Claude Code's Biggest Update in Months

Alright, so after about two weeks, we're  back with more Claude Code updates and I'm   gonna go through most of the features  that were released about an hour ago,   of which there are quite a few. Anyways, this  video is sponsored by myself and my Claude Code   Masterclass. There is a new year sale going on  right now and there will be a link down below,   but it's basically the most comprehensive class  on Claude Code that you will find on the internet. Many people have taken the class and gone on to  be the best Claude Code users at their company,   so you may also find it helpful. Anyways, the  first feature that caught my eye is running skills

and slash commands in forked subagent contexts.   Using this thing in the skill frontmatter. So basically the way that this works is that  if you go to any of your Claude Code skills,   so for example, if I go to my GitHub  release notes skill, then at the very top,   I can add this to the skill frontmatter. So  it looks like this context fork. And when I   then run that skill, it will then  run that skill within a subagent. So let's make sure Claude Code is  updated to correct version `claude   update`. Can you run the GitHub  release notes generator skill from   2. 16. 1 till now? And then Claude Code  will ask me if it can use the skills. So press yes. And now you can see the skill  is running within a separate context window   because the amount of tokens that it's using  is not going up in the main session. And now   that it is done, you can see the token count  only increase slightly compared to before,   where if it was running in the main session,  then it would basically like use a lot of   tokens because this skill is looking through  a lot of files within the commit history. So anyways, for any of the skills of yours  that can end up eating a lot of context in   the main session, if you add this to very top  context fork, then it will run within a separate   context window where like a subagent does. Now,  one advantage is that if you make this change,

previously if you made a change to skill or made  a new skill, you'd have to restart Claude Code,   but now you don't have to because it's like  automatically monitoring any changes that   are happening. Now, this should also work with  specifying the model in the skill frontmatter,

where if you don't have a model specified,  then it uses the model of the main session. But if you do, then it uses that  particular model instead. And in   that case, it can run faster  if you're using a lighter,   faster model. And of course, you can find  the model aliases by going on this page.

They also added support for an agent field  and skills to specify the agent type for   execution. And after a bit of experimentation  with this feature to figure out how it works,   basically you need to have the context  fork enabled on your skill frontmatter   and then any of the subagents that are  available in Claude Code. So for example,   I have a macOS log analyzer  subagent that is a Haiku subagent. If you then specify this in a skill,  such as this example find to do skill,   what will happen is it will first  spin up that particular subagent.    With all the information you provided to  subagent as the system prompt. And then   after it launches as subagent, it will  trigger the skill within the subagent.

And this should not be confused with the reverse  behavior where you can get subagents to use   and trigger skills by passing in the skills  that the subagents should have access to into   subagent frontmatter by doing something  like skills and then your skill names,   which is like find to do and like GitHub release  notes. You can get Claude Code to now reliably

respond in a different language by adding a  language parameter to your settings file. So   I can do like Japanese for example, and then  running Claude, tell me a story for example,   it will then give the output in Japanese  for whatever story that it comes up with.

Okay, now the next one is pretty  interesting because subagents can   now have their own hooks from pretool use,  posttool use, and stop hooks. And this also   applies to skills and slash commands. So  essentially it looks kind of like this. You would have your usual like  subagent name, description,   and then you define hooks over here. So  pretool use, posttool use if you want to   and stop. You define any matches that would be  required, and then what exactly the hook does. Okay, so here's a quick example. We  have a Cloudflare deploy subagent,   which has a couple hooks over here. And  this is for my application, HyperWhisper. There will be a coupon code down below,  and it's for the cloud version of the   application. So it's hosted on  Cloudflare Workers. So basically   what happens before it deploys is it will  check whether the types compile properly. After it deploys, it will then do a quick  health check, depending on whether deployed   to dev or prod. Using this health check  URL. And then finally it will send me a   notification using AppleScript on my computer  saying that the deployment has been completed. And here is where I call the Cloudflare deploy  subagent. And before doing the deploy, it does   a check to make sure the types work properly.   And then after calling the deploy command,   it then does a health check right down over here  to make sure it works, and then passes that back   to main session where like deployment is  complete, the health check and so forth. I made a similar one for macOS notarization. So  if I go through this, I can just tag the macOS   notarization subagent, press enter. So you can see  that it actually finished early because I did not   specify which file I want notarizing, but it did  send me a notification on my computer over here. So you can see that previously we had hooks  that were in the `settings. json` file and   that would run for all tasks. But now we  can make agent scoped hooks by defining in   the frontmatter of each agent and then also  do for skills and slash commands as well. So basically you could have like a deploy  subagent that would then like auto validate   stuff and do health check like I just showed.   You can make a research explore subagent that   would basically block any write operations  within a specific directory on your computer,   for example. And you can basically  have like test runner subagents. If you have a lot of tests in your code  base, you could split up 100 tests across   10 different test runner subagents and then  capture the results of any tests that failed   by using one of these hooks and then pass it  back to main session to then investigate why   that particular test has failed. Now, those  are just a few of the ideas I have had so far,   but I'm sure many people will be coming up  with more ideas of how this can be used. Claude

Code now also has a specialized subagent  when it comes to running batch commands. So if you do @, then you can type in  bash and you can see the specialized   subagent right over here. So I can say like  `npm run build` in the Next. js application. Then it will spawn up that particular  subagent, run through all the bash   commands that are required, and then it will  give the final result back to main session,   which keeps the main session's context  window clean. And then I can move this to the   background by doing `control B`. And as I showed  earlier, I can combine it with other things. So for example, my macOS appcast generator  skill, I can just add like agent type in bash   and then it will use the specific bash sub-agent  that's specialized to running bash commands. So I   would probably use this for bash commands that are  even multi-step or can take up a lot of context.    So for example, I could make myself a new  skill that sets up Docker and installs relevant   dependencies and does a bunch of other things  and that can be assigned to a bash sub-agent.

