What’s the First Thing You Do When You Move Into a New Place?

Picture this: you move into a new apartment. The layout is fine, the natural light is nice, but the art on the walls isn’t yours, the bookshelf is three centimeters too short, and the shower pressure is never quite right. It’s livable, but it doesn’t feel like home.

Boris Cherny — the creator of Claude Code — dropped a 12-tweet thread on X about how to turn Claude Code from “a nice apartment” into “your home.” It all comes down to one word:

Customizability.

Reflecting on what engineers love about Claude Code, Boris found the #1 answer isn’t “it’s smart” or “it’s fast” — it’s that you can customize literally everything. His opening message:

Every engineer uses their tools differently. We built Claude Code from the ground up to not just have great defaults, but to also be incredibly customizable. This is a reason why developers fall in love with the product.

Clawd Clawd 吐槽時間：

Boris using the phrase “fall in love” isn’t marketing fluff. Have you ever used a tool and thought “oh my god, this was made for me”? That feeling comes from customizability. VS Code crushed Sublime Text because of its extension ecosystem. Now Claude Code is walking the same path — except this time you’re not customizing syntax highlighting, you’re customizing an entire AI agent’s behavior (◕‿◕)

Now, I’m not going to walk through all 12 tweets one by one like reading a slide deck. Instead, let me reorganize them into a few story arcs — from “moving in” all the way to “inviting friends to see the finished place.”

Step One: Make Yourself Comfortable

When you move into a new apartment, you don’t start by picking curtains. You check that the water works, the Wi-Fi connects, and the toilet flushes. Boris’s first two tips follow the same logic — get the basics right.

Terminal side: /config for light/dark mode, iTerm2 native notifications (other terminals use hooks), /terminal-setup for Shift+Enter newlines, /vim for Vim mode. All one-liner commands.

Then “how hard should the AI think.” Run /model to pick effort level — Low (fewer tokens, faster), Medium (balanced), High (smarter but slower). Boris himself? High for everything. No surprise — this is the guy who ships 27 PRs per day (see CP-14). For him, quality always beats speed.

It’s like ordering ramen and choosing broth intensity: light, regular, or extra rich. Boris always orders the “extra rich with extra pork fat.” Not everyone’s stomach can handle it, but the results are undeniably better (￣▽￣)⁠／

Clawd Clawd 碎碎念：

Shift+Enter for newlines sounds tiny, but it’s a massive quality-of-life upgrade. Writing multi-line prompts with backslashes felt like texting with a flip phone in 2026. And when Vim mode dropped, Reddit was full of Vim users crying tears of joy — their devotion is about as intense as “people who use Linux making sure you know they use Linux” (⌐■_■)

Plugging Into the Ecosystem: Plugins, Agents, Skills

Basics done, but you’re still living in an empty apartment. Next step is like going to IKEA to furnish the place — except you don’t need to assemble anything. The community already did it for you.

Claude Code now has a plugin marketplace. Run /plugin to install pre-built LSPs, MCPs, Skills, custom agents, and hooks from Anthropic’s official marketplace. You can also create a private marketplace for your company, then commit settings.json to your codebase — teammates auto-load all marketplaces when they clone the repo.

Clawd Clawd 溫馨提示：

Hold on, let me process this. Claude Code now has a plugin marketplace ヽ(°〇°)ﾉ

History lesson: VS Code’s killer advantage in 2015 was its Extension Marketplace. Chrome beat Firefox partly because of the Extension Store. Every great developer tool eventually grows a plugin ecosystem. Claude Code has reached that stage. This isn’t the “write a bunch of config yourself” world anymore — it’s the “the community already built it, just install” world.

And private marketplaces for companies? Imagine a team of 50 people where your code review agent, testing skill, and deployment hooks are all packaged as plugins. Day 1 onboarding becomes a single install command. Now THAT’s onboarding (๑•̀ㅂ•́)و✧

But the plugin marketplace is just “furniture.” What Boris talks about next — custom agents — is how the apartment develops your personality.

Drop .md files in .claude/agents to create custom agents. Each agent gets its own name, color, tool set, permission mode, and model. Boris also revealed a hidden feature: set the "agent" field in settings.json or use the --agent flag to make a custom agent your default main conversation agent.

Imagine having one agent for code review, one for writing tests, one for refactoring — each with different models and permissions, like specialized engineers on your team with different clearance levels. You could create a “Security Auditor” agent with Read-only permission — no Write access. Way safer than “use one Claude for everything and pray.”

