Anthropic Just Added These Features to Claude Code

Okay, so earlier today, some interesting  new features were added to Claude Code,   so we will be going over them, as well as many  of the new features from the last two weeks. But   before getting started, it's worth me mentioning  that I have a Claude Code masterclass that teaches   you every single feature in Claude Code, as  well as many techniques that I've picked up   myself from over 1000 hours of usage, many of  which that you will not find anywhere else on   YouTube. It will be linked down below alongside  a coupon code, and it does have a seven day money   back guarantee if you're not satisfied. Okay,  first of all, they made auto-compacting instant,

and I noticed this earlier today because I  compacted this conversation and it basically   happened immediately. And usually when it came to  compacting conversations, it would take about two   minutes, but now it's much faster. And basically  the way that this works behind the scenes is that   your. claude folder in your home directory has a  projects folder. And inside of that, if you go to   any of your projects, you can then click into  any of the sessions that you've had or any of   the chats, go to session memory, and then you can  see a summary of that session. And basically every   time you send a message to Claude Code or every  time it responds to you, it always adds onto the   session memory, the summary. So you can see it  kind of looks like this. You have the session   title, and this is a title that Claude Code came  up with itself, the current status of the session,   which includes anything completed, any discussion  points, and open questions. And then like a bunch   of other things that you can see described over  here, some learnings, some key results, and then   a work log, which is automatically updated every  time you send a message. So if I quickly send a   message to chat and say something like, can  you tell me a short story? And you can see. Now it says, user requested short story, Claude  wrote a short story about the conversation itself.    So basically it's now interesting that every  time you're using Claude Code, it's automatically   keeping a memory of the entire session. And I  believe what's happening is that when you do   compact and it's instant, what it's actually doing  is it just loads the summary of that session into   a brand new session that it was writing as you  were going along in the previous session. And   if you are interested in what's stored in the  session memory, then you can basically just ask   Claude Code, where is your session memory stored?   And it will tell you the exact location because it   knows its own like session ID. And you can see in  my case, here it is, and I can just click on that   and then see the current status of the session.   Anyways, I thought that was a pretty interesting   decision and I guess you can have fun looking  at the session memory if you are interested in   what Claude Code is doing behind the scenes.   Okay, so now agents and bash commands can run

asynchronously and send messages to wake up the  main agent. So for example, I can spawn a brand   new subagent by saying, can you spawn a general  purpose subagent to do intensive research online   and explain to people why they should take my  Claude Code masterclass with a link? And then   press enter. And now you can see it spawns up this  agent with this task given to it. And then I can   press control B to move it into background. And  now you can see it's looking for the agent output,   but I can press exit because it's listening  for the agent output. I can press exit,   continue on saying like hi, tell me a story, for  example. And then the chat can continue as normal,   like it just told me a story. I can then do slash  tasks and see that agent is running async in the   background. Press enter, click on it to see how  many tokens it's used so far. Then I can also   see the token count go up. And everything else  that it's doing here. And then I can continue   doing my work as normal. And finally, once that  sub agent is done, it will then wake up the main   agent again. And then finally, once it is done,  it will say agent completed. The research agent   has completed synthesis and Claude Code will use  a new tool called the agent output tool to then   look at the response that the agent then gave  it. And then it will like combine that response   into the main agent's task at hand. So you can  see now the comprehensive research for why you   should take the Claude Code masterclass is now  ready. And then I can use this for marketing   material or something like that. So this can be  pretty useful in situations where you're coding   on something. You can get a subagent to then like  search to find information in the background and   then come back and resume the session once they  found that information. Likewise, the same thing   also applies for bash commands. So in certain  situations, you may be running a npm install   command or ffmpeg command or something like that,  which may take a while in the background. You can   just leave that running in the background and then  continue on with any work that you want to do. And   the way of doing it with a bash command should be  pretty similar to what I just showed you earlier.

