Transcription
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.
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.