Clawd Clawd 忍不住說：

The words “hidden feature” are basically catnip for engineers. Every time someone says “there’s a little-known feature…” the bookmark count explodes (¬‿¬)

But here’s the deeper point: Boris designed agents to work like engineers on a team — different people handle different things, different people get different access. This is basically principle of least privilege applied to AI. You wouldn’t give an intern root access, so why give every AI agent the same permissions? (ง •̀_•́)ง

How Safety Works: From Keys to Sandboxes

So your AI toolbox is stacked now. But here’s the problem — clicking “Allow” for every single command is like entering a password, scanning your fingerprint, and saying “it’s really me” into a camera every time you open your own front door. Secure? Very. Annoying? Extremely.

Boris gives two layers of solutions.

Layer one: allowlists. Run /permissions to precisely open the commands you need. Wildcard syntax — Bash(bun run *) approves all bun run commands, Edit(/docs/**) approves editing anything in the docs folder. Boris said in CP-14 that he doesn’t recommend --dangerously-skip-permissions. The right approach is carrying a key, not removing the door.

Layer two: sandboxing. If allowlists still feel like too much work, Claude Code has an open source sandbox runtime. Run /sandbox to enable it — supports file isolation and network isolation. Commands inside the sandbox need zero permission prompts because they’re already isolated. They can’t break your system.

Think of these two layers like home security: allowlists are “giving family members a key,” sandboxing is “setting up a fenced play area for kids — they can do whatever they want inside and nothing breaks.”

Clawd Clawd 吐槽時間：

Permission management reflects Claude Code’s design philosophy: Security is the default. Convenience is opt-in.

A lot of people find the permission prompts annoying on day one and skip straight to skip-permissions. That’s like unbuckling your seatbelt because it’s uncomfortable — 99% of the time nothing happens, but that 1% you’ll wish you hadn’t ┐(￣ヘ￣)┌

The sandbox is the ultimate solution. You no longer have to agonize over “do I trust this AI.” It’s in a sandbox — build castles, start fires, whatever — it won’t burn down your living room. For people running hundreds of AI sessions a day, this is the cure for permission fatigue ╰(°▽°)⁠╯

The Tiny Things You Can’t Live Without

Next, Boris covered three features. Individually, none of them deserves a whole article. But together, they answer a bigger question: how does a tool go from “very useful” to “mine”?

Status line — an info bar below the composer. Show current model, working directory, remaining context, cost, whatever you want. Boris says everyone on the Claude Code team has a different status line. Run /statusline to set yours up. The context window is like your phone battery — you always want to know how much is left before you get the dreaded mid-conversation cutoff.

Keybindings — run /keybindings to remap any key. Settings live reload — changes take effect immediately, no restart needed. Live reload means you can tweak and test in real time, like adjusting an audio equalizer while listening. Many tools require a restart for setting changes, which makes the experience so bad you never bother customizing.

Spinner verbs — that spinning text while Claude thinks? You can customize it. Practical value: approximately 0.01 out of 10. But Boris’s exact words: “It’s the little things that make CC feel personal.”

Clawd Clawd 補個刀：

Okay, I’ll admit spinner verbs are pure “engineer aesthetic.” But notice that Boris put them in the same thread alongside hooks and plugins? That’s his message — a tool feeling like “mine” doesn’t come from one killer feature. It comes from a hundred small details adding up.

Think about it: iPhone’s haptic feedback, the typing vibration intensity, keyboard sounds — the “practical value” of these features is basically zero. But without them, the phone feels “not mine.” Claude Code’s spinner verbs work exactly the same way (◕‿◕)

I can already see someone setting their spinner to “secretly judging your code…” or “pretending to work…” Me? I’d go with “questioning life choices…” (￣▽￣)⁠／

Two Features That Change Your Relationship With AI

Boris’s thread so far has covered looks, ecosystem, security, and small details. But these next two features operate on a completely different level — they don’t change what Claude Code looks like, they change how it interacts with you.

Hooks let you deterministically inject custom logic into Claude’s lifecycle. Boris listed some use cases: auto-route permission requests to Slack or Opus; nudge Claude to keep going when it finishes a turn (you can even use another agent to decide whether to continue); pre-process or post-process tool calls with your own logging.

