ЭЗ

How I Use Claude Code (After 1,600 Hours)

41:30

How I Use Claude Code (After 1,600 Hours)

Ray Amjad 02.03.2026 124 просмотров 18 лайков

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

Level up with my Claude Code Masterclass 👉 https://www.masterclaudecode.com/?utm_source=youtube&utm_campaign=AzmnaoVP8sk 🎉 Use BIRTHDAY at checkout for a Claude Code Birthday discount. Learn the AI I'm learning with my newsletter 👉 https://newsletter.rayamjad.com/ To avoid bias, I've never accepted a sponsor; my videos are made possible by my own products... —— MY CLASSES —— 🚀 Claude Code Masterclass: https://www.masterclaudecode.com/?utm_source=youtube&utm_campaign=AzmnaoVP8sk —— MY APPS —— 🎙️ HyperWhisper, write 5x faster with your voice: https://www.hyperwhisper.com/?utm_source=youtube&utm_campaign=AzmnaoVP8sk - Use coupon code YTSAVE for 20% off 📲 Tensor AI: Never Miss the AI News - on iOS: https://apps.apple.com/us/app/ai-news-tensor-ai/id6746403746 - on Android: https://play.google.com/store/apps/details?id=app.tensorai.tensorai - 100% FREE 📹 VidTempla, Manage YouTube Descriptions at Scale: http://vidtempla.com/?utm_source=youtube&utm_campaign=6omInQipcag 💬 AgentStack, AI agents for customer support and sales: https://www.agentstack.build/?utm_source=youtube&utm_campaign=6omInQipcag - Request private beta by emailing r@rayamjad.com ————— CONNECT WITH ME 🐦 X: https://x.com/@theramjad 👥 LinkedIn: https://www.linkedin.com/in/rayamjad/ 📸 Instagram: https://www.instagram.com/theramjad/ 🌍 My website/blog: https://www.rayamjad.com/ ————— Links: * https://cognition.ai/blog/devin-sonnet-4-5-lessons-and-challenges * https://github.com/aparajita/claude-code Timestamps: 00:00 - Intro 00:19 - Your Imagination 02:14 - Mental Models 02:45 - Front Loading Context 03:08 - Overcoming Code Bias 03:54 - --dangerously-skip-permissions Sandbox 04:40 - Ghostty 06:08 - Factorio 07:03 - Obsidian with CC 07:51 - Voice to Text 08:24 - /add-dir 09:20 - Verbose Mode 10:01 - Hierarchical Claude.MD Files 11:25 - At File References 11:59 - Iterating on Claude.MD files 12:26 - Modularisation 13:40 - Deferring Searches to Subagents 14:40 - 'Why' for Subagents 15:46 - Speaking the Agent's Language 16:55 - Screenshots as PRDs 17:48 - Clarifying Questions 18:07 - /rewind 18:42 - Plan Mode 19:17 - Spec Developer 20:25 - Spec Developer for Understanding 21:07 - Explore Subagents Before Specs 22:03 - Verifying Plans 22:24 - Mermaid Diagram 23:45 - Separating Plan Sessions 24:07 - Saving Plans for Later 25:16 - Re-entering Plan Mode 25:40 - Avoiding Context Switching 26:12 - Compacting vs a New chat 26:43 - Documenting Failed Attempts 27:19 - Avoiding Long Contexts 27:41 - Turning off Auto Compact 27:55 - Context Awareness 28:25 - Ticking off Plans 28:38 - Auditing Context Window 29:06 - Disable Model Invocation 29:48 - Models for Skills 30:14 - Statusline 30:38 - Session Management 31:16 - Multi Proposals 31:41 - Resuming Subagents 31:59 - Custom Slash Commands 32:59 - Logs for Debugging 33:32 - Using Sonnet 33:44 - Using Well-Known Packages 34:00 - Exa MCP 34:08 - Official MCPs 34:32 - Mixing Models & Modes 35:26 - Headless Mode 37:26 - /insights 37:44 - Task Lists 38:13 - Worktrees 39:30 - Chrome MCP

Оглавление (57 сегментов)

0:00 Intro 82 сл.
0:19 Your Imagination 456 сл.
2:14 Mental Models 140 сл.
2:45 Front Loading Context 88 сл.
3:08 Overcoming Code Bias 181 сл.
3:54 --dangerously-skip-permissions Sandbox 162 сл.
4:40 Ghostty 364 сл.
6:08 Factorio 218 сл.
7:03 Obsidian with CC 174 сл.
7:51 Voice to Text 132 сл.
8:24 /add-dir 209 сл.
9:20 Verbose Mode 164 сл.
10:01 Hierarchical Claude.MD Files 294 сл.
11:25 At File References 130 сл.
11:59 Iterating on Claude.MD files 110 сл.
12:26 Modularisation 275 сл.
13:40 Deferring Searches to Subagents 207 сл.
14:40 'Why' for Subagents 259 сл.
15:46 Speaking the Agent's Language 251 сл.
16:55 Screenshots as PRDs 203 сл.
17:48 Clarifying Questions 74 сл.
18:07 /rewind 127 сл.
18:42 Plan Mode 140 сл.
19:17 Spec Developer 248 сл.
20:25 Spec Developer for Understanding 167 сл.
21:07 Explore Subagents Before Specs 190 сл.
22:03 Verifying Plans 71 сл.
22:24 Mermaid Diagram 313 сл.
23:45 Separating Plan Sessions 80 сл.
24:07 Saving Plans for Later 260 сл.
25:16 Re-entering Plan Mode 103 сл.
25:40 Avoiding Context Switching 129 сл.
26:12 Compacting vs a New chat 109 сл.
26:43 Documenting Failed Attempts 133 сл.
27:19 Avoiding Long Contexts 81 сл.
27:41 Turning off Auto Compact 52 сл.
27:55 Context Awareness 106 сл.
28:25 Ticking off Plans 52 сл.
28:38 Auditing Context Window 108 сл.
29:06 Disable Model Invocation 153 сл.
29:48 Models for Skills 98 сл.
30:14 Statusline 91 сл.
30:38 Session Management 149 сл.
31:16 Multi Proposals 89 сл.
31:41 Resuming Subagents 69 сл.
31:59 Custom Slash Commands 221 сл.
32:59 Logs for Debugging 129 сл.
33:32 Using Sonnet 44 сл.
33:44 Using Well-Known Packages 55 сл.
34:00 Exa MCP 27 сл.
34:08 Official MCPs 91 сл.
34:32 Mixing Models & Modes 199 сл.
35:26 Headless Mode 470 сл.
37:26 /insights 73 сл.
37:44 Task Lists 115 сл.
38:13 Worktrees 300 сл.
39:30 Chrome MCP 490 сл.

