25 Hidden Settings That You Need In Your Claude Code Setup

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: 25 Hidden Settings That You Need In Your Claude Code Setup · Channel: AI LABS · 15:57 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Claude Code, Prompt and context engineering, Software engineering with AI

Summary

The video presents Claude Code settings, commands, and keyboard shortcuts intended to tailor its behavior and improve coding workflows. It covers creating and auditing project instructions, importing setups, managing skills and permissions, and reducing context use. It also explains options for planning, output styles, focus mode, parallel sessions, notifications, and model selection. The video ends with shortcuts for managing prompts and a command that offers interactive lessons.

Key points

The video describes a setting that changes the init command so it asks about a user’s workflow before writing project instructions, skills, and hooks.
An import command can bring selected instruction files and MCPs from another coding agent into Claude Code.
Prompt audit and skill doctor are presented as ways to review outdated instructions and identify skills that are rarely used or consume context.
Auto mode setup adapts auto mode rules based on a project and past sessions; fewer permission prompts adds commonly used read-only commands to an allow list.
Settings can disable Claude app connectors and artifacts, and can restore the option to clear context when accepting a plan or specify where plans are saved.
Build eval creates tests for an app, while hill climb tries changes and keeps those that score better in the tests.
Output styles, thinking summaries, and focus mode offer ways to adjust Claude’s responses and what appears in the terminal.
Branch, fork, and subtask support different ways of working on parallel tasks, while sessions can message one another.
Question timeouts and local notifications help users manage sessions that need input or have finished.
The video also covers choosing models for planning and subagents, prompt shortcuts, and the powerup command’s interactive lessons.

Chapters

Start	Chapter	Summary
00:00	Project setup and importing instructions	The video introduces hidden Claude Code settings, then covers an init setting, importing another agent’s setup, and auditing project instructions.
02:37	Auditing skills and configuring permissions	The video explains how to review outdated instructions and unused skills, then describes adapting auto mode and reducing permission prompts.
05:11	Context, evaluations, and plan settings	The video covers disabling connectors and artifacts, testing app behavior, and changing how plans are cleared and saved.
09:45	Output styles and parallel work	The video demonstrates output styles, thinking summaries, focus mode, and ways to branch, fork, or delegate tasks.
12:20	Session communication, notifications, and models	The video describes session messaging, question timeouts, notifications, and model choices for planning and subagents.
14:55	Prompt shortcuts and interactive lessons	The video wraps up with shortcuts for stashing, restoring, and undoing prompt edits, plus the powerup command and closing remarks.

Quotes

Claude Code has a lot of hidden settings and most people don’t know they exist — 00:00

It shows how often each skill gets used and when it was last used. — 03:39

To use focus mode, you first need to run the TUI command — 10:47

Branch copies the chat you’re in and opens that copy as a new session. — 11:18

It gives you short lessons inside Claude Code. — 15:26

Concepts and entities

Claude Code, CLAUDE.md, Auto mode, Skill Doctor, build eval, hill climb, Codex, Cursor, Scribba, Explain

Full transcript (original language)

00:00 Claude Code has a lot of hidden settings and most people don’t know they exist, even the ones who open it every day. So, we went through every setting and picked the ones that changed how we work. One of them is a setting that even the people who build Claude code use themselves and it completely changes a command you probably run a lot. Another one fixes the biggest problem with running more than one session on the same project. And there’s a command that teaches Claude how you work, so it performs better. We also found some keyboard shortcuts for the little things that get in your way when you write a prompt. If this is your first time,

00:31 we’re a software company and this is our channel AIAs. And in this video, we’re going to show you each of these hidden settings and how to use them. [snorts] Starting with the one setting that creator of Claude Code themselves use in their project, which is a fix toinit command. This command writes claude.md file for you. To do that, it reads your project first and then writes down everything it learns from it. But that’s a problem because the claw.md file is only for the things the agent needs to remember all the time. So if the agent can learn something just by reading the code, it doesn’t belong in that file. To

01:03 fix this, there’s a setting called claude code new in it. You need to set this to one in your terminal right before you type the claw command to launch claude code. It has to go before the claude command because this way the session can access this setting. With this setting on, init straight away like it did before. Instead, it asks you a few questions about how you work. First, it asks if you want to improve your current setup or start fresh and if it should make skills and hooks. Then, it also asks how you use Git and how you’re planning to put your app online. From

