The argument in one line.
Claude Cowork's power is locked behind unintuitive defaults, so storing all instructions in CLAUDE.md files, using projects only as folders, loading skills on demand, and scheduling tasks during off-peak hours transforms it from a frustrating tool into a genuinely useful agent.
Read if. Skip if.
READ IF YOU ARE…
- A non-coder using Claude Cowork who wants to optimize file organization, instruction management, and scheduling efficiency across multiple concurrent projects.
- Someone considering Claude Cowork adoption who needs practical patterns for structuring folders, instructions, and agent workflows before investing time in the platform.
- A power user of Claude Cowork who has hit friction points with project bloat, instruction fragmentation, or scheduling costs and seeks specific tactical fixes.
SKIP IF…
- You're still deciding whether Claude Cowork is right for your use case—this assumes you're already committed to the platform and want advanced optimization.
- You work primarily in code-first environments or use Claude through API integrations rather than the desktop app interface.
The full version, fast.
Claude Cowork rewards a clean folder-based setup, not bloated project instructions. Treat projects purely as chat folders so your workflow stays portable to future agents, and push all guidance into CLAUDE.md files inside each working directory, with subfolder CLAUDE.md files for project-specific context. Select multiple folders per task to scope information precisely, set a default working folder, and rely on toggled voice dictation to brain-dump messy thinking the agent can structure. Schedule heavy tasks for off-peak hours to preserve usage limits, offload slide decks and sites to Claude Design's separate quota, and use Obsidian as a free cross-platform viewer. Load skills on demand instead of stuffing instructions, add connectors to centralize tools, then prune unused skills and connectors monthly to protect performance.
Members feature
Chat with this breakdown.
Modern Creator members can chat with any breakdown — ask for the hook, quote a framework, find the exact transcript moment. Unlocks at T2: refer 3 friends + add your own API key.
Create a free account →Where the time goes.
00:00 – 00:19
01 · Cold open
Establishes the gap between most users and full Cowork potential after recent updates.
00:19 – 02:28
02 · Tip 1: Projects = folders only
Controversial take: never put instructions in project settings. Keep instructions in CLAUDE.md files so they travel when you switch tools.
02:28 – 03:31
03 · Tip 2: CLAUDE.md in every folder
Store all agent instructions in a claude.md file at the folder level. Each subfolder gets its own scoped instructions file.
03:31 – 04:31
04 · Tips 3 & 4: Multiple folders + default folder
Select exactly the folders each task needs. Set a default parent folder so new tasks start there automatically.
04:31 – 06:06
05 · Tip 5: Dispatch limitations
Dispatch requires computer on + Cowork open + linked. One continuous conversation, no project switching.
06:06 – 07:12
06 · Tip 6: Voice dictation toggle mode
Hidden drop-down behind the mic icon switches from hold-to-record to toggle mode. Brain-dump ideas and let Claude structure them.
07:12 – 08:10
07 · Tip 7: Off-peak scheduling
Peak hours are 5AM-11AM Pacific. Schedule automations for off-peak or when limits will reset.
08:10 – 09:51
08 · Tip 8: Skills over long instructions
Skills load on demand; a bloated CLAUDE.md loads every time. Replace long instruction blocks with modular skills.
09:51 – 10:18
09 · Tip 9: Use connectors
Connect the apps you already jump between. Goal: one central Cowork command hub.
10:18 – 10:51
10 · Tip 10: Claude Design has separate limits
Slide decks, websites, motion graphics built in Claude Design do not touch Cowork usage limits.
10:51 – 12:14
11 · Tip 11: Obsidian as file viewer
Free cross-platform app. Point it at your agent folder. Renders markdown, HTML, and nested subfolders.
12:14 – 13:08
12 · Tip 12: Prune skills and connectors monthly
Too many active skills or connectors degrades performance and burns token limits faster.
Lines worth screenshotting.
- Project instructions in Claude Cowork create lock-in — putting all instructions in CLAUDE.md files instead makes the setup portable to any future tool in one find-and-replace operation.
- Use Cowork projects as folder systems for organizing chat history, not as instruction containers — the instruction layer belongs in CLAUDE.md files that travel with the files.
- Multiple folders can be selected as the working directory for a single task, giving Claude access to exactly the files it needs without exposing the entire file system.
- Setting a default folder means every new task starts in the right directory without a manual selection step — the highest-leverage Cowork configuration for daily users.
- Dispatch requires the computer to be on and Cowork to be open, making it impractical as a mobile solution — its value is limited to simple single-context tasks.
- Toggle dictation mode using Command-D so that voice input stays active without holding the button, which is the correct mode for extended brain-dump sessions.
- Scheduling tasks during off-peak hours avoids the cost surcharge that Cowork flags with a visible warning — the same work runs cheaper at 3am than at 2pm.
- Loading skills on demand rather than storing them permanently in instructions keeps the context window clean and avoids bloating every session with irrelevant SOPs.
- Obsidian is a free cross-platform markdown viewer that lets you browse Cowork's file outputs on any device, including phone, without a separate app or subscription.
- Voice dictation into Claude Cowork converts unstructured brain-dumps into structured output — the agent is optimized for extracting signal from rambling, making talking often better than typing.
- A CLAUDE.md file in each subfolder with subfolder-specific instructions creates a hierarchical instruction system that scales as the project grows without touching the root file.
- A cloud-based coding agent that continues running when the laptop is closed is the correct tool for overnight builds — Dispatch is not a substitute for a properly hosted agent.