Intro

All right, so about one year ago today, Claude Code was released. And since then, I've used it for over 1,600 hours and have made dozens of videos about it over the last few months that you can go back and watch. But in this video, I will be sharing 60 tips that will make you a Claude Code Power user because it's becoming increasingly clear that this is one of the most important skills for 2026. Now, before getting started, I

Your Imagination

want to make it really clear that the only limitation is your imagination with Claude Code. So, Claude Code edited 95% of the video that you're watching right now. I use it for my video editing. It solved pretty much every minor problem in my life. So, for example, at my nearby Starbucks, it automatically disconnects you from the Wi-Fi every hour. So, you have to manually reconnect. I just told Claude Code, "Can you make a script to do that for me automatically? " People have even been exploring their DNA with Claude Code and making visualizations for it as well. Now, to prevent me from spending 14, 15 hours every single day in my room using Claude Code, Claude Code has been helping me find a girlfriend and also set up dates for me. Many people seem to think you can't do this for some reason. But in reality, you can because cloud code reverse engineered the API endpoints of a pretty popular application for me. And every 10 minutes it looks at all the new users who have signed up and classifies them according to this prompt here. And then it will automatically reply to any simple messages that it knows answer to and it will ping me on Telegram if there's a message that it doesn't have the answer to. Now, there are two benefits here. Firstly, there are thousands, if not tens of thousands of people signing up every single day that it's physically impossible for a human to read through all the new profiles. And secondly, you can avoid the algorithms of these applications because actually working against you instead of in favor of you because if they were really helping you, then you would leave the application pretty quickly and they would make less money from you. But the value of AI is firstly it can sift through a lot of data very quickly and secondly you can align it to your goals. So honestly this has been pretty successful for me and the few of the friends who have tried this approach out. So, I'm actually 100% serious because the only limitation to using Claude Code is your imagination. And the reason I like it so much is it basically feels like playing life on Minecraft creative mode. And if you want to learn more about this, then you can go to my Claude Code master class where I cover a lot of these ideas. It is the most comprehensive Claude Code class that you will find online. There's a sale going on right now to celebrate the one-year birthday of Claude Code. So, if you use a coupon code birthday at checkout, then you should get a

Mental Models

discount. Now, next up is you always want to be trying to update your mental model of what the models are capable of. I find that many people have old habits and assumptions and they're like, "Oh, I tried that a couple months ago and the model could not do it. " And then they never bother to retry again. And I find my biggest limitation are my bad assumptions of like, oh, the model would never be able to do that. But then when I actually give the task to model, it gets like 80 90 95% of the way there. For anything that cloud code struggles with or is not able to do, I make a note of that in a separate document. And then once there's a new model release, I basically try that prompt again. Another

Front Loading Context

mistake that I see people making is they try to frontload the model with a bunch of files that they think may be relevant for the task. But this can fill up the context window way too quickly and degrade the performance of cloud code. So it's much better that you let the model figure it out and search through the codebase to find any relevant files. And this helps because it means that you don't end up polluting the context window with a bunch of irrelevant files.

Overcoming Code Bias

Another tip is for avoiding code bias in claude code. So, one problem is that claude code can get stuck in existing coding patterns, which is really handy when it comes to adding a new feature because you want it to follow existing patterns in your codebase. But if the patterns in your codebase are really bad, then it can't really think outside the box and come up with new ones because it just ended up loading in all the existing ones into its context window. So often what I do is I make a brand new folder empty folder with cloud code inside and I say okay can you implement this feature like just make this feature in isolation and then once it makes that feature and it works in isolation then I would get it to copy over that information into the older project and integrate that new code into the older project. So when cloud code is stuck in a loop isolating the problem in a new clean environment can help it get a fresh perspective. Another problem

--dangerously-skip-permissions Sandbox

that I see happening is that many people use dangerously skip permissions. basically all the time on their local machine and they never actually go ahead and restrict internet access for the model to specific websites. So what you can actually do is you can enable a sandbox and then make a HTTP proxy and with a proxy script you can put in an allowed list of domains that cloud code will be able to access which may be relevant for your work and you can be sure don't contain any prompt injections. So you can either get cloud code to implement a script like this yourself or you can go to my cloud code masterass and download it from there. But the main benefit here is that you can have a white list of company websites or websites that you trust for particular tasks. So this is essentially restricting cloud code to fixed number of domains that are relevant for that

