These New Claude Code Settings are INSANE!

Key Takeaways

These new Claude code settings are insane. Most people just install Claude code and run it with the defaults, but there are hidden settings buried in config files that change [music] everything. We're talking better agent control, less annoying permission prompts, a context window that doesn't fall apart halfway through a big task, and nobody's talking about this stuff. Today, I'm breaking down 12 settings that will make [music] Claude code actually work the way it's supposed to. Let's go. Right, Claude code, it's Anthropic's AI coding agent that lives in your terminal. It understands your whole code [music] base, it can write code, fixes bugs, manage workflows all through natural language. It's powerful, but here's the thing most people miss. The default settings [music] are just okay. The real power is in the config, and most developers never touch it. So, let me show you what's actually possible. First, let's talk about where all this stuff lives. Cuz this is where people get confused. There are multiple config files, and they interact in ways that aren't obvious. You've got the legacy file at at Claude.json, which is kind of a chaotic grab bag of global settings, project configs, and history. Then you got user settings at /user/claude/ settings.json, which apply to every project. And then you've got project-level settings that live inside your repo. Hey, if we haven't met already, I'm the digital avatar of Julian Goldie, CEO of SEO agency Goldie Agency. While he's helping clients get more leads and customers, I'm here to help you get the latest AI updates. Julian Goldie reads every comment, so make sure you comment below. Claude uses a scope system to determine where configurations apply. If a permission is allowed in user settings, but it's not allowed in project settings, the project setting takes precedence, and the permission is blocked. So, the hierarchy matters. Keep that in mind as we go through these settings. Higher scope wins. You can also open the settings interface directly, run the config command when using the interactive REPL, and it opens a tabbed settings interface where you can view status information and modify configuration options. So, you don't always have to edit JSON by hand, but knowing the underlying config means you actually understand what's happening. All right, let's get into the settings. Setting number one, the auto compact threshold. This is huge. Claude's context window is its working memory. As it fills up, Claude starts losing track of earlier details, and quality drops. The default auto compaction threshold is around 95%. By then, you've already lost coherence. You're most of the way through a complex task, Claude is starting to forget stuff, and it hasn't even compacted yet. That's the default behavior. The fix is to drop the threshold. Setting it to 60 to 75% works well for most tasks. Claude will update the Claude settings.json for you. Under the hood, he adds the Claude auto compact APC override environment variable. You want that memory fresh. Don't let it get to 95% before it cleans up. Set it lower, and your sessions will feel way more coherent. And speaking of what's possible with AI automation in your business, if you want to learn how to deploy Claude code and tools like it to automate your business, save hundreds of hours, and get ahead of your competition, that's exactly what we do inside the AI Profit Boardroom. It's where serious business owners come to learn how to actually apply AI, not just talk about it. We break down the tools, show you the workflows, and help you implement them in your specific business. If you want in, check the description and comments below. Setting number two, the terminal output limit. The default terminal output limit is set to 30,000 characters, which can restrict your ability to fully use the 1 million token context window. So, Claude can technically hold a million tokens of context, but the output is being choked at 30,000 characters by default. You're leaving capacity on the table. If you're working on big code bases or doing deep analysis, bump this up. It's a simple config change, and it makes a real difference when you're dealing with large outputs. Setting number three, path-specific rules and the file reading limit. You can configure path-specific rules within the .claude folder. These rules allow you to define specific behaviors for directories, such as prioritizing certain file types or excluding irrelevant files. And the default file reading limit is 2,000 lines, which is good just for handling larger data sets or code bases. Setting number four, data retention and privacy. Claude code retains conversation data for 30 days by default. However, this could be adjusted in the settings.json file. For long-term projects, you might extend the retention period. For privacy-focused workflows, disabling data storage entirely is an option. If you're working on anything client-sensitive or anything you don't want going anywhere, this matters. One setting, eight, done. Now, here's where it gets really interesting, because we're only scratching the surface of what's possible. Hooks with exit codes. Hooks are tools for enforcing workflows. You can configure them to prevent undesired actions, enforce task-specific rules, and ensure iterative processes are completed as intended. The way they work is simple. A hook runs before or after a tool call. Based on what it finds, it either lets Claude continue or stops it dead. Two blocking hooks in particular are really useful as defaults. One blocks rm -rf commands and suggests