The argument in one line.
Claude Cowork lets non-technical users automate routine tasks by giving AI agent control of their files, browser, and applications through natural language commands, making agentic AI accessible to anyone regardless of technical skill.
Read if. Skip if.
READ IF YOU ARE…
- You're a non-technical founder or operator who wants to automate file and browser tasks but has avoided Claude Code because the terminal felt like a barrier.
- A creator or consultant managing repetitive workflows—research, data organization, content processing—and wants to see concrete use cases beyond coding.
- You're comfortable with AI tools but new to agentic workflows and want to understand how to structure prompts and sessions for compounding results over time.
SKIP IF…
- You're already deep in Claude Code or terminal-based AI workflows—this is positioned as an entry point, not an advanced optimization session.
- You're looking for production-scale deployment guidance or enterprise security considerations; this focuses on personal and small-team workflows.
The full version, fast.
Claude Code's creator Boris Cherny runs five to ten parallel sessions simultaneously — not sequentially — and that shift from one-task-at-a-time to tending multiple agents is the biggest workflow change most users haven't made. The method pairs plan mode with auto-accept execution: work with the model until the plan is solid, then let it run uninterrupted, because Opus 4.5's planning accuracy means a good plan almost always produces good code. A CLAUDE.md file checked into git and updated by the whole team every time Claude makes a mistake functions as compounding team memory, eliminating repeated corrections. The third key is giving Claude a way to verify its own output — running tests, spinning up a server, checking a browser — because the model codes blind without it.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 03:26
01 · Intro + what is Cowork
Greg frames Cowork as Claude Code for non-technical people and introduces Boris. Boris explains the product is Claude agent SDK under the hood.
03:26 – 05:51
02 · Agentic defined
Boris draws the line between chat-AI and agents: agents take actions — files, tools, browsers, not just text. This direction has been core to Anthropic from before the models were good.
05:51 – 08:23
03 · Demo: Folder access + receipt renaming
Boris grants Cowork access to a receipts folder and asks it to rename files by date. Introduces reverse solicitation — model asks clarification when unsure instead of assuming.
08:23 – 10:52
04 · Demo: Receipts to spreadsheet
Boris asks Cowork to put the receipts in a spreadsheet. It creates an Excel file, notices its own formatting error mid-task, and self-corrects.
10:52 – 15:52
05 · Demo: Google Sheets + browser control
Boris asks for a Google Sheet instead. Claude opens Chrome, navigates to Drive, creates the sheet, and pastes the data. First live browser-control moment in the demo.
15:52 – 22:07
06 · Demo: Email + parallel tasking
Claude drafts a Gmail to Amy, finds her in contacts, shares the Google Sheet with viewer access, and sends. Boris kicks off a second parallel task mid-demo to prove the workflow.
22:07 – 24:22
07 · How to start with Cowork
Boris recommends starting simple: install Chrome extension, grant one folder, play. Skills are for later when you hit specialist software workflows.
24:22 – 28:44
08 · Where agents go next
Boris: 100% of his code written by Claude Code in the last 2 months. Exponential vs linear planning trap. The toil-some work disappears; you tend to an army of Claudes.
28:44 – 41:12
09 · Boris viral Claude Code setup (13 tips)
Greg shares Boris 99K-bookmark thread. They walk through: parallel terminal tabs with separate git checkouts, iOS + web overflow, Opus 4.5 with Thinking, shared CLAUDE.md as team memory, GitHub Action tagging for compounding engineering, plan-first execution, and verification loops.
41:12 – 42:07
10 · The Claude pronunciation debate
Closing bit: Greg (French-Canadian) says Claude in French. Boris commits to trying it around the Anthropic office.
Lines worth screenshotting.
- Claude Cowork is not a simplified version of Claude Code — it is a different product with a GUI that controls files, Chrome, and desktop apps from plain English, designed for users who will never open a terminal.
- Boris Cherny building both Claude Code and Claude Cowork means the GUI and the CLI share the same underlying agent architecture — Cowork is not a wrapper around a less capable model.
- CLAUDE.md as compounding team memory means every session inherits what all previous sessions taught the agent — the file grows with the project, and the compounding is multiplicative not additive.
- Parallel Claude Code sessions let you run planning in one window and implementation in another — which means the planning session's output becomes the input to the implementation session without waiting for one to finish.
- Plan-first execution prevents the model from committing to an approach before the human has seen the design — which is why the plan mode gate exists: the model writes the plan, the human approves, and only then does the implementation begin.
- Verification loops — checking the output against the intended behavior before ending a session — are the quality gate that prevents the accumulation of undetected errors across long agent sessions.
- Claude Cowork controlling Chrome from plain English converts web-based workflows (form fills, data collection, platform interactions) into tasks that can be delegated to an agent without API access or browser automation code.
- The 13-tip viral setup thread that got 99K bookmarks from Boris Cherny is the document that separates Claude Code beginners from power users — because the tips are not features but disciplines: verification, plan-first, parallel sessions, CLAUDE.md maintenance.
- A GUI-first AI agent that your dad and mom can use without terminal access is the democratization vector that takes AI-assisted development from a developer skill to a general productivity skill.
- The difference between a one-session task and a compounding agent relationship is CLAUDE.md: without it, every session resets; with it, every session adds to a growing behavioral contract that makes the agent more accurate over time.
- Running multiple Claude Code sessions simultaneously on the same project is not a parallel processing trick — it is the correct architecture for separating concerns: one session plans the architecture while another implements the previous plan.
- Cowork's ability to control desktop apps means it can take actions in software that has no API — which is the category of automation that previously required either RPA tools or manual human operation.
- The 99K bookmark count on Boris Cherny's tip thread is the signal that the information gap between 'using Claude Code' and 'using Claude Code well' is large enough that people are archiving the document for repeated reference.
- Verification as a habit — not a feature — is what Boris Cherny identified as the highest-leverage behavior change for Claude Code users: run the output, check it, and only close the session when you've confirmed it does what you intended.
- Claude Cowork solving the terminal-fear problem for beginners is the same product strategy that every developer tool has eventually deployed: build the power version first, then build the accessible version on top of it once the capability is proven.
Build the verification loop. Build the CLAUDE.md.