Ghostty

particular task. Next up is using ghosty as your terminal and then customizing it to use claude code better. So the cloud code team personally loves using ghosti and one of the ways that I have it set up. I can make a brand new tab pretty easily by doing ctrlt and then also rename any tab using keyboard shortcut that I have which is control shiftr. type in a name like the project name and if I want to have a sidebyside panel then I can just do control D or control shift D and then split the window pretty easily kind of like this. If you get cloud code search online and look at the ghosty docs and it can do all these adjustments for you such as setting up the control shift R to rename a terminal window and you can combine this with hooks as well. So often I'm working on many different projects at the same time. So I'd have like five tabs open and often claude code would be done on one particular tab or require permissions but I don't know which tab to actually go back to. So what I have set up in my root folders claude settings JSON file is basically a permission dialogue hook. When claude code requires a permission then it sends me a notification on my system and a stop hook which will basically show a tiny bell near the tab once cla code is done. So if I ask it to do something. So, let's say I just said hi. Switched over to another tab. Then you can see right at the top over here, there's a little bell icon to say that cloud code is done. This helps me make sure that I don't miss anything that is on another tab. Of course, a better option would be to have multiple sessions running side by side, but if you're at a cafe and your monitor is pretty small, there's only so much that you can have running side by side. I do also cover hooks in much more detail as well in my cloud code class if you're interested. Now

Factorio

when I have parallel cloud code sessions running, I basically treat it like the game factorial. Having one session, one task at a time is pretty slow. So what I usually do is I have three, four or five instances per project running in parallel. And the benefit here is that there's pretty much always something for me to do. Like I'm not sitting down and waiting for the agent to complete a task. There's always one agent that kind of has a question or needs attending to or like a plan that I could be developing or something like that. So across like three or four projects, I probably have 10 12 cloud code sessions running in parallel and there's literally always something that I can be doing. And one of the fun parts about this is that you can basically be spotting bottlenecks that cloud code consistently makes across different projects and sessions, coming up with any slash commands or skills to automate particular workflows. And the bottleneck kind of shifts from how fast can a single session of cloud code complete a task to how well can I design my pipeline. I can spot mistakes and improvements to cloud code more often and then I can improve my pipeline across all my projects. Something I

Obsidian with CC

started doing recently is using Obsidian with Claude Code. And the benefit is all your notes are basically just markdown files on your disc that Cloud Code knows how to navigate really well. And it can do stuff like make back links for you, generate graphic views, generate HTML files that like explain ideas better. You can link it to Skills and MCP servers as well to make like a live operating system in Obsidian. And I basically use Cloud Code to edit all my files for me and generate these kind of images as well. And as of yesterday, it's become even easier for Claude Code to do that because they released Obsidian CLI and agents such as Claude Code are just better at running CLI commands. So you can make Cloud Code even more powerful by giving it a skill that helps it use Obsidian CLI effectively. So I basically won't have to read through files as often and can discover back links automatically and just save on context. Now when giving

Voice to Text

prompts, I basically only ever dictate my prompts now. I very rarely type on my keyboard because I use my own application called Hyposper. If you set the accuracy to highest, then you will find that it is the most accurate voice to text tool. And many people have been switching over from bigger competitors for exactly that reason. You can try out for free using the link in the description for Mac OS and Windows. And a pro license is also included for free for those who join the Claw Code Masterass. So essentially, I would have all these sessions running in parallel and then dictate prompts to all of them and switch between them really quickly. And I find that to be much more productive than typing out anything. Now

/add-dir

because I am working on many different projects at the same time, I may implement something in one project that I simply want copied over to another project and integrated into that. So essentially I would get cloud code to add another directory and then do yes remember for the session. Now it has access to both of these directories as working directories. And I would basically say something like hey can you look through the agent site codebase identify how stripe has been integrated and then copy over that integration and functionality into hypersper plan this approach out ask me any clarifying questions if needs to be. So after saying that and going to plan mode I press enter and then it would come up with a plan to basically transfer features over and that saves it from having to figure out something from first principles that was already figured out in a different project. I've used this quite a lot because I've been migrating a lot of my projects over to trigger. dev for background jobs and workflows and I had it set up on one project and then I just told Cloud Code, can you copy that implementation from that project into this one? Now, when

Verbose Mode

the models weren't really that great, Cloud Code always showed you the thinking process of the model. But as the models have gotten better, they found that this is no longer necessary for most people because they don't have to like interrupt the model as much or steer it as much because more often than not, it's doing the right thing. Now you can turn this back on by going to slashconfig searching for verbose and then setting this to true. So by default it will be false but setting it to true will show you a token count on the right hand side over here. And it also shows you the thinking process grayed out kind of like this. And I just enjoy reading the thinking process because it makes me feel like wow this model is like really small or they've gotten really good recently because often the thinking process reveals something that I would have never thought myself. Now I make

Hierarchical Claude.MD Files

extensive use out of hierarchical claude MD files. So you have your one in your root user folder. So right over here, this is the one that applies to all projects. And then you have your project level ones. And then you can also make personal preferences ones that exist only on your machine and should not be committed to GitHub. And I do this even when I'm working alone for the different development environments that I have. So on my Windows computer, I always run cloud code in Windows Subsystem for Linux. I just have a local MD file which basically just says you're running in VSSL. And you have to remember that these are loaded top down into claude code. So your most important instructions should be in the highest relevant one. I also have cloud MD files across my entire project in different subfolders for different kinds of tasks. So rather than having a massive cloud MD file with things about NexJS, Mac OS like uh general rules and stuff. I split this up as much as I can into separate ones. So for example in my project hypo I have the root one and then I have a Mac OS specific one. window specific one as well. I have one for the back end. I have another one for a task management system. And the benefit here is that your claude MD files are progressively disclosed. So only the most relevant ones are loaded in. So if my task to claude code was only making edit to the landing page website of the application, then the Mac OS one and the backend one would never be loaded in. And that keeps the context window narrow, focused, and as sharp as

