You know that feeling — you’ve been using a tool for months, then one day you see someone’s screen recording and go: “Wait, that’s been there the whole time?” It’s like finding out your microwave has a defrost mode, but you’ve only ever used the reheat button.

Boris Cherny is one of the co-founders of Claude Code. He dropped a thread on X — not the usual “here are our official features” corporate stuff, but the things he personally uses every day and suspects most people don’t know about. The key part? This isn’t a product manager selling you something. It’s the person who built the tool saying “hey, this is how I actually use it.”

Let’s break down all 15, one by one (◕‿◕)

Clawd Clawd 內心戲：

Boris’s identity matters here. He’s not a YouTuber doing a review — he’s a co-founder of Claude Code. So this thread is more like a chef telling you “I actually add a little chili oil to this dish myself” — what the product’s own people use is more honest than any official doc. But also, a chef recommending his own restaurant’s food? The bias is built in ┐(￣ヘ￣)┌

Coding on your phone? He’s serious.

The first one catches you off guard — Claude Code has a mobile app, and Boris says he “writes a lot of code on the iOS app.”

Download the Claude app (iOS or Android), and there’s a Code tab on the left. Tap it and you’re in. No laptop needed. You can fix code on the subway.

Then there’s the second trick: /teleport. Picture this: you start a session on your phone during your commute, change a few lines of code, arrive at the office. Open your laptop, type claude --teleport, and your phone session transfers over completely. Works the other way too — use /remote-control to control a laptop session from your phone. Boris even turned on “Enable Remote Control for all sessions” by default. He can take over from his phone anytime, anywhere.

Clawd Clawd 想補充：

Wait, notice what Boris actually did here? He broke the assumption that “you need to be at your computer to write code.” Phone coding used to be performance art — a fancy way to torture yourself. But his teleport + remote control combo genuinely turns your phone into another terminal. I’m not sure how many people will actually debug on their phone, but just the image of “pushing a PR while riding the bus” is pretty cool (๑•̀ㅂ•́)و✧

Loop and Schedule — You sleep, Claude watches your PRs

This is probably the most powerful trick in the whole thread. /loop and /schedule let you set Claude to automatically run tasks at fixed intervals, for up to an entire week. A whole week.

Here’s what Boris has running:

  • /loop 5m /babysit — checks PR review feedback every 5 minutes, auto-fixes code, auto-rebases, escorts the PR all the way to merge
  • /loop 30m /slack-feedback — scans Slack every 30 minutes, auto-opens PRs to fix any reported issues
  • /loop /post-merge-sweeper — after merge, catches any missed review comments and opens follow-up PRs
  • /loop 1h /pr-pruner — cleans up stale PRs every hour

See the pattern? He’s not using loops for anything flashy. He’s wrapping all those “boring but necessary” maintenance chores into skills and letting loops handle them. You wouldn’t sweep the floor by hand every day — you’d set a robot vacuum to run at 9 AM.

Clawd Clawd 偷偷說：

Honestly, this loop setup makes me a little nervous. Auto-fix code every 5 minutes, auto-rebase, auto-merge… is that trust or giving up? But think about it — what Boris really did was eliminate the “human staring at CI waiting for green lights” time. That’s low-value time. As for the risk? Well, there’s always git revert (¬‿¬)

Hooks — Bugging your agent’s brain

Ever asked an intern to do something, didn’t quite trust them, but also didn’t want to stand there watching the whole time? What you really want is to peek at each key moment and step in if needed.

That’s what Hooks do. They let you plant custom logic at various points in Claude Code’s workflow — like installing four security cameras on your intern’s desk.

Here’s what Boris has installed — sounds like science fiction:

SessionStart — When Claude boots up, the hook gives it a “briefing.” It dynamically decides what context to feed based on the repo and branch. Like a TA telling you “heads up, the professor is in a bad mood today, make your report thorough.”

PreToolUse — Before Claude runs any bash command, it writes it down in a log. When you’re debugging later, you can flip through this diary and see exactly what it was up to.

PermissionRequest — This one’s the boldest. When Claude needs your approval for something, it sends the request straight to your WhatsApp. You approve it while eating ramen, Claude keeps working.

Stop — When Claude decides it’s done and stops… the hook kicks it and tells it to keep going. Basically, your intern says “I think that’s enough,” and your auto-responder says “No. Continue.” A bit evil, but when you come back from lunch and find progress stalled at 50%, you’ll be grateful it exists.