01:33 your answers, it makes a small plan and waits for you to check it. When you approve the plan, it starts writing everything. When it’s done, your claw.md file is a better one that has your own way of working in it. And it also creates skills and hooks to give you a better starting point. If you’re working in another coding agent like Codex or Cursor, you might already have a setup there with the MCPS and instruction files you’ve set up for that agent. If you want to move all of that setup to claude code, you don’t need to start from zero. Instead, you can move that setup into claude code using the import

02:04 command. To use it, you just need to run import with the name of the agent you used before. Claude code will then show you all the things it can copy, including your instruction files for that agent and the MCPS you’ve connected to it along with anything else you’ve set up there. Then it waits for you to choose which settings you want to import. After that, it makes a skill that contains all the instructions to copy the things you’ve selected and uses that skill to copy everything into clawed code. So you can use the same setup you had in your previous agent in claude code as well. No matter how carefully you set up your claw.md and

02:37 other project files, they still get outdated over time. That’s because the models keep getting better. So many of the instructions you wrote for older models don’t help the new ones anymore. Instead, they just fill Claude’s context with things it doesn’t need, which makes the models answers worse instead of better. Your setup probably has many skills and claude.md files. If you changed all of them one by one, it would take a very long time. So, there’s a command that checks them for you. To use it, you type the clawed API command and then add prompt audit after it. This command reads your whole setup and uses

03:08 sub aents to check each part of it. Then it gives you a report of every problem it found along with the fixes it suggests. When you ask Claude to apply those fixes, your setup only keeps the instructions that still help Claude. You’ll need this the most right after a new model comes out because that’s usually when the most instructions become outdated. Just [snorts] like your instructions files, you need to frequently update your skills, too. Most of the skills that you have set up don’t even get used very often, but they still sit in Claude’s context for no reason. To find those skills, there’s a command

03:39 called skill doctor. It shows how often each skill gets used and when it was last used. It also shows how many tokens each skill takes in the context window. With that list, you can ask Claude to delete the skills you don’t use. Then only the skills you use stay in Claude’s context. There’s also a skills command that lists every skill you have. And when you press T on your keyboard, it sorts that list by how many tokens each skill takes in increasing order. So you can tailor your setup accordingly. But before we move on to the next setting, it would be great if you subscribe to

04:09 the channel and hit the hype button. This small gesture of support goes a long way for us. You might already know that when you start Claude code without adding anything after it, it starts in auto mode. In this mode, Claude allows all commands that are safe on its own. So it doesn’t ask you for permission for every single thing. But auto mode doesn’t know how you work and which commands you usually let pass. There’s a command called auto mode setup that teaches it exactly that. This command reads your project and your past session history. Then it changes auto mode’s rules to match the commands you usually

04:40 allow. When it’s done reading, it shows you the new setup to check. If you’re happy with it, you need to approve it and then it updates auto mode settings. From then on, Claude Code follows these new rules in auto mode. Now, [snorts] that command was for editing auto mode’s own config. But if you don’t want that config to change, there’s another command called fewer permission prompts. This command searches your past sessions for the commands and tools you use the most. It only picks the ones that read things and don’t change anything. Then it adds them to an allow list, which lets Claude know that these commands are

05:11 allowed to run without asking you first. And Claude Code saves this list in the settings file of whichever project you’re working on. So you can let it work on long tasks without worrying that it’s going to change anything that you don’t want it to. When Claude works on a long task without stopping to ask you, it needs as much free context as it can get. But part of that context is already used before you even type anything. That’s because there are connectors already added in the Claude app like notion and Slack. And they also load into every Claude code session, even when you never use them. So if you don’t

05:41 need them, a setting called disable claude AI connectors stops them from loading. You set it to true in the settings.json file in the main.claude folder. This setting applies to all the Claude Code sessions you open. After that, the Claude apps MCP servers don’t appear in Claude Code anymore. You don’t even need to open the file yourself. You can just ask Claude Code to add the setting and it will edit the settings file for you. There is also a setting to turn off artifacts. When you work on a design, Claude first makes an artifact which opens the design as a page in your