At File References

possible. To prevent my Claude MD files from getting too long, I also have references within them to other files. So, for example, in this project, I noticed that Claude code really struggles when it comes to handling custom vocabulary. So, rather than having this long reference file within the Claude MD file, I basically have it in a separate file instead, and it's only used when it's doing something related to vocabulary stuff. And this is really handy for situations where Claude Code doesn't really know how to work well with a particular package or library, or there's like a weird quirk inside it that it always struggles with. Once you figure out, just make a reference file and then reference it from your claude MD file. Whenever I

Iterating on Claude.MD files

notice cloud code make a mistake, then I tell it to find the relevant Claude MD file and make a change there instead of the root one. I try to have my Claude MD files as short as possible. And every time there's a brand new model update, then I look at what I can remove. I very rarely follow these like random Twitter hype threads where they just talk about like, oh, you should add all of this to your Claude MD file. I do share more Claude MD best practices in a free video in my mastercloud code class if you interested. Next up is I noticed that

Modularisation

many coding agents basically like to keep adding to the same file instead of making a brand new file and modularizing your codebase. So I regularly modularize my codebase because it just keeps my file shorter. If you don't do this and all your files are like many hundreds or like over a thousand lines, every time cla or another coding agent has to make an editor file, it will always first read the entire file. And this can be problematic if most of your files are really long. So for example, you can see I came up with a plan and I started implementing the plan over here and then it's like okay, let me just start reading all the files I need to modify in parallel. It ran through all these files, some of them being 700 lines. And after reading through like 2 3,000 lines, the context window ended up being about 47% full before it decided to make any changes. And the main problem here is that all my files are really long. If they were much shorter, then claude code would just be reading shorter files instead and the context window would not fill up as quickly. So I often get times where I tell claude code, can you identify all files that are over 500 lines and then make them smaller or modularize it more? And that means anytime I make changes in the future, claude code does not like pollute the entire context window immediately before it started making changes because it had to read a bunch of files that were really long. Now I'm also aggressive in

Deferring Searches to Subagents

deferring any searching related tasks to sub agents. Now cloud code does this automatically in planning mode because it searches for your codebase to establish patterns and finds out which files need changing. But I find that outside of planning mode, if you get it to quickly check some docs or something or like try to find in the log file the error that just happened, then it will load that into main context window which will quickly fill it up. So it often jumps from like 30% to 55% because it just loaded in a bunch of log files. So in my user level cloud MD file, I basically have this always and aggressively offload online research eg docs codebase exploration and log analysis to sub aent. When you're about to check logs, defer that to a haiku sub agent. For complex problems you're going around in circles with, get a fresh perspective by asking sub agents. And I find that this means I can stay within the optimal range of the context window for longer, which is between the 0 to 50% mark. Now, since that is in the user cloud MD file, this will apply across all the projects on my computer. I also

'Why' for Subagents

get Claude code to include a Y whenever spins up a sub aent and I found that led to better results especially when you have multiple sub aents running in parallel in different parts of the codebase. So without a Y if you were adding a off related feature then cloud code would often just search how off works with an explore sub aent and then you would get pretty generic results. If you have multiple sub aents they may be overlapping and if your authentication system is really big then you may end up getting a lot of noise. But by including a Y or a specific purpose for each sub agent in its system prompt, you now have how off works for rate limiting because you're improving your rate limiting. And that means that you will have less overlapping summaries. If you do have multiple sub agents, the right stuff will be filtered and returned back to the main session from the sub agent. And overall, you just see better results. By default, cloud code doesn't really include a Y. So, uh you can see this is an explore sub agent that is spun up. And there isn't really any Y over here except this part where it says to understand what a scraper would see versus a real user. So in my user level cla MD file I have when spawning a sub agent include a Y in the sub agent system prompt because it will help it filter the signal from the noise. Now I

Speaking the Agent's Language

often try to speak the agents language as well. Though one of the benefits of having verbose output on and seeing the reasoning trace and reading for it is that you get a sense of how cloud code refers to different things in the codebase because the way that you refer to something may be different from how cloud code refers to the thing. And by you using like weak terminology or inaccurate terminology, it can end up wasting the agent's time when it comes to searching and it may just end up doing the wrong thing. So you may use the word download URL in your prompt, but in reality you actually have a couple different potential download URLs and the one that you were referring to was a CDN download URL. So if you read through the reasoning trace or just what cloud code is thinking, then it would say like I updated the CDN download URL. And then if you use that word or phrase next time, the agent will be faster and better at that particular thing. And this is the same reason why developers and also designers have an easier time using claude code because they know all the terminology that will help claude code better understand what they mean. So designer would explicitly refer to a button as like hey can you edit the secondary button or the danger button on that page whereas a non-designer may just refer to that as a button instead.