Two systems worth stealing today
Boris works differently than almost everyone: he never writes code by hand, runs 10 sessions in parallel, and his team gets smarter every PR — because every mistake becomes a permanent rule.
- Create a CLAUDE.md for every active project right now. Check it in. Update it every session.
- Every time Claude does something wrong, add a never-do-X rule. That is the compound flywheel.
- Tag @claude on your own PRs via the GitHub Action to auto-update CLAUDE.md — never repeat feedback twice.
- Start every non-trivial session in plan mode. Do not auto-accept until the plan is solid.
- Give Claude a browser, a test command, or a bash script to verify its output. Quality jumps immediately.
- Run sessions in parallel: kick off 3-5 morning tasks, tend to them between other work.
- For Cowork-style workflows: grant one folder, give one task, let it surprise you before customizing anything.
Terms worth knowing.
- Claude Cowork
- A desktop application from Anthropic that provides a graphical interface for an AI agent that can control files, browsers, and apps through plain-English instructions, currently Mac-only with Windows planned.
- Claude Code
- A command-line coding agent from Anthropic that runs in a terminal and helps developers write, edit, and manage code through natural language, also available in IDEs, browsers, and mobile apps.
- Agentic AI
- AI systems that don't just generate text but take actions in the world — using tools, interacting with software, controlling browsers, and executing multi-step tasks rather than only answering questions.
- Claude Agent SDK
- A programmatic toolkit from Anthropic that lets developers build their own applications on top of the same agent technology that powers Claude Code and Cowork.
- MCP
- Model Context Protocol — an open standard that lets AI agents connect to external tools, services, and data sources through a consistent interface so the same agent can work across many apps.
- Reverse solicitation
- An AI behavior pattern where the model proactively asks the user for clarification when it's uncertain, rather than guessing and proceeding with a wrong assumption.
- Alignment
- The field of AI research focused on making sure models behave in ways their users actually want, including studying model internals to verify safety properties.
- Mechanistic interpretability
- An AI safety discipline that studies the internal neurons and structures of a model the same way neuroscientists study a brain, to understand how it reaches its outputs.
- Prompt injection
- An attack where malicious instructions are hidden in content the AI reads (a webpage, document, or email), tricking the agent into doing something the user didn't ask for.
- Sonnet / Opus / Haiku
- Anthropic's tiers of Claude models — Haiku is fastest and cheapest, Sonnet is the balanced middle tier, and Opus is the largest and most capable, with version numbers like 3.5 or 4.5 indicating generations.
- Skills
- Prepackaged instructions and resources that teach an AI agent how to handle a specific task or file format, like creating Excel spreadsheets or working with a particular software tool.
- Plan mode
- A Claude Code feature where the agent first drafts a written plan for what it intends to do and waits for approval before making any changes to files.
- Auto accept edits
- A Claude Code mode where the agent's file changes are applied automatically without the user confirming each one, typically used after a plan has been approved.
- CLAUDE.md
- A plain text file checked into a code repository that holds project-specific instructions, conventions, and lessons for Claude — effectively a shared knowledge base the agent reads on every session.
- Pull request (PR)
- A proposed set of code changes submitted to a shared repository for review and merging, usually on GitHub. Often abbreviated PR.
- Spec-driven development
- An approach where you write a detailed specification of what software should do before writing any code, then build to match the spec.
- Git checkout
- A working copy of a code repository at a specific branch or version. Keeping multiple checkouts lets you work on several independent changes in parallel.
- Worktrees
- A Git feature that lets you have multiple working directories from the same repository checked out simultaneously, useful for juggling parallel branches without switching.
- Lint rule
- An automated check built into a codebase that flags or rejects code violating a specific convention, used to enforce standards without manual review.
- Compound engineering
- A practice popularized by Dan Shipper where every code review or AI interaction feeds lessons back into the system's instructions, so the agent gets steadily smarter on your codebase over time.
Things they pointed at.
35:40conceptDan Chipper Compounding Engineering
Lines you could clip.
05:01
“The biggest difference with agents is it can take action.”
15:15
“Now is the age of multi-Clauding, of parallelism, of not going super deep on stuff, but kinda being more of a generalist and tending to your Claudes.”
24:22
“In the last two months, Claude Code has written 100% of my code. I haven't written a single line by hand.”
33:00
“Once the plan is good, the code is good.”
39:00
“Give Claude a way to verify its work and it'll be much better.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00Claude Cowork is here. And if you understand how to use it, you're gonna be able to outperform 99% of people on this planet.
00:07It is an easy way for you to use Claude Code. You might have heard of Claude Code. It's gone viral.
00:13But the problem is it feels technical. You have to go into the terminal and it's not fun for a lot of beginners.
00:19So Claude has come up with Cowork, and it's a brand new product that harnesses the power of Claude Code in a UI that is simple, that anyone could use, that your dad can use, your mom could use, hey, even you can use. So in this episode, I brought on Boris, the maker of this product.
00:37I'm so excited that he shows you the best practices for how to use Claude Cowork, and at the end, how he sets up his Claude code to get the most out of it.
00:49You're gonna love this episode.
00:59We are lucky. We've got Boris. He's the creator of Claude Code, and I would say the cocreator of Claude Cowork today.
01:07And today, what I wanna get accomplished is everyone's talking about Cowork, and I wanna best understand what are the use cases, how can we get started, what are some non obvious ways I can use it. So Boris, thank you so much for coming on the show. And I have one question for you.
01:24By the end of this episode, what are people gonna get out of it?
01:29Oh. I think people are gonna start to get some more ideas for how to use co work.
01:35And hopefully they're gonna tweet about it and maybe they'll even reach out to me so I could learn how they wanna use CoWork. It's funny.
01:43I feel like QuadCode from the beginning, it was sort of built to be it wasn't actually built to be a product at all. But when we first started thinking about it as a product, we thought that the thing people would use it for is coding.
01:55But we've learned very quickly that people didn't just use it for coding. They used it for, like, all sorts of stuff. So I I kind of feel the journey of Cloud Code has been very surprising.
02:04And I have been learning so much just watching how people use the product and how they abuse it and kinda like what what they actually want to use it for and even if it's not designed for that. So I kinda feel the same way about Coworka. Know, I have some ideas about what people are gonna use it for.