using trash instead. The other blocks direct pushes to main or master branches and requires feature branches instead. Both read the bash command from stdin, match with a pattern, and exit with an error that tells Claude what to do instead. So, you're not just hoping Claude does the right thing, you're enforcing it at the system level. That's a completely different level of control. Setting number six, the sandbox. Claude code has a native sandbox that provides file system and network isolation using OS-level primitives. I'm not sure what OS it uses. I'm not sure what OS it uses. You enable it by typing sandbox in a session. Without the sandbox, deny rules only block Claude's built-in tools. Bash commands can bypass them. With the sandbox enabled, the same rules are enforced at the OS level. So, bash commands are also blocked. You can configure specific network domains that are allowed, control which paths Claude can write to, and even block reads from sensitive files like gcloud credentials. If you're running Claude code in any kind of production-adjacent environment, or if you're letting it operate with less supervision, you want the sandbox on. Setting number seven, sub-agent configurations and agent teams. Sub-agent configurations and agent teams allow for collaborative workflows between multiple agents. This is where Claude code goes from being a single AI assistant to being a system. Think about what that means for running a business. You can have one agent focused on content, one on technical SEO, one on client reporting, all running in parallel, all operating under the same rules you've defined in your settings. We use workflows like this inside the AI Profit Boardroom to show members how to run lean operations with AI handling the repetitive heavy lifting. The managed settings.d drop-in directory lets separate teams deploy independent policy fragments that merge alphabetically, so different teams can have their own config fragments without stepping on each other. Clean, scalable. Setting number eight, claude.md. This isn't technically a settings.json setting, but it has more impact on Claude's behavior than almost anything else. It's a markdown file in your project root that Claude reads at the start of every session. You can tell Claude what to preserve during compaction by adding instructions to your claude.md. For example, when compacting, always preserve the current file paths being edited, test failure messages, and architecture decisions made this session. When compact runs, either manually or automatically, the important context survives. Claude.md is where you define how Claude should think about your project, what matters, what to avoid, what conventions to follow. It's your standing brief for the agent. Setting number nine, custom output styles. You can create custom output styles as files in /claude/output_styles. You create a markdown file with YAML front matter defining the style, and it shows up as an option in right-click config. The styles are markdown files with YAML front matter defining styles that are structured. Setting number 10, the effort command and effort levels. Effort levels have been simplified to low, medium, and high. There's an effort/ command to set the model effort level for the current session, and effort auto resets it to default. This matters because different tasks need different amounts of thinking. A quick bug fix doesn't need the same effort as architecting a new system. By setting effort explicitly, you get faster responses when speed matters, and deeper thinking when quality matters. You're in control. Setting number 11, the BTW feature. This one is worth its own video, but the short version is that /btw lets you send context to Claude mid-response. Say Claude is working on something, and you realize you need to add a detail, you don't have to wait. There's a fix so that our BTW now correctly includes pasted text when used in an active response. You can drop a note in, and Claude weaves it in without starting over. For long-running tasks, that's massive. Setting number 12, permissions and the bypass flag. The dangerously skip permissions flag is intentionally scary. It lets Claude run any command and edit any file without asking. I'm not telling you to use it right now. The name exists for a reason, but it exists. It's a setting. And once you genuinely understand what Claude code does and how it behaves in your code base, it becomes a real option for fully automated workflows. There are also two other permission modes worth knowing about. You can alias Claude permission mode plan to start in plan mode, which is read-only with no edits, and Claude resume to pick up the last session right where you left off. Start in plan mode when you're scoping, jump to full mode when you're ready to build, resume when you're continuing a task. These three aliases alone will change how you work. If you want to go deeper on this, inside the AI Profit Boardroom, we break down exactly how to configure AI tools like Claude code to automate real business operations. We show you how to set up agents that work autonomously, save you hours every week, and handle the tasks that don't need a human. Links are in the description and comments. Come and see what people are building in there. And if you want the full process, SOPs, and over 100 AI use cases like this one, join the AI Success Lab. It's completely free. 38,000 members are already in there using AI to work smarter. You'll get all the notes from this video, access to the community, and a library of AI workflows you can start using today. Link is in the comments and description.