Screenshots as PRDs

Likewise you can massively shorten your prompts by having screenshots especially the right size screenshots. Often some people have a really small screenshot kind of like this in the top right and they're like, "Oh, this is wrong. " Like, "You should change this. " And Quad Code can't really find that part of the codebase, but if it was a little bit bigger, kind of like this, then that would be a good size because it can rely on the related text to find that part of the codebase more easily. But if your screenshot is like way too big, then it has no idea what you're referring to. So having the right size screenshot, especially for debugging UI related tasks, is pretty important. The same thing also applies to like using competitor references. You can have a big screenshot and then a smaller zoom up of the thing that you're actually referring to. And in situations where you find yourself going around in circles, you may just want to switch over to Gemini CLI instead because I find that the Gemini models have a better visual understanding and or better like just visual aesthetics compared to the Claude models. Next up

Clarifying Questions

I almost always ask Claude Code whenever building out a brand new big feature to ask me any clarifying questions about how I want something implemented. Now, it will do this sometimes. It will ask you up to four questions, but I find that is not enough. And I'm often like, can you ask me 20 clarifying questions using the ask question tool for how this should work? I also use the rewind tool

/rewind

a lot. So, when I find that Claude Code has made a mistake or it just kind of went down the wrong track, but the conversation up until that mistake was still really good. Instead of arguing with it and saying, "No, go back to the previous thing. You should not have done that. " I just rewind to the earlier part of the conversation and then just give it better guidance from there because otherwise you have the bad decisions lingering in the conversation that have essentially now polluted the conversation. So if you do slashreind then you can go back to any earlier points in the conversation and you can do something like restore code and conversation restore code restore conversation and summarize from here.

Plan Mode

Now, my workflow for adding any feature, no matter how small it is, is to start a brand new conversation, switch to plan mode by doing shift tab, and then describe what I want using Hyper Whisper, my voice to text tool, and then let Claude code explore the codebase. Often, it will ask me some clarifying questions as well, and then I review the plan. And if it was a really small feature, then I just let it proceed with the plan, implement the thing. And nowadays, I don't even check the code because like Opus 4. 6 text gets it almost perfectly every single time if the plan is robust and good enough. I know a lot of people just don't read the plan anymore, but like you should still be reading the plan so you know what's

Spec Developer

going on. Now, when it comes to much bigger features, I basically use the spec developer skill and that looks kind of like this where Cloud Code will ask you a lot of non-obvious questions, some of which are actually really insightful. So, what I basically do is when I'm working on a brand new feature, so for example, I have a tasks folder in this project. The very initial plan that Claude Code comes up with is often kind of short. So it's like 150 lines, 200 lines and it may not cover all the edge cases that I actually care about. So what I do is if I'm making a big feature when reviewing the plan for implementation, I often just reject the plan and then I trigger the spec and developer skill. It will then massively improve on the plan for a lot of questions and then my plan will often end up being like 5, 6, 700 lines, but it covers basically every single edge case that I care about. Now, I do go into this into much more detail inside of a free video in my Cloud Code Masterass called adding new features. And this is part of my uh daily workflows class, which I'm updating on a month-to-month basis whenever one of my workflows significantly changes. So, if you want to stay up to date with what I'm currently doing, you may find this class useful because it has a bunch of

Spec Developer for Understanding

free videos. Now, the biggest benefit of using that skill is you have peace of mind and you have an understanding of how your codebase actually works. You don't have to read through the code through the questions that cloud code has asked you. You already understand many of the key decisions that needed to be made in regards to adding that new feature. So I can explain how a lot of my codebase works at a pretty deep level even though I've never read through the codebase through the questions that Cloud Code decided to ask me when it came to developing plans. And this piece of mind is really important because before I started doing this and having those questions asked with each update, I was always kind of worried like, okay, is this going to suddenly break everything for end users? Like, but now because of the spec developer skill, I basically never worry about this anymore. Now, what I also do with the

Explore Subagents Before Specs

spec developer is that if I feel that it needs to understand the codebase a bit better before asking me those questions, I use a different slash command, which is spec developer explorer. And because I've saved everything into a tasks folder, often I would get Cloud Code to plan out four or five new features at the same time, save them all as tasks, and then load one of these task files into brand new chat. And then I trigger my spec developer explore, which is exact same as spec developer with one difference being uh read the plan, use it to spin up three to five explosive agents in relevant parts of the codebase to understand it, and then interview me in detail. And I find that these questions are much better than just generic spec developer questions. But of course, you can still use a generic spec developer if cla had previously explore the codebase before earlier on in the conversation. Now, one of the benefits of using the spec developer to make these really long plans is that after this big feature has been implemented, I

Verifying Plans

can then tag the plan file and then tell cloud code, can you spin up like three, four, five different explore sub agents to make sure that everything listed in that plan file had been successfully implemented. And then you can also tag a nova agent for example like codeex instead to just confirm all the changes required had been made from the long plan file that you used spec developer

Mermaid Diagram

to write. I also do have a mermaid diagram generator skill that is shared across all my projects that helps me understand how data flows for my applications especially for critical features. So I would either tag the mermaid diagram generator skill at the very end after a plan has been implemented to understand how things work or if I'm just kind of worried about something then I will tag it and then ask it to look into something for me. Can you describe how the organization creation flow works in this application and how it integrates to stripe as well and now the skill seems to be done and made me a mermaid diagram. So, I have the auto organization creation and then the manual one and then a bunch of other similar logic as well like the member invitation flow. And this basically gives me and anyone that I'm working on a project with a good understanding of what Cloud Code actually did behind the scenes. And honestly, I'm really surprised by how detailed that is because it says, for example, all subsequent TRPC calls include the organization ID. Now, you can either get Cloud Code to make a skill like this yourself or download it from the relevant video in the Cloud Code masterass. But the main benefit here is that I just get a better mental model of my codebase so I know how things work behind the scenes, which makes me feel more secure when shipping to customers. And this is really helpful when it comes to diagnosing a problem that a user may be facing. So if I get an email complaining about their latency or something, then I can get Clauddio code to make that diagram for me and then see all the steps involved and then maybe condense it or simplify it down a bit.