02:19I kind of view these as hypotheses. And so, you know, happy to talk about it. Happy to maybe I'll do a quick demo to to show you kind of some of the some of the things that we use it for.
02:29But I think that it's gonna be pretty surprising, and I hope that I will be surprised when we see how people actually use it in the wild.
02:38Yeah. I think that's always the case with with platforms especially if you think about it like when when the makers of the App Store created the App Store, you know, the the initial apps was, like, you know, a beer drinking app and, like, a bunch of random apps like that. Did they know that, you know, Uber was gonna come out of that, DoorDash was gonna come out of that, TikTok was gonna come out of that?
02:56Probably not. So I think that what's really cool about the phase that we are in right now with Claude Code and also Claude Cowork is we're all figuring it out at the same time.
03:08And none of us, including the creator, have all the answers, but I agree with you. Like, what's it's it's a good time to be sharing some you know, sharing what's working, what's not working.
03:20And, you know, if you're open to it, let's let's screen share and and get our hands dirty. Yeah.
03:25Yeah. Let's do it. So what are we looking at right now, Boris?
03:29Yeah. So this is the this is the cloud desktop app. So you just download it.
03:34Cowork is only available for Mac OS, you know, Windows coming soon. There's a few different tabs in the desktop app.
03:41So there's the chat, that's the default. There's co work, that's the new one, and there's code, and that's just quad code. Co work under the hood, it's actually just quad code.
03:51And so, you know, the agent that makes quad code awesome, we call it the quad agent. It's also available as the quad agent SDK, you can use it programmatically. You can you know, all all sorts of companies build all sorts of cool things on top of it.
04:03We actually use that same exact SDK directly in Corework. So it's like, you know, it's pretty cool. It's just kinda one way across everything.
04:11We have the best agent, have the best agentic model, might as well use it. And so what I'll do is just to kinda show how how to use this thing. And, you know, like, when I think about agentic AI, this word agentic has sort of lost all meaning because it it's just, like, used so much.
04:30So I I feel like probably a lot of listeners have heard the word agent, but they don't actually know what it means or, you know, they think it's, like, some, like, cool AI or or something. But it it actually has, a very specific meaning in the AI world, which I think has kind of been lost because a lot of the products that people have released in cold agentic in the past are not actually agentic.
04:48And so, like, when you think about the AI products that everyone's used, you know, like, obviously, Cloud Code, chat based apps where you just chat with app kind of back and forth, send some messages. The biggest difference with agents is it can take action.
05:02And it's not just text and it's not just like web searching, but it can actually use tools on your computer. It can it can interact with the world. And so for Anthropic from the very beginning, since before our models were good, like before like, you know, like Cloud three or whatever, this is a thing that we wanted to get really good at because we felt that it's very important.
05:25And so from the start, wanted our models to be really great at coding and then really great at tool use and then really great at computer use. So it's it's kinda cool like, you know, the last year seeing how people have been hacking quad code, it's pretty obvious this is kind of the place that we should go.
05:41And so for people that have used quad code, you know, none of this will be, you know, like too surprising. These are actually things they can do. So really what we're trying to do is make this something that everyone can use in a way that's safe.
05:51And so what I'm gonna do is you know, here on my desktop, I have a I have this like receipts folder. I have a few receipts in it.
05:59So I'm just gonna give co work access to my desktop. And you you have to pick like which specific folders I can see.
06:07By default, they can't see anything. So you have to kind of opt in to see to let it, uh, access specific folders. And so I'm gonna say, you um, know, I can say, I have a receipts folder.
06:21Can you rename the files to match the dates on the receipts?
06:29So so I think one interesting thing is when you're using CoWork, it's really like, it's operating with your files.
06:37Like, that's a that's a big sort of mindset shift that people should have. It's almost like your operating system. Is that right?
06:44Yeah. Exactly. It's like it it has your files, so the ones that you give it access to.
06:48But actually, the even cooler thing is that it can use all sorts of tools. And so files actually is like you know, that's like that's useful, but it's not it's not like that cool. What's actually much more interesting is it can generate files for you.
07:01So, you know, it can make like presentations and things like this. It can interact with any tool over MCP, and it has built in support for Chrome based browsers.
07:10So it can actually control your browser to do to do stuff. And so I'll I'll kinda show that a little bit. So this is kind of the first step.
07:18When you're first getting started with CoWork, the thing I recommend is, you know, do exactly what I just did. You know, dismount the folder, give CoWork access to it, and just, like, play around with it.
07:27It's it's super useful for cleaning up files, organizing things like that. And so here, it it found these four receipts that I have. It's asking me if one one of the receipts, I guess, is missing a date.
07:40So should I just rename the others? So I'll just say, know, for one, it's we'll say it's, like, up to you.
07:50And then for two, I'll say, don't rename it.
07:55And we actually call this reverse solicitation in the AI world.
08:00So so what this means is when the model is unsure about something, it's gonna ask you for clarification. And we we've sort of taught the model to be pretty good at this. So instead of assuming it's unsure about something, it's just gonna ask you.
08:13And so yeah. In this case, so it renamed the receipts.
08:17So I'm just gonna open this up to double check. Yeah. Cool.
08:19And so the receipts were named a little bit better organized. And so maybe what I can try next is let's, put this in a spreadsheet.
08:30So it takes control of your computer, basically, in that sense, right, if you allow it to take control?
08:37Yeah. Yeah. That's right.
08:38And so we put so much work into kind of safety and making sure that as this happens, you accidentally shoot yourself in the foot and delete files or whatever.
08:48So there's just a huge amount of work that went into this. And it kind of starts at the model side where, you know, for Anthropic, from the very beginning, we were the AI safety lab and not the reason that we exist.
09:00And so there's a lot of work into, like, alignment and mechanistic interpretability and kinda all these ideas to to make sure that the model does what you want in a way that's safe, kind of at the model layer. And this literally means, like, studying the neurons, kind of the same way that you would study neurons in, you know, like, in a in a human.
09:18And so we you know, you can, like, identify structures, and you can kinda study in a very scientific way as a black box also to make sure that it's safe.
