they have data centers but this is going to be a pretty big data center in uh northern part of Norway. The facility will target to deliver 100,000 Nvidia GPUs by the end of next year. So, I thought it was pretty cool, but I'm of course very biased on this this uh this part. And I think it's pretty good for Europe. Yeah overall I think. So, what I think I'm going to do is I'm going to try to finish up my application here. I'm just going to see. Someone says that there was some issues with the uh password checking. But I think I fixed it. So I can do password. Okay. So here's my event. I can look at my dashboard. Create a new event. Let's say we do uh stream or like a stream clothes giveaway. The event type is going to be yeah I don't know other. So I fixed kind of I added like a date selector. So this is a bit better I think. Uh giving giving away five pro claw codes subscriptions. Let's try to create this event. Yeah, I got the toast here. Event created. Uh let me try to join an event.com. Yeah, it's looking pretty good. So, if you have time, maybe you can uh help me check it out. Just try to join a few events. So, I'm going to link the app in the chat now if you have time. Either if you're going to create a new user, don't use uh don't use your real email. Just do like a mock email or password and password. I will have to change up the text and stuff here. And but other than that, I think it's pretty much ready now. So, I was thinking for next week, um, I'm going to do like a giveaway. I think I lost my chat here. It should be fine. So, I was thinking next week to do like a giveaway on uh Clo, but I'm not quite sure if it's going to be What plan is best? Uh, telling me that my first name contains potentially dangerous characters. Really? So, you use the that upper Okay. So, what name did you use? Just Tron or did you use like a special character? Really? Okay, that's not good. If I sign out and I do register. Huh. Okay. I have to fix I have to look into that. H maybe the security is a bit too strict then potentially dangerous characters. Uh okay, that was a bit too strict. Uh I'm going to note that down. Thanks. I'm just going to do like uh this. Where is it? Here it is. I'm going to do a note here. potentially uh dangerous characters uh in register. Okay, that's strange. Okay, I need to check that. Okay, Salvador, I can see you joined. That's good. So, I'm just going to run a poll here. Uh, what you guys think is the best way to do this? So, the app is just for uh Yeah. Hello. Nice to see you. The app is just for um it's just for me for this um this uh fall or late summer. Uh I'm planning to do some events. Uh it might be online events, it might be like uh physical meetups and stuff. So instead of just using like the other platforms, I just wanted to make my own. So I can create events and people can join. Uh this is probably going to be online events and uh uh physical events. Uh I might have to do a bit more like uh location maybe and URLs and so uh thanks. Yeah, this project was a part of trying to learn sub agents uh with cloud code. So I basically I use cloud code for this project with kind of the front-end developer agent. We have we created a security agent yesterday and yeah worked. This is a very small project, so I would say it worked pretty good, but if I think if I added more complex stuff, the agents I think would have uh had a bit more issues. Uh I'm going to do a poll uh because I want to do next week I want to do like a giveaway on stream. I might do like one giveaway. Either I'm going to do uh best giveaway. So either I was thinking doing uh one max uh one 5x max plan on anropic or I could do like five pro plans. So, I'm not quite sure what is the best if it's better to give away 5x max plan or to do one giveaway each day for a pro uh sub plan. Personally, I think maybe pro plans because then people can try clawed code uh if they don't have access now or they can just use Antropic then more people get to try it. Of course, I don't have any like I can't really give away the plan, but I can give away like I think the cost is is it so monthly. It's going to be I guess $20 then max users are willing to pay anyway. Yeah, I think so, Sean. So, maybe doing that five pro plans. Uh but of course, anyone can join, right? There's no restriction. If you already have a plan, you can join. Uh, everyone that's on the stream can join. So, it's basically just going to be PayPal, I guess. But, uh, like $20 for a pro plan. Let's say you don't. So, not everyone lives in countries that has good the salaries uh, compared to the pricing of the American companies could be a bit skewed. So, it's quite expensive to have a pro plan, but of course, anyone can join. Still no deep think model here. Okay, so we have few more people. Benjamin joined learned a learned learned a. Okay, that's cool. Yeah, it seems to work pretty good. Uh, but this event uh I'm going to do like a we're going to do something else, but I just wanted to see if the app works and I need a name for it and I need a domain. So, I'm probably going to fix that over the weekend. So, that's going to be pretty good. Yeah, this is a bit funny. I saw this post on claw code here. Uh, it always says you're absolutely right when you try to correct it. It's pretty funny. Claw artifacts. Might have to take a look at this new API from Antropic with the artifact part. Haven't really tried it too much, but I don't really know if I'm kind of the targeted audience. Is there any documentation here? I might check it out later. Let's see. Ma joined. Nice. So, I just need to think about should I do just a random draw giveaway next week or should I do some kind of uh I'm not quite sure. I'm going to think about this weekend. So, yeah. 75% says five plants is the best. And I think I'm going to go for that. I think so. So, yeah, this seems to work pretty good. And I'm just going to have to check up the register here if people having problem entering their name. So, I the only video I kind of catched my eye here was uh vibe coding in production, but I don't think I have time to watch this now. I wonder what he means about that. should use off and have them just log in with their Google. Yeah, I thought about that. Uh that's what I use on my uh video course platform. Uh I'm just going to I'm just going to use Google. But the idea um behind the app here or like the web app was that uh to join an event you don't have to create an account but to create an event you have to do that. So yeah, I might do uh Google oo off later here. Uh maybe, but I don't know