Separating Plan Sessions

Now I pretty much always plan in one session and then execute in another one. So for example, Claude Code came up with this plan over here. I almost always do yes context and auto accept edits and then it will load that plan into brand new like session make all the changes are required and then that means I still have a good amount of context left to make any minor adjustments after checking out the implementation. Now I

Saving Plans for Later

save every plan that Claude code came up with into the project folder. So in this project I have basically all the past plans from the last 2 weeks. And the reason that I do this is because nowadays I'm moving much faster with cloud code and sometimes I add a feature add five more features after that and then I decided the first feature I added I no longer want and I can't just revert the get commit history because otherwise the five other features that I did want would be lost. But by having all the plans in a folder, I can tell cloud code, okay, find the plan where we implemented that particular thing also use a get commit history as reference and then undo that particular feature change add. And with that additional context, it just does a better job at removing any feature that I added on a whim. And I like archiving the plans because I can get Claudico to analyze it in the future and then give me a history of how this project has evolved over time and it also compare different models in the future to see how the planning process or planning abilities have improved. If you want some better plan management organization for past plans, then there is a pretty good skill that you can download from this project that will be linked down below and it's this plan manager skill and it has a whole bunch of features that you can pause and read here. Now, when Claude

Re-entering Plan Mode

Code has added a feature and then they have a pretty good follow-up feature, I just stay in the main session, switch back to plan mode to make a brand new plan. And this is better because it has a bunch of context earlier on to help him make improvements. And then after it comes up with a brand new plan, I would clear the context and have it be implemented in a brand new context window. So, essentially, I just re-enter plan mode whenever I want to make some more changes to what Cloud Code has done. Likewise, I'm pretty aggressive in

Avoiding Context Switching

starting a brand new chat when I'm working on something new to avoid context switching the model. So, I would get it to do like a profile API fix in a different chat to a landing page redesign and so forth. I know many people do it in the same chat for some reason and then they just see worse results. I instead batch similar tasks together and when it seems the next task is going to be a big task, I then re-enter plan mode. And the way I think about this is basically, is the context that we've accumulated so far related to next task? If so, I would reenter plan mode and do that. If not, then I would just start a brand new chat. Now

Compacting vs a New chat

previously, when the models were not that good, I basically never compacted the conversation. But nowadays, I've been compacting a little more often because they have gotten better. And my methodology is basically, is a model still doing the same task? progressing clearly, as in editing the right files? And have I compacted under free time so far? If so, then compaction is safe. But if the model is making consane or confused edits and I keep having to correct it to the right part of the codebase or if I already compact it too much, then I would just start a brand new chat. Next up, when it

Documenting Failed Attempts

comes to some pretty tricky bug fixing, then I would get Claude Code to make a document of all the attempts that it made so far to fix that particular bug and then start a brand new chat instead of compacting, load in the document of all the fixes that we've attempted so far. And then eventually on the fourth or fifth attempt, it would find the right problem and fix that instead. And then because I prefer to have the smallest possible fix in the codebase, if I've logged everything that was done in previous attempts, then I can remove the unnecessary changes that happened in previous attempts that didn't actually fix the root cause of the problem. It's basically like Okam's eraser where I would want to have the simplest possible

Avoiding Long Contexts

fix. Likewise, if you have a really complicated problem and cloud code just can't figure it out, then instead of trying again over and over in the same session, just start a brand new session. Because it has been observed in many agents that as the context grows significantly, then the agent shows a tendency to favor repeating actions instead of symphasizing novel plans. And that is exactly why starting a new chat may just solve the problem for you. Now

Turning off Auto Compact

I also do have autocompaction off because I like to manually compact instead. And you can do that by going to /config, searching for autoco compact, and turning that to false. And that's basically because I just like having more control over the way that the context window is behaving. Now, the

Context Awareness

cloud 4. 5 models and onwards are aware of their own context window. And there's a good blog post by the team at Cognition AI who made Devon, and they found that this causes context anxiety where the model begins taking shortcuts or leaving tasks incomplete when it believes it's near the end of its context window. They found the best way of addressing this issue is to use the 1 million token context window model, but don't go beyond 200,000 tokens. But if you don't have access to 1 million token context window, then you can put this in your Claude MD file instead. And

Ticking off Plans

essentially, when combining this with one of the long plans that I have, I would then say, okay, can you now tick off everything that you've done from the plan so far and then compact the chat and then have it continue on by tagging the plan file again. Now, I also

Auditing Context Window

regularly audit my context window by doing /context. And nowadays, you don't actually have to worry too much about MCP servers because they are loaded on demand when needs be, but any memory files and also skills here can be pretty important. Some people can end up having a lot of memory files that probably need deleting. Some people end up having their skills taking up 10 20% of the context window because they have like a 100 skills installed and that can be bad because you have way too much for the model to think about at every turn about what it should consider using. Which is

Disable Model Invocation

why for any skills that you don't ever need the model to use, you should be adding this to your formattera, which is disable model invocation. So this can prevent cla code from ever loading the skill and can be handy for situations where you would only ever trigger that skill manually. So it would look something kind of like this where I have the skill formatter added this new line and set that to true. And that means that if I do /context, then I won't see the Windows release skill here, but I can still tag the Windows release skill by doing / Windows release. So this kind of draws this distinction between the skills that you would ever invoke or the skills you want the model to invoke as well. And also reducing the amount of skill description that you have loaded in at all times in your context window.