09:25So this is called alignment. And then we do a whole bunch of other stuff. So there's actually a whole virtual machine running under the hood, and this is just to make sure that any actions taken are safe and don't affect your broader system.
09:37And then as of last week, there's also deletion protection. So if you accidentally delete something, then you're gonna get prompted first.
09:45So the model can kinda make sure that that's actually a thing that you wanna do. Obviously, also as we start interacting with the Internet, something like prompt injection is quite scary. And so we built in a lot of protections against that.
09:57Obviously, you know, it's it's not perfect, it's something we're iterating on. But this is also part of the reason that we released this pretty early is we wanna see how people use it. And a big part of making models safe isn't just studying them on a lab, but studying them in the wild to see how it's useful.
10:13And so this was like pretty cool. Right? So we we have this folder.
10:17First it kinda renamed all all my receipts. I asked it to make a sheet. It just made the spreadsheet, so it's already here.
10:24Maybe I don't want a spreadsheet. Maybe I I want like a Google sheet. So actually
10:28That's so interesting. Right? Because, like, I think where people stumble with tools like Cowork and Cloud Code is because you can do anything, making it a of course, that makes sense, making it a Google Sheet.
10:44But, like, that's where it's so interesting that you can you really can treat it like a teammate in that sense. Right?
10:50Like, go and do this thing. The world is your oyster. Woah.
10:53What is happening right now?
10:55Yeah. So it's like it's opening the browser. And so here it's gonna ask you for permission.
11:00Just for this demo, gonna say always allow for the site. But you can just say allow once or deny.
11:05And so Quad's taken the wheel. It's making a spreadsheet for me. This is gonna sort of take a while.
11:11And so this is one of the things that we're iterating on is making this kind of computer use a lot faster. I I remember man, this is when I first joined Anthropic and this was maybe this is like Sonnet 3.6 or I forgot what the what we called it, like Sonnet 3.5 new.
11:29And that was the first model that I think we really started to crack computers. And I remember I was sitting with my team and there was literally a researcher that ran into the room in the movies and was like, oh my god, Quad knows how to use computers.
11:41And we just had to order a pizza and it picked a pineapple pizza and then ordered it and it asked us for a credit card.
11:48And then we we got that delivered to the office. But it was tedious. It took like an hour or something for it to click around.
11:55And since then, we've been improving the model's ability to use a computer. And so you can kinda see that here. So here the, you know, Quad is typing and it's interacting with the spreadsheet.
12:05And it can see what's on the screen. It can interact with it. Anything that's in your browser, it it can just use.
12:11And could it use, like, email know, let's say we wanted to send this to somebody, you know, finance team.
12:20Like, can you maybe we'll get into it later, but, you know, with things like connections and MCPs, is that possible?
12:28Yeah. Yeah. Totally.
12:29Okay. So let's, like, let's make this, like, kind of a nice spreadsheet because we don't wanna send, like, a badly formatted spreadsheet to our coworker. But let's make it nice, and then we can ask Claude to send it.
12:39And so here it's kinda cool. It's, like, you know, it's it's still early days and so there's some, like, formatting mistakes. It didn't paste it exactly correctly, but it noticed it.
12:47So notice that, you know, it's not actually split correctly, and so now it's trying to format it. So kinda notice this.
12:55And, yeah, Greg, like, to your point, it's funny. These, like, these use cases and the way that people are gonna use these tools, like I said at the start, it's gonna be so surprising. I'm just so excited to see how people use it.
13:07And sort of the crazy thing about them is they're so general purpose. It's like it's sort of like a computer itself or like the Internet or something. Like like you said, like, when when you first, like, had an iPhone, like, you would never have predicted that, you know, there would be an Uber app at some point.
13:20But, you know, that's that's what happened. And I sort of feel like we're at the beginning of that, but for for agents.
13:27Yeah. I think it it it's it sounds so obvious in hindsight.
13:33Like, of course, you have a phone with GPS. Of course. Right?
13:37You're gonna you're gonna have things like Uber and DoorDash and stuff like that. But I feel like, yes, you're right.
13:44I think, like, what I'm trying to figure out in my mind like, I'm watching this and I'm like, this is really cool, but I'm thinking about, like, okay. You know, how do I audit, like, my entire company? Like, what are all the tasks that the company is doing?
13:57How are they interacting with files? And how are they using the Internet? And how are they sending things?
14:04And then what are opportunities I can use to, like, make my team and also my life more productive. That's what's going through my head right now. Yeah.
14:12Yeah. Totally. That's a great way to think about it.
14:15And by the way, something that's kinda cool is, you know, as this is running, I'm just gonna make a new task and do something else. So, you know, what are some cool episodes of startup?
14:38Cool. So while while this one's running, we can, like, we can let this one go too. And I I often have a bunch of tasks running in parallel too.
14:46But, yeah, that's exactly it. It's like you should just think about, like, what what's all, like, the tedious stuff that you do every day? And you can just throw all of this stuff to co work.
14:54And I I feel like this happened for coding over the last year because, of course, engineers and programmers are the earliest adopters. So when the tooling first became able to do that kind of work, engineers adopted it first.
15:07But now now this is coming now this is possible for everyone else to use too.
15:13So that's super exciting. And when we talk to engineers about the way they like to use quad code, they run a bunch of quads in parallel. They use it to automate the tedious stuff.
15:21And I don't know, man. Like, for me, this is just the most fun I've ever had as an engineer because I get to do the stuff that I enjoy and I just feel so productive because quad does all the, you know, the stuff that I didn't wanna do. Okay.
15:35So now now we have this, like, we have the spreadsheet. This looks alright. It's not too bad.
15:40The data looks correct. You know, one thing that's missing though is well, we we can add maybe like the totals, but I think for the sake of the demo, what I'm gonna do is I'm gonna see if it can email it.
15:56So just reading for audio listeners. Can you open Gmail and send the sheet to Amy? And wow.
16:05Yeah. So I'm gonna say always continue.
16:09So how is it gonna know it's Amy. Right?
16:12It's gonna pull up the contacts within Gmail?
16:16Yeah. Yeah. Exactly.
16:19Yeah. That that's exactly it. It's gonna interact with it the same way that, you know, anyone else would interact with a computer.
16:26And it's the same thing. It can it can, like, click stuff. It can read the screen.