06:13 browser. That way it shows you if its plan is right. But the tool that makes artifacts also loads into every session and uses context. So if you don’t need it, you can add enable artifact to the same file and set it to false. Then claude doesn’t make artifacts anymore, no matter what you ask it to do. Now if the app you’re building with claude code uses AI itself, you also need to test whether its AI gives good answers. For that, the same Claude API command from before has a build eval command, too. This command builds a set of tests. These tests check if your app works well

06:44 in different situations. This command writes these tests only once. That way, every new version of your app gets checked against the same tests. Another command that Claude API has is hill climb. This command keeps making changes to your app and runs the tests after each one to see how those changes affect performance. It then scores the responses. It only keeps the changes that score higher than the previous one. But both of these commands cost real API money every time they run because every test calls the model. And if your app doesn’t call claude or any other model,

07:15 build eval still helps. It basically goes through your app and plans tests for you without using the API. So you can check if your app is working as it should. Now when you start building a feature, you usually ask Claude to plan it first in plan mode. Claude writes the whole plan and then it asks you to choose a permission mode and start building. Claude code also used to give you an option to clear the context before building, but it doesn’t show that option anymore. That option helps because all the back and forth from planning stays in the context while Claude builds. If you clear it, Claude

07:45 starts building with only the plan in the context. To get that option back, you change a setting called show clear context on plan accept. It’s in the same settings file in the main.claude folder, and you need to set it to true. You can ask Claude to do that for you, too. Then every time you plan something, Claude gives you the option to clear the context first, the same way it used to. But when you clear the context, the plan doesn’t disappear. The only difference is that you won’t see that file in your project anymore. Instead, it saves every plan in a folder called plans inside the

08:15 main.claude folder. So no matter which folder you’re working in, all your plans go into that one folder. But you usually want them in the same folder that you’re working in. So you can share them with the people working on your project. To do that, you add a setting called plans directory to your settings.json file. Then you set it to the name of the folder inside your project where you want your plans to go. From then on, every plan you make in plan mode gets saved in that folder. But before we move on to the fix for the output Claude gives, let’s have a word by our sponsor, Scribba. They’ve just launched a new

08:45 texttovideo AI tool called explain. When you build something fast, understanding the code later is the hard part, and your standard AI chatbot just hands you more text to read. Explain turns your code, question or document into a narrated video instantly. We connected it to clawed code, pointed it at our little URL shortener, and asked for a walkthrough. Seconds later, we had a real video tutorial that traced the whole flow from shortening a link to the redirect with narration and a moving cursor guiding us through the actual code. It even drew a diagram of how the

09:15 whole thing fits together, and every bit of it came from our own repo. Explain plays more like a personalized YouTube lesson, not a chatbot. and it runs right inside the tools we already use from chat GPT plug-in to a Chrome extension. So instead of writing heavy docs nobody reads, we shipped something anyone could just watch and it’s free while it’s in beta. Try it now. The links in the description below. You can also change how Claude works with you using the output style command. When you run this command, it will list all the styles you can choose from. There are five styles.

09:45 The default style is the one that’s set if you haven’t changed it. Proactive is basically where Claude acts first instead of planning. so it does things straight away. Concise is for when you want shorter answers than Claude usually gives. Explanatory is where Claude explains more of what it’s doing. And the learning style is where Claude asks you to write some of the work yourself so you learn while you build. To switch to any of these, you just need to type the command followed by the name of the style you want. From then on, Claude uses that output style. Claude [snorts] Code used to show you a summary of all

10:16 the reasoning it did when you pressed control and O. But Claude Code doesn’t show these summaries anymore. To get them back, you just need to add a setting called show thinking summaries to the settings.json file in the main.claude folder. Then you need to set it to true. When [snorts] you leave Claude working on a long task and then return, you have to scroll up a lot to find your prompt and Claude’s answer. That’s because Claude code normally shows you every change it makes at each step. But Claude code also has a focus mode which only shows your prompt and Claude’s answer. It also shows a short

10:47 summary of what Claude did. To use focus mode, you first need to run the TUI command, which basically controls how Claude code is laid out in your terminal, and set it to full screen because focus mode only works in full screen mode. Then you need to run the focus command. Now, when you give Claude any task, you can still see what it’s doing and which files it’s reading. But those steps leave the screen when Claude is done. So, when you return after a long task, you only see Claude’s answer, and you don’t have to scroll through a lot to find things. Sometimes in the middle of a chat, you want to try a