Models for Skills

Likewise, I also like to define a model for a skill by doing model and then like haiku for example because some of the skills that I have don't really require like much reasoning or thinking. It's just a set of steps that the model should be following instead. If you don't set this, then this by default will be your main sessions model. But since my main session is always an opus model, I prefer to have this set for all my skills. So they're like sonet models or highq models. I also like to have the

Statusline

status line at the very bottom that you can make by doing / status line and then saying show the model name a battery showing how much of the context window has been filled up so far the branch that we're currently on and whether we have any unstage changes. So this asterisk is for unstaged changes. If you do have a burst mode on then you can see the number of tokens here as well. But the model here can be pretty handy especially if you switch models quite a

Session Management

lot. I also do slreame and sl resume quite a lot in my ray OS folder in particular because I want to come back to certain sessions where claude code was doing online research for me or drafting emails or something like that. But I very rarely use these commands outside of my ray OS life OS folder and you can just do / rename and then pressing enter cla code will think of a default name based on what the conversation was about so far but if you want to specify another name then you can do / rename and then like mcp workers or something. The handy thing here is it also renames the tab if you're using a different terminal like warp and it also renames the input bar here which can be handy if you have many sessions side by side. When I am torn

Multi Proposals

between a few different ideas for how something should be implemented, then I get clawed code to spin up a set of sub agents to then explore that possibility in more detail by exploring the codebase or doing research online. For example, have all the information then passed back into main session and then I let the main session decide what should be implemented and present a plan to me rather than trying to have all the approaches happening within the same context window. And if I'm curious about

Resuming Subagents

the details of one particular approach, then I would tell Claude Code to resume that sub agent. Have it do a bit more searching online and then present me an updated version because Claude Code supports resumable sub aents. So you can just tell it to resume a sub agent, do a bit more and then give the result back to you. Now I also do use custom slash

Custom Slash Commands

commands and skills for any repeated prompts pretty heavily in cloud code. So for example, I have a slash command called branch commit PR which basically does exactly that. it creates a new branch, does a commit, and then makes a PR into main. But as of a couple months ago, Claude Code began to treat slash commands as skills, which means that if you do like /context, then you will see that as a skill instead. But of course, with skills, you do get a bit more flexibility. So if I move this into a brand new skill instead, kind of like this. And then opening it up, I can add front matter to the very top, kind of like this, including the disable model invocation, which means that claude code will never trigger this itself. only I will be able to trigger this. So these ones I basically repeat across many different projects and my user level and then I have some project level ones mainly when it comes to doing releases, syncing localizations which I do in a forks context window because I don't want to pollute the main context window. Then a couple other skills such as syncing different models. Also for writing blog posts, looking at logs and a bunch of more things. Something I do

Logs for Debugging

when I find that claude code is stuck in a loop rather than trying to have it guess where the error is happening. I basically tell it to add logs throughout that entire part of the application as much as it can at each point and then run through it, click the buttons that are required and then either copy and paste the logs over or if the logs exist on my computer then I get a sub agent to then go ahead find those logs load them in find the irrelevant error pass it back to cloud code it fixes a problem and then I remove all the logs and that is exactly why I have a Windows log explorer sub aent and a Mac OS one. I

Using Sonnet

also find that when the sub agent is not able to correctly explore the logs to find the relevant error, then upgrading it from the haiku model to the sonet one instead seems to solve that issue in many cases. I also use well-known

Using Well-Known Packages

packages whenever possible because there's more training data for them, which means that agent will get it right first time when using that particular package such as TRPC or next. js. But in many cases when there are packages or libraries that I really want to use that are very recent but don't have much

Exa MCP

training data for I always have the XMCP server enabled for all my projects which allows cloud code search online and find any relevant documentation. And for

Official MCPs

libraries that I use pretty heavily such as like trigger. dev which I've been using quite a lot. I have the official MCP ser installed. So when I do like /mcp I have trigger. dev here. And then I can see that I have the search docs tool which is really handy. And these other tools help Cloud Code do a good job at just debugging any errors when it comes to using trigger. dev because it's still pretty new and like not readily available in the training data. I also

Mixing Models & Modes

do mix models and modes pretty heavily as you may have saw earlier. So the vast majority of the time I'm using Opus and then I only switch down to Sonet when I have like any quick questions about how something was implemented. And very rarely do I use Haiku unless the thing that I'm doing is really well defined like a workflow inside of a skill. And instead of always switching between them, I have the model that I want defined at the very top. So for example, the Mac OS log explorer uses this model. The localization syncer uses the HiQ model because it does not need anything more complicated. The SEO blog writer uses the sonet model. And then I use something like the HiQ model for the Windows release and the Mac OS release. My heristic here is basically if the skill has a really well- definfined workflow that doesn't require much thinking, then a haiku model is perfectly fine because it just runs through the steps. But if it does require a bit more thinking because it's doing something slightly more complicated, then I prefer to use a more powerful model. I personally use