16:31Okay. So compose window open. Let me type Amy.
16:34So you have, like, multiple when you're when you're, like, locked in, so to speak, you're kinda you have multiple of these windows open at or multiple virtual machines open at at the same time.
16:45Yeah. Yeah. I usually have a bunch of these, you know, like five or five or 10 or whatever.
16:50And so we can see the models, you know, trying to interact with this. And so it filled out Amy.
16:57And you know, by the way, just so because this is, you know, going on YouTube, you know you're gonna have some haters in the comments, so we're just gonna address them right now. Someone in the comments is gonna be like, but that's so slow. I can do it faster myself or something.
17:11What is your reaction?
17:12Of course you can. Yeah. Could definitely do this much faster.
17:16And it's actually the same thing for a lot of quad code. Right? Like, especially at the beginning, I I could do it much faster.
17:22But but I think sort of two things happen. One is that the model just gets better at doing it quickly, and this is just something you would I I would expect over time.
17:31But then the second thing is because you can do multiple things in parallel, it's actually a big time saver. And so usually my workflow is I'll kick off like a few different tasks in parallel, and I'll just kinda like go back and forth between them and kinda tend to my quads, make sure they're in a good place, see if they have any questions or anything like that.
17:49So it's really a different kind of workflow. Like, I feel like now is the age of like kinda multi multi clotting, of of parallelism, of like not going super deep on stuff, but kinda being more of a generalist and more tending tending to your clots.
18:04And so this was this other chat that we kicked off earlier. And in this case, it sort of did some research, so that's pretty cool.
18:13So it didn't just go to, you know, like startup ideas podcast website, but it searched the Internet, checked a bunch of different places, checked like someone's notes about it. Yeah. Ten ten rules for for quad code.
18:24Yeah. I'll find that interesting too. And let's see.
18:30So it seems like this is the email is probably all drafted.
18:36The draft has been saved. Yep. Please send.
18:40And I also use this for, like, Slack too. Like, a pretty common use case I've been doing maybe a couple times a week is I have a spreadsheet where we track all the team's work for the month or for the week or whatever. And instead of having to bug everyone on the team to fill out the status, what I do is I ask co work to look at the spreadsheet.
18:59Any column that's not filled out, just message the engineer on Slack. And it does that really well. So, you know, I I just ask it to do that, and then I go get a coffee, and I I don't have to do that anymore.
19:10Do you think that co work is gonna be the gateway drug to Claude Code? Because for a lot of beginners, nontechnical people, Claude Code feels a bit overwhelming. But when I'm when I'm watching you do this, this feels, you know, approachable.
19:25Yeah. I I mean, I think it's, like, probably two kinds of users. Well, and and it's also it's funny.
19:31QuadCode is built originally in a terminal. And nowadays it's available in the Quad mobile app.
19:38It's available on the website, in the IDE, on Slack, on GitHub. So nowadays it's available everywhere, but originally it was built in a terminal.
19:46And I never thought that most engineers would want to use a terminal. Because, you know, it's scary. It's like this thing that's sort of hidden away on the computer.
19:53Like, you know, only the most hardcore engineers want to actually work in a terminal day in and day out. And so it's surprising to me that most engineers wanted to use it. What's even more surprising is a lot of non engineers started to use it, and that was the craziest thing.
20:06And, you know, looking at, like, the sales team at Anthropic, like the GTM team, half of them use quad code every week. And looking at other nontechnical people at Anthropic, like designers, product managers, data scientists, pretty much all of them use quad code every day.
20:20So that that's been like pretty surprising. But my hypothesis is what they would prefer is actually something a little bit more like this, where you don't have to deal with a terminal and all that kind of thing.
20:31You have a you have an nice UI because you don't need access to bash and all that kind of stuff because the the model will just do it for you.
20:38Yeah. I think what I would love with something like this is, you know, automating some of these processes.
20:46It's like instead of doing one offs, it's like, anytime this happens, I want you to do this. That that's my dream.
20:54Yeah. It's a cool idea. I'm I'm excited to see, like, how it's tied into skills also.
20:58I think it's something we're we're thinking about. We haven't quite figured out yet. Yeah.
21:03Because like a skill, what what is it? It's essentially kind of a repeatable way to do something. And so you saw that a little bit earlier up in this conversation where when Quad was generating the spreadsheet.
21:15I don't know if you saw it, but it actually loaded the skill. And so we prepackaged the skill for Excel and that's the way that Quad knows how to do it. And so if you have some weird file format, don't know, if you work with AutoCAD or if you work with Salesforce, whatever it is, the tool that you use, just make a skill and then Claude can can do it for you.
21:35That's interesting. So if you have more skills, co work, and you're asking it, you know, to do things related to those skills, co work is more likely to use those skills.
21:47So you might be able to get a better outcome out of co work if you, you know, do some upfront work to actually create some skills.
21:57I think so. Yeah. Wow.
22:01Yeah. I think I think that's exactly right.
22:03Yeah. Okay. Well, that's good to know.
22:05And and, you know, what else do we need to know about co work? Because I know there's, you know, there's, you know, extensions, there's there's skills.
22:14Like, walk us through some of is there anything else that, you know, we need to know in order to get the most out of of the product?
22:23You know, we we try to build core work just to be pretty simple. When we think about the different audiences we serve, for an audience like engineers, they don't actually like simple.
22:32Simple is is good by default, but engineers love to hack their tools. They love to customize their tools. So that's actually really important.
22:39So quad code is just like the most customizable coding agent. It has just like so many extension points. There's like there's skills.
22:46There's custom agents. There there's hooks. There's an insane amount of configuration and settings that you can set.
22:52You know, there there's a a very sophisticated permission system. So there's all sorts of ways to customize it. And this is because we know every engineer works different, and everyone has different preferences.
23:03Everyone uses a different tech stack. Everyone uses a different editor. Everyone use a different OS.
23:08We So wanted to make sure it's really customizable. For some of my co work, I think we're kinda starting in the opposite direction and just keep it really simple at the start.
23:16And so if you use co work, I would probably not customize too much. Like, if you have co work installed and you install the Chrome extension, that's pretty much all you need and it'll do everything else for you. I think over time as you find yourself using maybe software that Core Work is not great at, like, that's the point to think about writing skills.