how many people basically it's just for me. So I don't know how many people are going to use it to create events. But uh if I see like a need for that, I think you're right. would be better to just switch uh maybe to like a but it's a like a like I said it's just going to be very simple for my use cases and I kind of used it as a project to learn more about the the sub agents and it was pretty interesting. at least what I did yesterday with the uh the front end uh security validator agent. It worked pretty good. It created this uh did we get a log of that? I don't know. But I think if we do something like what was it? Let's see if I have it here. Is it mpn run test security? We ran this and we have 80 or 43 tests passed. Uh we failed on seven tests. So for the front end security part we passed. Yeah, this is like 80% right coverage. So I thought that was pretty good. We kind of instructed the security validator agent to set it up so all the input fields and stuff and password validation. Oh, here's the name validation, right? It should prevent from injecting SQL scripts and stuff into the the forms and fields. I don't see any other security issues with this app. There might be, but I'm not quite sure. I thought 43 out of 50 was fine for like uh let's say Sean or uh Sean, I guess. Uh what is kind of the if you if you work like at a very serious software company is there like a set do you need 100% uh cover on tests or is like n 90 90 good enough? Is there like a standard there or is that just different from every company? I'm not quite sure. I haven't really worked at like a serious I've done some consulting on different software jobs or companies but not uh I haven't really worked on any big projects from scratch. If there's like a we need at least 90% test coverage or something. Can I ask what chat GPT says? Is there any uh like um standard test coverage limit for like big software companies? Hello Bandis worked in a few big ones so it was 90%. Okay, that's pretty interesting. Is that like over average or maybe that's like average? I guess it varies from company to company, but uh Okay, pretty interesting. at least 90. Okay. There's no there's no cross industry standard. Yeah, I see different somehow up 90% average I guess. Yeah. And I guess part of the code base is maybe more uh vulnerable maybe. So maybe that's higher. I'm not quite sure. Just trying to learn. What allows um user to uh use their own email? I don't like uh I don't like it. I don't like it being linked to my Google account. Yeah, I see what you mean. Now it's easy. AI generates them. Yeah, I guess that's a good part. I just had a look when we did a test driven testdriven uh workflow experiment with the sub agents. I just had a look at kind of the the test the agent wrote. It was just crazy. It looked really complicated and boring, but it was pretty fast. So I can imagine doing that from scratch like 20 years ago. Can definitely see that how we structure our claw MD file and why agents and the cursor rules first to sit situate the agent with basic context by explaining uh the what and why of the app. Uh okay, that's pretty explain how to do basic development task, add packages, run tests, explain how any MCP servers should be used, playright for make progress on tricky UI task. Hm. Yeah, some of you mentioned that. Continue explaining how and where uh and where of our app debugging is like half the job. Yeah, again I don't think I need this comprehensive, but at least a lot of people is doing experiments with different kind of setups uh in like the claw MD files, cursor rules files. It's going to be interesting to see if it's going to be like a best practice, best standard way of setting this up. I feel now everything is up in the air. Like with everything in this space, nothing is really settled down. Memory files in sub trees for token efficiency. Cloud code reads memories recursively. starting in the working directory recurses up uh but not including the root reads any cloud MD uh when it finds it. This is especially convenient when working in large repos where you run cloud code in fu/bar uh and have memories in both fu cloud MD and in bar cloud MD. Claude will also discover Claude in the nested sub trees under your current working directory instead of loading them at launch. They're only included when they find them in the sub trees. Yeah. Isn't that what it talked about in the in the the video? I think so. If it's instructed to write test, it will include the relevant cloud MD files in the test slash sub tree. H cool. Yeah, it's going to be interesting to see if there's going to be like a not a protocol, but like a standard way of setting up your cursor rules cloud MD. It's probably just going to be individual based on uh the type of project you're working on, right? I guess so. But yeah, uh I think I'm kind of have to go soon. Half an hour I have to be somewhere. But um thank you everyone for tuning in. We had some interesting discussions about CL code and stuff. There's a lot of noise around saving and remembering using the simplest solution uh MDF files as a backup method uh but use RAM uh databases like RIS to drive uh slashsave restore so you never have to worry autosave. Yeah. Okay. That's pretty cool. I think there's going to be a lot of different ways to do this saving and remembering. Yeah, I think we know at the end of the year maybe. Yeah, we'll see. We have some more standards. But yeah, thank you everyone for tuning in. didn't do so much today but uh was some nice interesting discussions around different topics I think uh and the video I watched was a bit wasn't that much to learn from it uh unless seems like a lot of people here knows some advanced workflows um from before but uh a few interesting discussions I think and I think I know a bit more about how I want to set up my workflows going forward. So yeah, wish you all a great weekend. If I don't have any big plans, I might fire up if I'm just going to do some work during the weekend. Uh I might just do a stream if I'm just going to sit on my computer anyway. We'll see. Uh but next week uh at least we're going to do some giveaways and just hopefully GT5. Uh I'm planning to do like a stream uh on GT5. Yeah, have a good weekend. Salvador Sean, thanks you for tuning in. Have a good weekend and everyone have a great weekend and hopefully I'll see you again maybe this weekend or if not next week. Hopefully uh there's going to be some cool releases. Gift 5, maybe some other things. So yeah. Bye Jake Benjamin. Have a nice weekend. Enjoy your day. Bye.