Stats now provides users with interesting Claude  Code stats such as favorite model, usage graph and   usage streak. So if I do stats, so these are my  stats and I think they only started recording this   recently because I don't see any stats from like  this time. So yeah, I'm not sure what's up with   that, but basically it seems my favorite model is  Sonnet 4. 5 and that's probably because I use a lot   of Sonnet 4. 5 based sub agents. I used 10 million  tokens so far. My longest session was one day and   eight hours and I've been basically using it every  single day so I can like copy these stats by doing   and then if I go to Twitter and then press Command  + V, then you can see I can just share the stats   online pretty easily. If I press Tab to then go  to Models Tab, I can then see how many tokens I've   used per day for all the different types of models  available. And I can do Control + S again to copy   this and then go to Twitter and then do Control +  V to paste it. So yeah, this can be a pretty good   way to compare people within the company maybe  for fun or something like that. They now added

named session support, so you can do slash rename  to name sessions and then resume sessions by a   particular name that you gave it early on. But  I think that may be broken. Because it doesn't   recognize a command slash rename. So maybe they  will fix it later. But what I have noticed is that   if you do slash resume, then you can see all your  previous sessions and you can also rename sessions   here as well. So if you do R, then you can just  rename the session and I can call this like R   type in YouTube, press enter, and now the session  has been renamed to YouTube. And if I want to,   I can do Claude dash resume and then  put in YouTube and that will resume that   previous session. Likewise, the resume screen now  groups forked sessions together and has keyboard   shortcuts like preview and rename. So if I go  to slash resume again, and I can go to previous   session like how does auto compact work for you?   Press P, I can see everything within that session   and I can choose to resume it by pressing enter  or pressing escape to go back. And for example,   this session over here has narrow. So if I press  B, then I can see any previous sessions that   existed in that main session. Because I compacted  that session, a brand new session was spawned that   was forked out of the main session. And then  I can resume any of those previous sessions from   before it actually compacts to the conversation,  which can be pretty handy. Another handy thing is   you can press A to then show all your projects,  because when you do resume, then you can only   see the sessions that happened in that particular  folder that you're currently inside. We have now

also added support for Claude rules. So basically  with Claude's memory, there are four types of   memory locations in a hierarchical structure.   So the enterprise project memory, project rules,   user memory, and project memory local. So this  is quite a lot, but the way it's loaded in is   Enterprise is loaded in first, then the project  CLAUDE. md file that you may already have. Then you   can now define project based rules. And this  basically allows you to modularize your real   files better, kind of like Cursor rules, if you  do remember that. So if you scroll down, you can   basically see this is a structure in your. claude  folder. You'd have your CLAUDE. md file inside your   project. And then you can have rules for different  types of things. If you don't define a path at the   very top of the rule, then it will be applied all  the time. But if you do define a path, such as an   APIs folder, then that rule will only apply to  files within that folder. So you can see how   the setup is. It's like pretty similar to Cursor  rules if you're already familiar with that. Now,   of course, the advantage is you can now organize  your CLAUDE. md file better and have it only apply   to specific folders. So for example, if you have a  big mono repo like me for this particular project,   I can extract some stuff that only apply to the  CloudFlare workers and then some other stuff that   would only apply to macOS application over stuff  that would apply to a website and so forth. I will   be covering this in more detail such as how to  get it set up, use sim links, user level rules   and stuff. In my Claude Code masterclass as well,  pretty soon. For multiple choice questions, Claude

Code will now use a recommended indicator with  the recommended option being moved to top of the   list. So this can be useful in planning mode when  you're not exactly sure which option you should

take. Whenever Claude Code now makes a commit,  it will then say which model actually wrote the

code here. You can now enable and disable MCP  servers quickly by doing /mcp enable and then   the server name. So for example, MCP, I can see  I have the Supabase MCP installed. I can do MCP. Disable. Supabase and that will quickly disable  the Supabase MCP. If for some reason you want to

launch Claude Code without any slash commands,  then you can use this. So you can do claude   --disable slash commands. And then if you  do something like slash MCP or slash resume,   you won't see any of the commands available. So  I guess this gives you a more minimal like just   chatting with Claude experience. Maybe they now  also added a dash agent CLI flag to override