23:35But that wouldn't be my starting point. Just start simple. See how it works, see what's useful.
23:40And I think also the other thing, it's just it's so early. This feels to me a lot like Quad Code a year ago where we released it before it was early.
23:50It was not very good yet. It was super buggy. It kinda barely worked.
23:55The model was not very good at coding a year ago. And it kinda feels like that to me, but I think Corework is actually a little bit better than QuadCode was when we first released it because it's actually useful.
24:06And it's already a thing that I use every day. And, yeah, I think in the first week, we've we've seen so much more growth.
24:13It's it's been like a multiple of the growth that we saw with QuadCode the first week. So that that's very exciting. It it seems like this is something that people are already finding useful.
24:22You know, we're recording this, what is it, January 2026. If we if we come back here in twelve months, it's January 2027, how do you think people are gonna be using Cowork?
24:35You know? I know you don't have a crystal ball exactly, but what sort of use cases, and and what is what does the product look like? Like, describe the world in January 2027 with with co work.
24:48Oh, jeez, Greg. I I plan in, a one week timeline. The model is just changing so fast.
24:54It's it's so hard. And I just feel like, you know, like, the model is it's it's advancing exponentially and just, my puny human meat brain, like, can't grapple with exponential.
25:03It's like we think in linear. And so I think this kind of exponential is just very, very difficult to plan around.
25:10Okay. But if if I had to speculate, a year ago, I made and I think Dario also made this prediction that by the end of the year, people wouldn't be writing code anymore.
25:21And I think that was like sometime mid last year or something.
25:27I code every day. I ship 200, 300 PRs every month or something like that. And in the last two months, Quad Code has written a 100% of my code.
25:37I haven't written a single line by hand. And this is something like I also predicted this way back middle of last year.
25:45It was sort of not intuitive because if you just think about kind of the experience at the time and you trace it linearly, there's just no way that the model would be at that point.
25:56So you really have to kind of believe an exponential and just literally like plot it out and kind of follow the way that the line the line goes. That's the only way that you would have predicted this.
26:05And it was just it was absolutely right. And so I think for coding, this is a work this is something that we're gonna start to see in more and more places for more and more kinds of code that the model is able to just do all of it. And when we think about co work, I think it's somewhat similar.
26:19I think it's a little bit earlier. And I think what we're gonna see is for all this kind of tedious tasks like connecting app a and b or kind of shuffling data back and forth or whatever, the model is just gonna be able to do it.
26:32And it's gonna get increasingly good at it. This is I think in some ways it's a little scary. And in some ways I think it's really exciting because you don't have to spend your time on this toil some work anymore.
26:44You can just focus on the work that you enjoy. And also everyone, I think, just becomes much much more productive because you have an army of plots that that can do this.
26:54Okay. I like that feature. I like that feature a lot.
26:57It's a hard question. I it's a hard question when you it's sort of like the genie's out of the bottle, and it's hard to predict, like, where the genie's gonna go.
27:06So I believe that co coworker is the gateway drug to clot code.
27:11I think that people are going to start using it, and they're gonna develop, like, vertical use cases, like, for whatever it is their business is.
27:22Like, you mentioned AutoCAD. Like, you probably weren't thinking about AutoCAD when you were dev helping develop Cobalt Cowork.
27:29So I think there's gonna be, like, these verticals, um, and I think I think that Cowork is gonna be similar to, like, vertical job boards where there's very specific roles that you can hire to people.
27:44And I do think that there's gonna be this combination between skills and these, like, digital teammates that you're gonna, quote unquote, hire.
27:53Yeah. That's that's super interesting. Maybe we should make a bet and just see where it pans out in the year.
27:57But, you know, I I have no idea. But it's it's definitely interesting to speculate because I I I just feel like the way this technology goes is so different than past technology waves.
28:08It's sort of similar to the Internet. It's sort of similar to computers, you know, like maybe like telephones or something before that, but the speed is just so much faster.
28:16And because it's kind of piggybacking on all of this, like, the Internet could not exist without telephones telephones and kind of like phone lines being everywhere. You couldn't have dial up without it.
28:25You know, mobile phones couldn't exist without the Internet existing. So it's sort of like every layer of the stack, it just gets more and more powerful and it spreads more quickly. And so it's like on the back of all of this that AI can exist.
28:38And, yeah, it's it's just gonna be very exciting to see to see where it goes.
28:42I wanna shift gears just for the last ten minutes because we mentioned Claude Code. You had a you had a post that went absolutely viral.
28:52I think you know what post I'm gonna talk about. 99,000 bookmarks, which is crazy.
28:58And I'm gonna share my screen. I wanna go over it a little bit. This is me learning to use Twitter.
29:05I think I made an account, like, a decade ago, and I haven't really used it. So It's funny. Yeah.
29:10It's been fun to learn. You have, like, just a few tweets, but, like, this, you know, this one, you figured it out. You say, I'm Boris.
29:18I created Cloud Code. Lots of people have asked for how I use Cloud Code, so I want to show off my setup a bit.
29:24My setup might be surprisingly vanilla. Cloud Code works great out of the box, so I personally don't customize it much. There's not no one correct there's no one correct way to use Cloud Code.
29:35We intentionally built it in a way that you can use it, customize it, and hack it however you'd like. Each person on the Cloud Code team use it very differently, so here it goes. Could you just talk about some of the just expand on some of these and and what you were saying, you know, what you were saying on here?
29:51Because I thought it was really interesting.
29:53Yeah. I'd love to. So this first one, it's very similar to what I was showing in co work where my job now isn't to go super deep on one task.
30:02It's to do a bunch of tasks in parallel. And so, you know, in the when I'm working on quad code, I I usually work in a terminal or on the mobile app. These are kind of the two services I use the most.
30:13But like I said, everyone on the team is different. Every user prefers something different, so we build all of these. And so usually what I do is I'll sort a task in kinda one tab.
30:21And once Claude is thinking about it and starting to work on a plan, I'll move on to the second tab. And I'll ask it to make a plan for the second thing. Then I'll move on to the third one, ask it to make a plan.
30:31And then finally, like, when, you know, I've run out of immediate tasks to do, I'll go back to the first tab. I'll see if the plan looks good.