There's now wildcard pattern matching at any  position. So if you go to your settings file,   for example, then I already have it right over  here. So it can do any `npm install` command. If I wanted to allow any install command,  then I can just put a wildcard at the front   and then it would allow that. I can also do  the same thing with git commands, for example,   as well, by doing like a `git` and then  asterisk. When you press `control B`,

then I will now move all bash commands  and subagents to background at once. So if I run five explore subagents simultaneously,  then you can see they're all running in the   foreground right now and I can press `control B`  and then all five of them have moved to background   right now. And this basically means that you  can continue chatting with Claude Code whilst   things are still running in the background.   They added a slash `remote-env` command.

So if you do `remote-env`, press enter, then  you can see these are the remote environments   that I have. And basically they are the exact  same environments that I see when I go to the   Claude Code desktop or Claude Code in the  cloud. And I can make a new one over here,   put in the environment variables and so  forth, and then it should appear in this list. So if I change this to default, for example,   and then move a command to background by doing  `&` then I'm typing in the prompt, for example,   and that will then go to Claude to that remote  version. Instead of deleting unused subagents,

you can disable them by adding this to your  deny list in your `settings. json` file. So for example, I can just call this like  `macos-notarize` and this subagent will   no longer be available to Claude Code.   You can also now enable planning mode

by doing `/plan` and that will turn it  on. And if you do `/plan` again in plan   mode and you created a plan earlier,  then it will show you the same plan.

They now added support for once for hooks, which  basically means a hook will only run once the   first time it is matched and won't run again on  successive matches. And you can just enable it   by doing like for example, `once: true` anywhere  you're using hooks. They have now added a real

time thinking block display in transcript mode,  which basically means that if you press `control   O`, then you can see like right over here, this  is what the agent is thinking about right now. And then it changed to slash context, they  added skills, which basically means that

you can see roughly how much tokens your  skills are now taking up. So previously

when a sub agent would ask you whether it  can use a tool or permission, for example,   and you press no, then the sub agent would just  stop there completely. Now if you press no,   then the subagent will try different  approaches and continue working.

You can now trigger skills as slash commands  by doing like just a skill name. So I can do   `github-release-notes`, press enter, and  then it will start running that particular   skill. And whilst the skill is running, I can  see what kind of tools that it is calling. For example, these logging tools  and diff tools. And if I do slash,   I can see my skills and these are ordered by how  frequently they are used. When it comes to things

that have tabs like slash stats, I can use my  arrow keys to go back and forth between the two. Rather than pressing tab to cycle. They now  fixed the escape key with queued prompts.    So basically what would happen is  that if you queued some messages,   if I press escape, then it will  remove that last message from the   queued prompts and it won't actually cancel  task unless the prompt has gone into task. Whereas previously it would like unqueue  that message and then also cancel a task,   which is like not what you wanted. When it  comes to using terminals like Ghostty or iTerm,

you can move onto new lines quite  easily by just pressing shift enter.    Without having to change like any  configuration as you did before.

And also the creator of Claude Code said  because thinking is on by default everywhere,   UltraThink doesn't really do  anything anymore. So for example,   like Claude Code desktop that has thinking  on by default. And then he'd say is that   UltraThink can only be useful if you  have thinking turned off by default. Because then if you put it in that prompt, it  will turn on thinking just for like solving   that particular problem. Anyways,  I'll now be updating my Claude Code

Masterclass with a lot of information  that I briefly mentioned, but in more   detail. So if you want to do get better at  using Claude Code, then do check that out. There will be a link down below with  a coupon code. And something else I   have been thinking about recently is the  fact that Anthropic will most likely end   up introducing something this year that will  account or like make multi-Claude Code agent   orchestration easier. So many people have  been making their own tools around this,   like Gas Town, for example, and that allows you  to run many Claude Code instances in tandem. There are other tools like, I think, AutoClaude  and like Claude Kanban Board and stuff like that.    Basically a lot of people are making like random  tools around this like domain and no one exactly   knows what correct or the final form of like this  multi agent orchestration thing you will take. So   I imagine that Anthropic will be introducing  something to do with that later this year. And basically what makes me think this is  that someone shared a like system prompt   from one of the preview versions of Claude  Code. I think this has now been removed,   but you can basically see that like  you can spawn teammates and messages,   team coordination, like add a task update  to the team. So they have this like kind   of idea of teammates that may be baked  into Claude Code in the future. And this has kind of happened before  with Claude Skills, for example,   or MCP. Some serious users were kind of  playing around with things and making   their own GitHub libraries like Gas Town.   And then Anthropic basically comes up with   their own official way of doing something  that's adopted by the rest of the industry. Anthropic just seemed pretty good at making  these format standards protocols that seem   pretty well adopted by everyone else. So I  imagine they will be coming up with something   this year about like multi agent orchestration  in a way. And whatever they do come up with,   I imagine will end up being the  default for many like coding CLI   tools and it will probably be available  in Claude Code before any of a CLI tool. Anyways, I do kind of share  my thoughts on this kind of   stuff in my AI newsletter that will be  linked down below if you are interested.