Headless Mode

headless mode a lot in cloud code. And the most powerful thing is that you can define a JSON schema. So essentially, when you have a $200 month plan of cla code, you're getting about $2,000 of API usage out of it. And that means that you can process a lot of data in the background through headless instances of cloud code. But essentially, I may run cloud code hundreds of times in the background to do like receipt validation or do some kind of data research for me where every hour it would trigger a research skill to collect any information and then save into a folder. And you can do that by defining chrons on your local computer if it is running all the time. And I also do use headless mode with cron and skill to basically make sure that my claude code class is up to date. So it would look at the change log and the transcript and then it would identify any videos that need archiving, any ones that need updating and any new videos that I need to record. And you can kind of do this by making like other workflows. But the benefit of using headless mode with cloud code is that firstly you get to use all these skills that you're normally using in cloud code and secondly it's much cheaper because it's going through your anthropic subscription instead. I also do use /copy pretty heavily to copy over any emails that cloud code may have drafted or results that I got from doing some experimentation. So, for example, recently there was some kind of like DOS attack that happened on a client server and then I got cloud code to basically look through all the logs, figure out what happened and then if I want to share this on Slack, I can basically just do /copy and it will copy this all to my clipboard which preserves the formatting as well in markdown which means I don't have to like do this like weird selecting and copying it because that may not be preserved correctly. So if I do /copy often if it's a return to function then I can choose a small section kind of like this or choose a full response and then I would go to slack or something just paste it in here to send over to the client. So you can see it says this looks like markdown and I can press apply formatting and then it formats it a bit better. And likewise if I want to export the entire session I can do /export then save to a file and just send that file directly to the client instead so they can kind of like know how I diagnose issue and stuff like

/insights

that. Next up, there is slash insights. And I actually have not found this to be that useful myself because I think I'm already really good at using cloud code, but you may find it to be helpful because Claude will look through your entire sessions. I think it looks through the last dozens of sessions and then gives you insights in a file of what you could be doing better. So often

Task Lists

Claude Code will do this automatically. But when I have a really long plan, then I tell it to turn that into a task list. So, for example, I have a pretty long plan that I made with the spec developer and I'd say turn this into a task list. And after doing that, I would either say complete each task itself inside of a sub agent or if the task list is like pretty long and big, then I would say spawn up an agent team to complete that task. So, yeah, you can see it's now made the task list and it's waiting for me to tell it what to do next. Okay.

Worktrees

Now, when it comes to using work trees, I use the built-in work trees inside of cloud code. And I would do claw-w feature one and then claw-w feature 2. Now, I only use git work trees when I'm having two features being built in parallel that will kind of overlap in some way in the codebase because the main benefit here is that when cloud code comes up with a plan, so let's say made a plan on the left hand side and on the right hand side for implementing each feature, it knows exactly what files to change. But if you were not using gate work trees, if the code base suddenly changed as it was implementing something because it expected a file to be in one location but it actually got moved to a different location or one session of cloud code refactor the file but the other session was expecting it to be in a specific format. The way that I think about it is that having work trees when two features are being implemented holds the codebase static for each feature and then once implementation has happened I would do branch commit PR and then it would open up a new PR. But I don't always use work trees. So for example, if I was making changes to backend API and I was doing some small changes to the homepage, then I know that these changes will not be overlapping with each other in any way because I know those two parts of a codebase are distinct from one another. So if I'm just changing some copy on the homepage in one session and making backend changes in another session, I don't mind if they happen on the main branch without work trees. And finally

Chrome MCP

I use a clinging chrome MCP extension quite a lot. Now, this is mostly for pretty simple background repetitive tasks, but also when it comes to debugging any network errors, for example. A pretty simple example is that in my rayos folder, I have my own to-do list that exists on to-doist and then I have a to-do list for Claude Code. So, I told Claude Code which to-dos I should assign to me and which ones I should do itself. And when I upload a video or make a post online or something, I then tell Claude Code, okay, add to your to-do list three days later, check how that posted in terms of views and likes and so forth and any learnings that you have from the comments. Then it would automatically open up Chrome, go to that relevant post or video, and then read through everything. It's also pretty handy for scraping data from websites that are pretty anti-scraping data, and also just filling in forms online, especially really long ones. So, what I would do is I would see a pretty long form that I got to fill in. I would then just dictate everything to cloud code and then get it to fill in the form for me. And I've actually been using it to do my taxes as well. So, it just fills out all the forms for me. I then give it any like materials that it needs. It does the calculations for me and then fills them out. So, anyways, this is a small fraction of all the ways in which I use Claude Code. And if you want to learn more, then I cover many of these topics in more detail in my Claude Code Masterass and a lot more things that have not been covered in this video as well. I do update it regularly with all the cloud code updates and anytime I have a new workflow or technique then I also share that in the class in the relevant my daily workflows class. There's also a lot more things that I haven't covered before on the channel like advanced context engineering if you are also interested in my cloud code newsletter which is completely free and gives you access to a bunch of free videos from the class. There will also be a link for that down below. And I do send it kind of regularly whenever I discover something new or like have a new tip to share or something. And many people do reply to newsletters sharing their thoughts which in turn informs my thoughts going forwards. So links to that and many other things that I mentioned will be down below if you're interested. And if you have enjoyed this video then do subscribe to the channel because it lets the algorithm know that this is a good one and it pushes it to more

Ещё от Ray Amjad

How to Make Anki Flashcards 10x Faster with AI (for free!)

Ray Amjad | 23.04.2025 | 6 сегм. | 267 982

Staying Organised for A-levels // Cambridge Student

Ray Amjad | 05.09.2021 | 11 сегм. | 117 382

How I Got an A* in Physics A-level (Cambridge Student)

Ray Amjad | 28.09.2021 | 14 сегм. | 68 131

Mistakes I Made in Sixth Form

Ray Amjad | 17.09.2021 | 10 сегм. | 68 002

Brutally Honest Replit Review (Not Sponsored)

Ray Amjad | 12.09.2025 | 16 сегм. | 49 405

How I Got an A* in Further Maths A-level (Cambridge Student)

Ray Amjad | 02.10.2021 | 14 сегм. | 47 938

Ctrl+V

Экстракт Знаний в Telegram

Транскрипты, идеи, методички — всё самое полезное из лучших YouTube-каналов.

Подписаться