30:38I might go back and forth a little bit. And then once the plan looks good, I usually go into just auto accept edits right away. Because I think with Opus 4.5, once the plan is good, the model can just execute it pretty much perfectly.
30:51This is definitely not the case with previous models. And so I think there was a lot of excitement about OPUS port 4.5 in you know, over the last couple months, and I I I think this is kind of one of the big reasons.
31:02It's just gotten very good at coding, but also excellent at planning. So once the plan is good, the code is good. And so, yeah, my my work now is just jumping between tabs, kinda tending to the quads, make sure they're unblocked, answering their questions.
31:14With co work, I think I think it's actually quite similar now.
31:17I like that. I'm gonna I'm gonna quote you on that. Once the plan is good, the code is good.
31:22Because that's so true. Right? Because if you nail the plan, the code should, you know the agents should do the work.
31:29Yeah. Yeah. Exactly.
31:31I think, like, sometime last year, there was all this buzz about spec driven development. Yeah. And it just feels like it feels a little too QT and a little too rigid to me.
31:40But I think this is sort of a form of spec driven development. It's like there there is some kind of spec. I think it's just like a plan.
31:46That's all it has to be. It's just a text file. It doesn't have to be in a particular format.
31:50Once you have that, you're good. Number two, you say, I also run five to 10 clods
31:56in in parallel with my local clods. As I code in my terminal, I will often hand off local sessions to web using and or manually kick off sessions in Chrome, and sometimes I will teleport back and forth. I also start a few sessions from my phone, obviously, from cloud iOS app, Picatu, every morning and throughout the day and check-in on them later.
32:16What do you mean by that? This morning, I kicked off, I think, like, three claws as soon as I woke up.
32:22I just had, like, some thought in the morning about, well, maybe I should, like, build this thing or fix this bug or whatever. I was, like, checking Twitter and someone had a bug report. So I just opened my phone and, you know, in the Quad app, you on on the left side, you click the little menu and there's a code tab.
32:36So you can just, like, access Quad code there. That's what I use for a lot of my code. And it's funny.
32:41I I never would have guessed that this is the way that I code. If you asked me a year ago, I would never have predicted that the way I code now is, like, probably half of it is just on my phone and that it sort of just works. And then web is kind of the other part.
32:54So once I've run out of tabs because it's just like, you know, it's kinda a pain to manage a bunch of different git checkouts because in each tab in my terminal, I actually have a totally separate git checkout. And I don't really use WorkTrees or anything like that. I just keep it pretty simple.
33:09And in web, if I if I just ran out of terminal tabs, I'll just start, like, overflowing to web and and starting tasks there.
33:15Beautiful. So, yeah, I also I so recommending use iOS, use web, open multiple, like, that's what that's you'll get the most out of it that way.
33:27Yeah. Absolutely. Yeah.
33:28And we have a Android app now too. So, yeah, just like use the mobile app and see how far that can take you. If it's not giving you good results, also make sure that you tune the environment setup and make sure that you invest in your quad dot m d.
33:40That's super duper important. And I talk about that in number four also.
33:45Mhmm. Number three, say, I use Opus 4.5 with Thinking for Everything. It's the best coding model I've ever used.
33:52And even though it's bigger and slower than Sonnet, since you have to steer it less and it's better at tool use.
33:58It's almost always faster than using a smaller model in the end.
34:03And cheaper. It's sort of counterintuitive and I had to explain this to people a few times. But because the model is smarter, it actually uses less tokens in the end.
34:14And it uses so many less tokens, it's often cheaper than using a smaller, less intelligent model even though the per token cost for that model is lower.
34:25A little counterintuitive, but, yeah, just use the smartest thing if you can. Well, we appreciate that. You know?
34:30We appreciate that.
34:33Number four, our team shares a single Claude MD for the Claude code repo. We check it into git. The whole team contributes multiple times a week.
34:41Anytime we see Claude do something incorrectly, we add it to the MD so Claude knows not to do it next time. Other teams maintain their own Claude MDs. It is each team's job to keep theirs up to date.
34:54And the QuadMD is just a text file, so there's no special format. People ask this all the time. Is there some special format that it has to be in it?
35:02No. It's just like it's just a text file, so you can put whatever you want in it. This is a screenshot of our QuadMD.
35:07This is literally it. So just really simple. And, you know, you you could format this kinda however you want.
35:13Love it. And then finally, I think yeah. Number five.
35:17Oh, no. You have a you got a few more. So number five, during code review, I'll often tag at dot claud on my coworker's PRs to add something to the CloudMD as part of the PR.
35:29We use the claud code GitHub action for this. It's our version of Dan Chipper's compounding engineering.
35:36What what do you mean? What is compound engineering?
35:39Yeah. Oh my god. So there's actually like two bugs in this tweet.
35:43I think I think Dan actually calls it compound engineering, not compounding engineering. And then the at dot quad, this is me learning how to use x. But it it's actually at quad.
35:52There's there's no dot. I just didn't I think there's like an actual user or whose name is Claude. I I didn't wanna tag them.
35:58But yeah. So so what you do is in Claude code, you run this slash command slash install GitHub action. And what this does is it installs the Claude app in your GitHub repo.
36:09And what that lets you do is you can then at mention Claude whenever you want and just have it make changes. And it can just work on PR. So it'll push back to your branch and it'll push the changes right back.
36:19You can also tag it on issues. You can tag it kind of wherever. I do this multiple times a day.
36:25It's it's really, really useful. I think one of the most common use cases is just like little fixes. I think the other one is updating the QuadMD to keep the knowledge base up to date.
36:36And, you know, you should never have to comment about something twice. Back when I was at Meta, something that I did is this is like in a in a previous life, a previous job.
36:46Something that I did is every code review that I did, I would keep a spreadsheet of all the issues that came up. And whenever the same kind of issue came up again, I would just like tally it up in the spreadsheet.
36:56And whenever something hit, I think, like five or 10 tallys or something, I would write a lint rule. And what that is is it's a way to automate that part of the code review so I don't have to comment about it again.
37:07And that was back in the days before LEMS and before the model was any good at coding. And so this is the equivalent nowadays. You just tag Claude, and you have it update your QuadMD, which is your team's knowledge base.
