Claude Code is Now the EVERYTHING Agent

I'll be going over the last week of Claude Code  updates and how you can be using them in your   workflows. First of all, we have Output Styles,  and that basically allows you to adjust Claude

Code beyond just software engineering. They allow  you to use it as any type of agent, whilst keeping   its core capabilities, such as running local  scripts, reading and writing files, and tracking   to-dos. And they already have the default Output  Style, which is baked into Claude Code, it is in   the existing system prompt, and that is designed  just for software engineering. And they added two   more additional Output Styles that you can switch  between, one being Explanatory Style, and that   provides you educational insights, whilst it's  helping you complete these software engineering   tasks. So it can help you understand why Claude  Code may be choosing one database over another   database when it comes to planning a project, or  why in some cases it may use a store instead of   using a React context, for example. And Learning  Style is kind of like pair programming, so it   helps you learn by doing, where Claude Code will  not only share insights whilst coding, but also   ask you to contribute small strategic pieces of  code yourself. It will add a to-do human in your   to-do list, and then basically get you to complete  that specific task, and then check your work,   and provide any insights depending on how you  completed that. So we can try both of them now.    Firstly, you want to make sure you're on the right  Claude version, so you write claude --version,   and you should be on 1. 0. 81. And if you're not,  then just write claude update, and then it will   update to the latest version. And then start  writing Output Style, and you should see it   over here. And then you can choose Explanatory or  Learning, so I'll pick Explanatory and get it to   do a task. And now you can see it added an insight  over here, but because I'm already very familiar   with the code base, it's not anything new to me.   And something else that I noticed that they added,

but they didn't actually mention in the changelog  over here, is this ability to say yes and always   allow modifying files in this project. So that  basically means it will never ask me again if   it should make a particular edit. It will always  make edits in this project. So I can press Enter   over here, and after pressing Enter, it adds a  brand new line in the settings. local. json file,   which is default mode accept edits. So if I  wanted to go back, I can just remove this line,   and then save it as it is. And now we can create  our own output style by writing output-style: new,

and then say something like, write everything  you're doing in the style of Shakespeare,   press Enter, and then it fires off a subagent that  creates an output style file for this particular   style that I requested. And one thing you want to  bear in mind is that output styles directly affect   Claude Code's system prompt, so non-default output  styles exclude instructions specific to code   generation, and efficient output normally built  into Claude Code, such as responding concisely   and verifying code with tests. So this means  that this output style that I generated will   actually perform worse when it comes to coding,  because it doesn't have everything else required   in the system prompts that Claude Code usually  has. And we can see that Claude Code made a new   output style in this file over here. So if I open  that file, then you can see it kind of looks like   this. And then I can select the output style from  the list over here, and then say something like,   describe the code base. And you can see it's  responding in Shakespeare-style English. Now,   a few things worth bearing in mind is that output  styles completely turn off the parts of Claude   Code's default system prompt specific to software  engineering. So when you're actually using the   CLAUDE. md file, what's happening is you have the  system prompt that is baked into Claude Code,   and then just after the system prompt, you have  a user message, which is a CLAUDE. md file. So you   can see it says it over here. And if you use  a --append-system-prompt, that adds whatever   you said at the end of the system prompt. So for  example, I can use a --append-system-prompt. So   for example, if I do claude -p, tell me a funny  story, --append-system-prompt, respond only in   Japanese, then that will be added right at the end  of the system prompt, baked into Claude Code. And   you can see that the story is right over here. And  because it's a system prompt, it's weighted more   than a user message. Whereas because the output  styles actually remove one part of Claude Code's   system prompt, it means it's worse at software  engineering, but can be better at other tasks.    So you can use it for things such as writing and  then have it do web searches for you, or help you   do research and connect to MCP servers that help  you with like finding papers or something. It   can be better for other non-software engineering  related tasks. And basically because of that fact,   it adapts Claude Code for uses beyond software  engineering, because you can make it really good   at research, for example, by actually editing that  part of the system prompt, or make it really good   at content creation or something else. And if  you actually want to see what the default system