11:18 different idea without losing what Claude has already done. For that, Claude code has a command called branch. Branch copies the chat you’re in and opens that copy as a new session. So, you can try the new idea in the copy and still keep your place in the first one. But with branch, you still have to prompt it yourself. When you just want a task done on the side, you use fork instead. Fork is basically branch, but it starts the copy in the background and gives it the task you want done. So the copy works on that task while you keep using your main session as you usually do. And when the task is done, the

11:49 background session ends by itself. But fork always makes a whole new session. If you want the side task done inside your current session, Claude hands it to a sub agent. Normally a sub agent that Claude starts has a fresh context. That means it doesn’t know anything about your chat and it only gets the prompt the main agent gave it when it started it. That’s where the subtask command comes in because it starts a sub agent with your full context. To use it, you just need to run subtask and tell it what you want it to do. Then the sub agent works on that task with your whole chat in its context. We often [snorts]

12:20 work on the same project in more than one claude code session and each session only knows its own chat. Now claude code has a feature that lets these sessions talk to each other. To see which sessions you can reach, you run the list agents command. So let’s say you have two sessions working on different parts of the same project. The first session needs to know something. the second session has already worked out. But you don’t want to copy that context over yourself. So you just ask the first session to message the second one. The first session gets the answer straight from the second, even while the second is still working. Then it continues its

12:52 work with that answer. Once [snorts] you have many sessions running, you can’t watch all of them. And when Claude needs something from you, it asks you a question and waits for your answer. So unless you’re watching that session, you won’t know Claude is stuck. You might return later and find Claude still waiting. To fix that, there’s a setting called ask user question timeout. You need to add it to the settings.json file in the main.claude folder and set it to 60 seconds or any other time that you want. When that time is up, Claude picks the suggested answer on its own and keeps working. So when you return,

13:23 Claude is still working instead of waiting on a question. But Claude only does this when it can move forward without you. If a question needs your answer before Claude can go any further, it doesn’t pick any option. Instead, it keeps waiting until you reply. And with [snorts] many sessions open, you also don’t know which one is waiting for your reply and which one is already done. To fix that, you need to run the config command and find local notifications further down the list. When you press enter on it, you can choose the kind of notification you want. We chose terminal bell, but you can choose any of them.

13:53 With this setting on, Claude Code lets you know when Claude is done or needs your answer. When you switch to plan mode, Claude code uses the same model you started your session with, but you can change that when you run the model command and type opus plan after it. Claude code uses opus only in plan mode. For everything else, it uses the sonnet model. So the plan gets the stronger model and the building runs on sonnet. Opus plan also saves your usage because sonnet uses less of your limit than opus. That matters most on the pro plan because on that plan you’ll hit your

14:24 limit a lot faster. The [snorts] sub aents Claude starts use your limit to and you can choose which model they use. By default, each sub aent uses either the same model as your session or a model Claude code picks for it. To make all your sub aents use one smaller model, you can set claude code sub aent model to the whichever one you want. Then you set claude code sub aent model force to one. You need this second setting because without it, a sub aents own file can still choose a different model. But if you do want a sub aent’s own file to choose it, you can skip this setting. Both settings go in the

14:55 settings.json file in the main.claude folder. Claude code reads these settings when it starts. So from your next session, all your sub aents use the model you chose. Aside from these settings, there are some keyboard shortcuts that also help in your workflow. The first one is prompt stashing. When you write a long prompt and then remember another prompt you wanted to send first, you can just press Ctrl and S. Claude Code basically saves the prompt and empties the prompt box. So when you send the other prompt, your previous prompt returns back and you can send it later on. If you clear the whole

15:26 prompt by mistake, you don’t have to type your prompt again. You can just press control and Y and your deleted text returns. Other times you change a prompt and want to undo that change. You can just press control and the underscore key and the prompt returns to how it was before your last change. And lastly, if you want to learn Claude Code’s features, it has a command for that too called powerup. It gives you short lessons inside Claude Code. And each lesson shows a feature with an animated demo so that you can learn interactively while working with it. Now all the skills and workflows and

15:57 everything else we build and show you in all our videos can be found in AIABS Pro which is our community. So if you found value in what we do and want to support the channel, this is the best way to do it. The links in the description. That brings us to the end of this video. If you’d like to support the channel and help us keep making videos like this, you can do so by using the super thanks button below. As always, thank you for watching and I’ll see you in the next one.