37:18So really simple. And what this means is you don't have to point anything out twice. I'm curious to see what this looks like for co work.
37:25I don't think we figured that out yet. But definitely for quad code, quad m d is kind of the this is the one the one thing that you should all be updating all the time.
37:34Love it. Number six, most sessions start in plan mode.
37:39If my goal is to write a pull request, I will use plan mode and go back and forth with Claude until I like its plan. From there, I switch into auto accept edit edits mode, and Cloud can usually one shot it.
37:51A good plan is really important.
37:53Yeah. Like I said, planning is just the most underused feature in Cloud Code. Actually, lot of people use it, but I I would say it's still underused.
38:00I use it for almost all my sessions.
38:03Yeah. It's it's a no brainer.
38:08So if you aren't already doing this, please do. Do we we've got you know, we're not gonna have time to do all of them.
38:17What are the I don't think because you have a hard stop. Do you wanna pick one or two to end with?
38:24Yeah. Let's do this would maybe number 13.
38:29Yeah. Number 13, I think this is probably in addition to using Opus four when people ask about how to get better performance out of quad code, there's three things that I recommend almost every time.
38:39Number one is use Opus 4.5 with thinking always. Don't try to use a different model because Opus will just give you better results and more efficiency overall. The second thing is make sure you have a good QuadMD.
38:50And then the third thing is this this tip number 13, which is give Quad a way to verify its output. And so we we just kinda saw with CoWork how good Quad is at using the Chrome extension in order to write email and in order to work with Sheets.
39:06And it's exactly the same thing. If I'm building an app, I always use the Chrome extension to have Quad test its own work. And if Quad can verify its own output, the result is gonna be way way better.
39:18And it's sort of like imagine that you're a painter and you make paintings and they have to be pretty good.
39:25They have to be maybe even photorealistic or something or just some kind of very detailed style. And you .
39:34You're just not gonna be that good. It's not gonna come out that great. Or it's the same thing for an engineer.
39:38Like, if you have to write code but you can never run the code or you can never see the output and you can never see the website. It's just not gonna be good. And so it's the same thing with Quad.
39:47As the model gets more intelligent, that first shot is gonna get better and better. But really, you wanna give it a way to verify the output, and it'll be much better.
39:56So this is like running tests if you're an engineer, starting a server also if you're an engineer, or, you know, seeing seeing the output in a simulator or in a browser.
40:06Amazing. And the ultimate tip, I guess, is just get your hands dirty. Right?
40:11That's it. Yeah. I mean, there's no right there's no one right way to use the stuff.
40:16Like, you know, see see what's useful, see what's not, find your own workflows. It's sort of a I I heard someone describing QuadCode as like a find your own path book.
40:25You know, like one of those books where you have to like, do you do do you go to the dungeon or do you explore the forest? Like this kind of thing. Or or like a like an old RPG video game.
40:34It's just kinda like that. It's like it's very free form. There's no one right way to do it.
40:38So just see what works for you. Boris, I appreciate you taking the time, the generosity to spill the sauce and just to to share with us
40:47how we can get the most out of both Cowork and Claude Code. Everyone's thinking about it, everyone's trying to especially on this podcast, you know, the Start Up Ideas podcast, you have millions of people who are trying to figure out how I can be more productive and how I can build businesses around some of these things.
41:02So thank you. Thank you for the time. I'll include links on where to follow Boris.
41:08He doesn't tweet a lot, but when he tweets, it's it's worth paying attention to. And I just have I wanna leave with one thing just because my audience would kill me if I didn't ask this. So as you know, Boris, I'm from French Canada.
41:21And we speak French, obviously, in French Canada. And I call Claude Claude.
41:28Am I the only one?
41:32Oh, man. You're probably not the only one. A of the a lot of the world speaks French.
41:38Which one sounds better? What what feels I mean, me,
41:42Claude sounds better. It's like Jean Claude Van Damme, even though people would say Jean Claude Van Damme. It's like,
41:49that's my take. That's my hot take. To me, it sounds better.
41:53You know, Greg, I'm I'm gonna today around the office, I'm just gonna call it Claude, we'll we'll see what people say. I'll I'll report back. Okay.
41:59Report back. Alright. Sounds good.
42:01Boris, thanks again for coming on,
42:03and I'll see you next time. Yeah. Thanks so much, Craig.
The Hook
The bait, then the rug-pull.
Boris Cherny built Claude Code before most people knew what an agent was — and now he's back with Cowork, a GUI wrapper that makes the same agentic engine accessible to anyone with a Mac. Greg Isenberg got the private lesson. You're about to watch it.
Named ideas worth stealing.
32:50model
Compounding Engineering (CLAUDE.md as team memory)
Check CLAUDE.md into git. Every time Claude does something wrong in a session, add a rule so it never repeats. Tag @claude on PRs via GitHub Action to auto-update the file. The knowledge base compounds with every session and every code review.
Steal forAny project with recurring Claude sessions. One shared CLAUDE.md that the whole stack reads.
34:50model
Plan-first then auto-accept execute
Start every session in plan mode. Go back and forth until the plan is solid. Then switch to auto-accept edits. With Opus 4.5, a good plan almost always produces good code. Thesis: Once the plan is good, the code is good.
Steal forEvery non-trivial Claude Code session. Stop diving straight into implementation.
39:00concept
Verification loops
Give Claude a way to verify its own output: Chrome extension, bash command, test suite, simulator. The model iterates on failures. Without this, you are asking a painter to work blindfolded.
Steal forAny automated build pipeline. JoeFlow tauri:dev as verification. MCN test suites.
15:11model
Multi-Clauding (parallel session management)
Run 5-10 Claude sessions simultaneously across terminal tabs, web, and iOS. Each tab has a separate git checkout. Tend to them: check in, answer questions, unblock. Parallelism beats raw speed.
Steal forMorning stack launch routine. Start 3-5 sessions before coffee, tend between them.
How they asked for the click.
VERBAL ASK
42:07next-video
“Boris, thanks again for coming on. I'll include links on where to follow Boris.”
Soft verbal close with social links promised in description. No explicit subscribe ask. Ends on a warm pronunciation joke.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