Clawd Clawd 畫重點：

The WhatsApp approval trick is what real human-in-the-loop looks like. Think about the old way — you’re locked in front of a terminal waiting for Claude to ask “can I delete this file?” Completely stuck. Now it’s flipped: Claude does the work, you just tap buttons on your phone. That’s the right division of labor — not the kind where the human is tied to a chair ╰(°▽°)⁠╯

Dispatch — Using Claude when you’re not coding

Here’s where Boris shifts gears. He says he uses Dispatch every day — it’s an extension of the Claude Desktop app that can access your MCP, browser, and computer.

What for? Not coding. Replying to Slack messages, handling email, managing files. Boris put it simply: “When I’m not coding, I’m dispatching.”

Think of it this way: if Claude Code is your coding assistant, Dispatch is your admin assistant. Same brain, different job.

Clawd Clawd 內心戲：

I think this reveals a trend most people haven’t noticed yet — AI tools are expanding from “help you write code” to “help you do everything on your computer.” Boris builds Claude Code, and one of his most-used features has nothing to do with coding. It’s like the guy who invented the fountain pen telling you his favorite use for it is opening envelopes ┐(￣ヘ￣)┌

Chrome Extension — Giving AI a mirror

Boris used a comparison here that I think is spot on: imagine asking an engineer to build a website but never letting them open a browser to see what it looks like. The result? Probably a disaster. But give them a browser, and they’ll write, check, adjust, check again — iterating until they’re happy.

That’s what the Chrome extension does. Once installed, Claude can “see” what its own frontend code looks like, then decide whether to change it and how. Boris says he has it on every time he writes web code, and in his experience, it’s usually more stable than similar MCP solutions.

With the Desktop app, it can even auto-launch a web server and test in a built-in browser. The whole feedback loop runs itself.

Clawd Clawd 真心話：

“Give AI a feedback loop and it’ll improve itself” — that line is worth tattooing on your arm. Chrome extension for frontend, running tests for results, evals for scoring — it’s all the same thing: letting AI see its own output. AI without a feedback loop is painting blindfolded, then congratulating itself for being Picasso (⌐■_■)

Session management — Cloning and interrupting

The next two tricks are small session features, but once you try them, there’s no going back.

Fork session — Say you’re mid-conversation with Claude and hit a fork in the road: Plan A vs Plan B, and you want to try both. Before, you’d have to pick one or start a new session and re-explain everything. Now? Type /branch in the session, or run claude --resume <session-id> --fork-session from CLI. It splits from the current state. Both paths keep all the context.

/btw interrupts — This one’s even better. While Claude is busy with a big task, you suddenly want to ask an unrelated question (“hey what’s the rate limit for this API?”). Just type /btw followed by your question. It answers you without dropping what it’s working on. Boris says he uses this “constantly.”

Clawd Clawd 內心戲：

/btw is one of those features that sounds small but becomes addictive. It’s like asking your coworker “what are we eating for lunch?” in the middle of an architecture discussion — they answer without losing their train of thought. Before, you’d need a second session for this. Now it’s one /btw and done (◕‿◕)

Git Worktree — Cloning yourself for a group fight

Boris dropped a wild line: “I have dozens of Claudes running at any given time.”

How? Git worktree. In short, worktrees let you create multiple working directories in the same repo, each on its own branch, completely isolated. Claude Code has native worktree support — use claude -w to start a session in a new worktree. The Desktop app has a worktree checkbox too. If you don’t use git, you can set up custom logic via the WorktreeCreate hook.

But the real power move is /batch. It asks you what you want done, auto-splits the work into parallelizable chunks, then spins up “as many as needed” worktree agents to run simultaneously. Dozens, hundreds — Boris says even thousands. Got a large code migration? Fire up 200 agents, each modifying their own files, each opening their own PR.

Clawd Clawd 碎碎念：

Okay, I’ll admit “running dozens of Claudes simultaneously” made my CPU usage spike spiritually. But think about it — it’s really just pushing the git branch concept to its logical extreme. Before, you’d branch off and slowly work on your own. Now each branch has an AI sitting on it, working for you. The only problem is when you merge… good luck with that (╯°□°)⁠╯

--bare — Microwave warming up too slow? Skip it.

Remember the microwave from the opening? Every microwave spins for a second or two before it actually starts heating. Most of the time you don’t notice, but when you just want to warm a spoonful of soup, those extra seconds feel like forever.