prompt is, then you can run this command over here  to use Claude Trace, press enter, and then it will   run a version of Claude Code. Basically, that  intercepts all the requests going to Anthropic's   API, and it shows you the actual requests. So  I can open up this file over here. And then if   I say hello, and then if I go to raw calls over  here, open up the last message that I just sent,   and then type in system over here, and then I  can see it says you're Claude Code, Anthropic's   official CLI for Claude. And then I can copy the  system prompt over here and go to a new doc. And   then I can actually see the system prompt baked  into Claude Code all over here. So this may be   useful for you in helping you create a brand new  output style. And now we also have a brand new

planning mode. So if you write /model over here,  then you can see it says Opus plan mode, which   will basically use Opus 4. 1 whenever planning  something, and it will use Sonnet 4 otherwise. You   can see the other modes are Sonnet 4 everything,  Opus for everything, or using Opus up to a certain   percentage depending on your plan, and then  using Sonnet after that. Something else that we

also have is the ability to load in multiple MCP  config files. So for example, in this project,   I have two MCP config files, one for context 7,  which is fetching docs from the internet, and   one for Supabase. It uses my API key, API token,  which will be expired by the time you're watching   this video, so don't use this. And basically, if  I do claude --mcp-config, then I can do mcp1. json   mcp2. json, and load in both MCP files. And then if  I change my mind in the future, and I don't want   to load in the second MCP file, I can just remove  that. And then I can say who is the last user who   signed up to the production version of my app,  press enter, and then it will use a Supabase MCP   to find out. And this is the last user who signed  up, and you can see it's still responding in the   Shakespeare style, which is kind of funny. And  if I do mcp over here, then I can list the two   MCP servers that I loaded from the two individual  files. Now this can be useful because sometimes   when you have a lot of MCP servers, and you have  a lot of different tools, you may only need one   MCP. json for one particular use case, for example.   But it is worth bearing in mind that even though   you have a lot of MCP servers with a lot of MCP  tools, Claude Code should be able to handle them   pretty well. Because there's a benchmark over here  called LiveMCPBench, and basically they compare a   lot of leading models to see which performs best,  when given 70 MCP servers spanning 527 tools. So   they asked each of these models to complete  a bunch of real-world tasks, such as booking   train tickets and stuff. And you can see that  Claude Sonnet-4 and Claude Opus-4 is basically   in a league of its own when it comes to accuracy,  when given a lot of MCP tools. So this setting may   not matter that much for performance, it's more  for saving on context instead. And something

else they added is a command to help manage your  permissions. So if I do /permissions over here,   then you can see these are the permissions  I always allow, which is npm run install,   npm run lint. I can have it always ask for certain  permissions. So I can do like press 1 over here,   do supabase. So whenever it uses a supabase  command, it will ask for permission. Always deny   certain permissions. And then also add in other  directories I can read. And if you go to your   coding editor, go to settings. local. json, then  you can see that everything is over here as well,   which you can make edits to. Now, something else  Claude also did is introducing a 1 million token

context window. And right now it's not built into  Claude Code subscription plan. So you have to   switch to API pricing for that, which means that  you need to log out by doing logout over here, if   you want to use it and then do claude again. And  then when it asks you what you want to be using,   you should choose Anthropic console account,  API usage billing, connect it to your Anthropic   console account over here. And then you want to  run /model sonnet, and then in brackets over here,   square brackets, 1m. And that will switch to  Sonnet, like 1 million token context window   model. And then you can say hello again. And then  it will start using the model with that context   window. And if you want to learn more about this,  then you can watch my previous video about it over   here that came out yesterday. And as a bonus tip,  if you want to increase the amount of thinking

that Claude can do, then you can go to your . claude/settings. local. json, and then add this as   an env over here. So it's MAX_THINKING_TOKENS, and  then do 32000 tokens. And then if you rerun Claude   Code after exiting, and then you can see I have it  in an override list, which is Max thinking tokens,

32000. And yeah, that was most of the changes  that they added last week. If you do want to   see more videos like this, then do subscribe  because I will be making another one next week,   and the week after, and so forth, until Claude  Code basically gets replaced by something better.