Boris shared in CP-14 that his PostToolUse hook auto-runs bun run format. Now hooks can do even more — including “using another AI to decide whether to continue.” That’s agent-of-agents territory, recursive enough to make your head spin.

Clawd Clawd 畫重點：

Hooks are one of Claude Code’s most powerful and underrated features.

Picture this pipeline: Before Claude runs any command, your hook checks “is this safe?” After Claude writes code, your hook auto-runs formatter + linter. When Claude finishes a task, your hook sends a Slack notification. You’re basically building a CI/CD pipeline for your AI.

And “using another AI to decide whether to continue” — you’re creating a structure where AI manages AI. Add custom agents with different models and permissions… congratulations, you just invented the AI version of an org chart ヽ(°〇°)ﾉ

The other one is Output Styles. Run /config to set how Claude talks to you. Boris recommends two built-in styles: Explanatory (great for exploring new codebases — Claude explains frameworks and patterns as it works) and Learning (Claude coaches you through code changes instead of just doing it for you). You can also create custom output styles.

Here’s the thing most people miss: how good an AI feels depends less on the model’s raw intelligence and more on how it talks to you. Same professor, slightly different teaching style, and the end-of-semester evaluation drops from 4.8 to 3.2. Model didn’t change. Knowledge didn’t change. But the student experience is worlds apart. Output style adjusts exactly this — you’re not swapping the AI’s brain, you’re swapping its teaching style.

Clawd Clawd 真心話：

Learning mode is genuinely amazing for beginners. Before, AI just dumped a wall of code in your face and you’d close the tab after three seconds. Now it holds your hand: “Look at this first… notice this pattern… what do you think should change here?” The difference is like a teacher writing the answer on the board vs. Socratic questioning ┐(￣ヘ￣)┌

And Explanatory mode is for veterans exploring unfamiliar codebases — like having a local guide walking you through a night market: “this stall has been here 30 years,” “that alley is a shortcut.” You could explore alone, but having a guide is a completely different experience.

37 + 84 = Turning AI Config Into Infrastructure as Code

Boris’s final tweet is the grand finale:

Claude Code is built to work great out of the box. When you do customize, check your settings.json into git so your team can benefit, too.

The numbers: 37 settings + 84 environment variables. No wrapper scripts needed for env vars — just use the "env" field in settings.json. Configuration levels span from per-codebase to per-sub-folder to per-user to enterprise-wide policies.

But the numbers aren’t the point. Boris’s most important line is “check your settings.json into git.”

What does this mean? Your Claude Code config becomes Infrastructure as Code. Your team’s AI tool settings live alongside tsconfig.json, .eslintrc, and Dockerfile as part of the codebase. New hire clones the repo, all AI settings are in place. Someone changes a setting, it shows up in PR review.

Clawd Clawd 吐槽時間：

37 settings + 84 env vars. The number of possible combinations is so astronomical that you’ll never find two Claude Code setups that are exactly the same (╯°□°)⁠╯

But this isn’t just customization — it’s standardization through customization. Sounds contradictory, right? But think about it: every developer’s VS Code looks different, but the same team shares .editorconfig and .prettierrc. Claude Code does the same thing, except the dimension goes from “formatting” to “AI behavior.” That’s the design genius right there (๑•̀ㅂ•́)و✧

Back to That Apartment

Notice how Boris’s previous viral thread (CP-14) was about “how I use Claude Code” — an open house, showing you around his place. This thread is about you. He’s telling you: “Hey, you can knock down this wall, adjust that bookshelf, set your own shower pressure.”

And when you’re done decorating, git commit your settings — your roommate moves in and gets the same setup. Add the plugin marketplace, and now there are even “community-designed model apartments” to browse.

Claude Code’s accelerating growth isn’t just because the AI model got smarter (though it did). The real flywheel is this: every user is turning Claude Code into their own tool. And once a tool becomes “my tool,” you don’t switch.

Clawd Clawd 忍不住說：

Not a single one of Boris’s 12 tips teaches you how to make the AI “smarter.” Every single one teaches you how to make the tool “yours.” From terminal themes to agent permissions to spinner verbs — yes, spinner verbs — his message is crystal clear: what makes you feel “this is mine” isn’t how powerful the AI is. It’s every small detail being exactly the way you want it.

Anyway, I’m off to customize my spinner verbs now. Don’t look for me for the next three hours ╰(°▽°)⁠╯