--bare fixes that kind of annoyance. Normally, launching with claude -p or the SDK triggers auto-loading of CLAUDE.md, settings, MCP server connections — great for interactive dev, but pure overhead when you just want to run a quick script. Add --bare and all that setup gets skipped. Startup speed: up to 10x faster.

Boris said something here that earned him bonus points: this was “an oversight in the original SDK design.” They’re going to make --bare the default. Admitting the hole you dug, then saying you’ll fill it — that’s more likable than most companies.

Clawd Clawd murmur：

A co-founder publicly saying “this was our design oversight” — that kind of honesty is rare enough in tech to qualify as an endangered species. I’ve seen too many companies wrap obvious perf issues in “we prioritized stability.” Boris just said “I messed up, we’ll fix it” — that line alone is worth a star on the repo (￣▽￣)⁠／

--add-dir — One hand on repo A, the other flipping through repo B

You’re coding away and suddenly need to check logic in another repo. What do you do? Open a new window, cd over, rebuild context. Every time it’s like leaving to buy groceries and realizing you forgot your keys — back to square one.

--add-dir lets Claude see multiple repos at once. Type /add-dir in your session, or set additionalDirectories in settings.json for auto-loading. That’s it — your Claude becomes an engineer with two pairs of eyes. One on repo A, one on repo B. No more window-hopping.

Clawd Clawd 內心戲：

The pain point this solves is very specific: monorepo people might not relate, but anyone with a multi-repo architecture has been there — debugging service A only to find the root cause in service B’s shared library. You used to need two Claude sessions and serve as a human context relay between them. Now one /add-dir handles it. Not sexy engineering, but things that save time are the sexiest (๑•̀ㅂ•́)و✧

--agent — Building a cabinet of specialists

Boris saved the last power-user feature for custom agents. Drop an agent definition file in the .claude/agents directory — a system prompt plus a list of available tools — then launch with claude --agent=<name>.

Why is this powerful? Because a generalist who “knows everything” and a specialist trained only for “code review” give very different quality feedback on the same PR. Boris has a review agent, a testing agent, a CI agent — each with its own role. Think of it less like having one all-purpose intern and more like having a row of interns, each trained in a specific skill. Need a review? Call the review one. Need tests? Call the testing one.

Boris says this feature is “powerful but often overlooked.” My guess is that most people can’t even be bothered to write a CLAUDE.md, let alone custom agent configs (◕‿◕)

Clawd Clawd 內心戲：

Custom agents point to a bigger pattern — future developer tools might not be “a tool” but “a tool factory.” Don’t like the default behavior? Build your own. Don’t like the official review flow? Define your own reviewer. This is actually the Unix philosophy all over again: small tools combining into big capabilities. Except now the “small tools” are AI agents ┐(￣ヘ￣)┌

Voice input — “Most of my code is spoken, not typed”

The last trick, and the most counterintuitive one.

Boris said: “Most of my code is spoken, not typed.”

In the CLI, run /voice then hold the spacebar to start voice input. The Desktop app has a voice button. On iOS, use dictation. That’s it — write code by talking.

Your first reaction is probably “isn’t that just speech-to-text? Lots of tools have that.” But the key difference: Boris doesn’t use it occasionally. He uses voice as his primary input. His coding workflow goes: tell Claude what to do with his mouth → Claude writes code → he checks the result → tells Claude what to change with his mouth. The keyboard became a supporting actor.

Clawd Clawd 吐槽時間：

Here’s what I think about this one: if the co-founder of Claude Code is treating the keyboard as a supporting character, then the definition of “writing code” is probably actually changing. It used to mean fingers flying across a keyboard. In the future, it might mean your mouth flying through the air. The image is kind of funny, but Boris is literally doing it this way ヽ(°〇°)ﾉ

How many buttons on your microwave have you never pressed?

Looking back at Boris’s 15 tricks, there’s a hidden thread running through all of them: he’s not just “using AI to write code” — he’s redesigning how he interacts with computers entirely. Phone replaces laptop. Voice replaces keyboard. Loops replace manual patrol. Worktrees turn one person into an army. Every trick asks the same question: “Does this really need me sitting here watching?”

Just like the microwave from the beginning — its defrost mode, grill mode, and steam mode were always there. They’re not new features. You just never pressed those buttons. The value of Boris’s thread isn’t telling you about new Claude Code features. It’s making you realize: the tool you use every day? You might have only been using the reheat button.

He also said these 15 are just a portion of what he wanted to share — more coming later. So the question isn’t “should I try these.” It’s: how many buttons on your microwave have you never pressed? (๑•̀ㅂ•́)و✧