the agent settings for the current session. And  that basically means that if you have some kind of   sub agent in your session, then you can run Claude  Code as that particular sub-agent. So for example,   I have a macOS log analyzer sub-agent that uses  a Haiku model because it doesn't need a more   powerful one. I can copy over the name of that  sub-agent, do claude-agent, put in the name, press   enter, and then you can see we'll automatically  run as the macOS log analyzer sub-agent. And I'll   say, what do you do? And then you can see it says,  I'm Claude Code, blah, blah, macOS unified   log analysis. So I think that can be pretty cool.   You can run Claude Code as any of the sub-agents   that you have in your project so you can have  all the same instructions that the sub agents   do and actually manually execute a particular  sub agent. I can also run Claude Code as any of   the default Claude Code sub agents by doing claude  dash agent and plan for example. And that is   the planning sub agent that is built into Claude  Code. And then if I say, what do you do? Then it   will let me know of the planning responsibilities  and capabilities that it has. So what this means   is by default, if you leave out the agent flag,  then Claude Code will run as the main agent that   is built into Claude Code. But if you specify an  agent flag, then you can use any of the   default subagents, like manually, like control  them, or you can use any of the subagents that   you have actually installed or created in your  project. And this can be pretty good because   usually when Claude Code spawns up subagents, like  you can't stay the subagent in any particular way,   it just runs and then comes back to you. But in  this case, you can probably diagnose subagents to   make sure they operate well by actually running  the subagent manually and trying to like get a   feel for it. They also now added the ability to  delegate tasks to Claude Code from Slack. So in

Slack, you can go to the app section, search for  Claude, and then click on it and press add   to Slack to add it to your workspace. And then it  will ask you for like permission and stuff. So you   can press allow. And now it says Claude Code is  now installed. So we can go back to Slack. And now   here we have Claude available. So I can press  connect account, then connect to my Anthropic   account, go back to Slack again. And then I can  choose a model that's used for conversation,   choose whether it's code only or code plus chat,  see the organization. And then inside a channel   I can do @Claude and be like, hey, can you add  a Chinese translation to a landing page for my   application, HyperWhisper? And then that,  and then HyperWhisper will transcribe what   I just said. And then Claude Code will be like,  Hey, I'm going to get started with this. But it   seems I have to add Claude to channel first. So  I'll and that's organizing   gathering information, and then it will generate  me a response. But it's now asking me the repo   it should work in because I can't identify that.   So this can be pretty useful if you want to work   with Claude Code on a bug fix or something in  your team, or if you have some kind of channel   on your team Slack where you're regularly getting  bug fixes or issues coming in, you can just tag   Claude Code. And if you want to extend the ability  then you can go to Claude, press connect to Slack   connector, and within the connectors over here,  you can add a custom connector to like the Sentry   MCP server or something or your own logging  system. And if you are interested in buying my   speech text application, HyperWhisper, then there  will be a link and coupon code for that down below   as well. I am also working on a Windows version as  well, so that will be out within the next month or   so. But if you press download now, then you can  put in your email and then you'll get notified   when the Windows version is out. Anyways, just  before ending the video, I remember reading the

measuring political bias in Claude like article  that Anthropic wrote, and this graph was pretty   interesting because it shows a bunch of different  models. It is one month old, so some of the models   are not on here. But Opus 4. 1 was the best at  considering opposing or counter arguments to like   political problems. And maybe the fact that Claude  is better at considering opposing perspectives and   counter arguments that can carry over to making  better decisions when it comes to coding. And I   imagine like since Opus 4. 1 was the best here,  it's quite likely that Opus 4. 5 is pretty good   as well. Or score slightly higher on this like  eval or benchmark thing that they have. Anyways,

hopefully that video was pretty interesting and  you learned something useful. If you do like this   kind of content, then do subscribe to the channel  because it lets YouTube know that this is a good   video and you will get recommended more videos  like this in the future. And if you are interested   in my Claude Code masterclass, then that will  be linked down below alongside a coupon code.    And I will be updating the class in the next two  days with all the new features that I just talked   about in the video. There's a seven day money back  guarantee if you don't like the class. And so far,   less than 0. 5% of people have asked for their  money back. So that's how I know the class is   really good because like people just like  enjoy to the class and they don't ask for a   refund. And yeah, because I've never accepted  a sponsor on the channel and never want to to   keep my videos unbiased, classes like  this and my software like HyperWhisper   basically helps keep the channel going. So  if you're interested then do check it out.