Twelve Structural Decisions That Determine Whether Claude Cowork Stays Useful as Your Workload Grows
Claude Cowork
Paul Lipsky's 12-tip framework argues that the decisions that matter most for long-term Cowork performance are structural — where instructions live, how skills are organized, when automations run, and which integrations stay active — not which features you use.
02Tip 1: Projects = folders only
- Never put instructions in project settings — they are locked to Cowork and will not travel if you switch tools
- Use projects only to organize chats — the organizational value is real, the instruction-storage value is a lock-in trap
03Tip 2: CLAUDE.md in every folder
- Every folder gets its own CLAUDE.md with scoped instructions — subfolder instructions apply only to tasks that run from that subfolder
- Folder-level instruction scoping prevents the rule bleed that makes large workspaces unreliable
05Tip 5: Dispatch limitations
- Dispatch requires the computer to be on, Cowork to be open, and the account to be linked — all three must be true for scheduled tasks to run
- One continuous conversation with no project switching is the current Dispatch constraint — design automations around this limitation, not against it
06Tip 6: Voice dictation toggle mode
- A hidden dropdown behind the microphone icon switches from hold-to-record to toggle mode — the toggle is more useful for longer brain-dumps
- Toggle mode lets you talk through a complex idea and let Claude structure it afterward — the input is speech, the output is organized text
07Tip 7: Off-peak scheduling
- Peak hours: 5AM to 11AM Pacific — schedule automations outside this window for more reliable execution and less rate-limiting
- Scheduling around the usage cycle (when limits reset) is more reliable than trying to squeeze into peak capacity
08Tip 8: Skills over long instructions
- Skills load only when invoked; a bloated CLAUDE.md loads every session — the difference is context cost per task
- Modular skills are the architecture that keeps long-term Cowork performance stable as your workflow grows
12Tip 12: Prune skills and connectors monthly
- Active skills and connectors that are not being used degrade performance and consume token limits on every session
- Monthly pruning keeps the system at the performance level you built it to — it is maintenance, not optional cleanup
Terms worth knowing.
- Claude Cowork
- Anthropic's agentic desktop application that runs tasks directly against files and folders on a user's computer, acting as an AI assistant that can read, write, and organize local content.
- AI agent
- A software program powered by a large language model that can take actions on its own — running tools, reading files, and completing multi-step tasks — rather than just returning a single chat response.
- Project (in Claude Cowork)
- A container inside Claude Cowork that groups related chats and lets you attach standing instructions that apply to every conversation started from within it.
- CLAUDE.md
- A plain-text Markdown file placed in a folder that an AI coding agent reads automatically when working there, supplying persistent context and instructions for tasks in that directory.
- Codex
- OpenAI's coding-focused agent product that, like Claude Cowork, operates against local files and uses its own equivalent of an instructions file.
- Task (in Claude Cowork)
- Claude Cowork's term for a single chat session or unit of work, listed under recent tasks once started.
- Dispatch
- A Claude Cowork feature that provides mobile access by linking a phone to a running desktop session, allowing remote interaction with the agent while the host computer stays on.
- Voice dictation (push-to-talk vs toggle)
- Two input modes for speaking to the app: push-to-talk requires holding a key while you speak, while toggle starts and stops recording with a single key press each time.
- Scheduled tasks
- Automations that run an AI agent prompt at a specified later time, useful for deferring work until usage limits reset or off-peak pricing applies.
- Off-peak hours
- Time windows outside the provider's busiest period when running AI tasks consumes usage limits more slowly; for Claude, peak runs roughly 5–11 AM Pacific.
- Usage limits
- Caps on how much an AI subscription can be used within a rolling window, typically measured in tokens or messages and resetting on a schedule.
- Skills
- Reusable prompt packages that an AI agent loads on demand only when a relevant task arises, keeping standing instructions short and reducing wasted context.
- Connectors
- Integrations that let an AI agent reach into external apps and services — email, CRMs, calendars — so actions in those tools can be triggered from one central interface.
- Claude Design
- A separate Anthropic product bundled with a Claude subscription for generating slide decks, websites, and motion graphics, with its own usage allowance independent of Cowork.
- Obsidian
- A free cross-platform note and file viewer built around a local folder of Markdown documents, useful for browsing and editing outputs that an AI agent saves to disk.
- Token limits
- The ceiling on how much text — measured in tokens, the chunks a language model processes — can be sent or generated in a given window before throttling or extra charges kick in.
- School community
- A paid membership site hosted on the Skool platform, typically offering courses, a forum, and live sessions to members around a single topic.
Things they pointed at.
Lines you could clip.
00:52
“We shouldn't fool ourselves into believing that it will be the best tool forever.”
01:08
“20 projects = 20 prisons.”
08:56
“Skills fix that because your agent will only load the skills that it needs.”
Word for word.
metaphoranalogy
00:00Claude CoWork, after receiving an incredible number of updates over the past few weeks, is now one of the most powerful AI agents for non coders, but most people are only using a fraction of its capabilities. If you just learned the 12 simple Claude Cowork tricks that I'm about to share with you though, you'll finally unlock its full potential.
00:19And I wanna start with what is probably my most controversial take. My first tip is that Claude Cowork projects should only be used as a way to organize your chats and not as a way to give Claude instructions.
00:32So inside of Claude Cowork, there are these things called projects. Here are my projects. And if you open up any of these, you'll notice that over here on the right, you can give the project instructions.
00:43The idea behind this is that you can put some instructions in here so every time you start a chat from within a project, it will automatically apply those instructions. The problem is, as great as Claude Cowork is, we shouldn't fool ourselves into believing that it will be the best tool forever. To future proof your setup, you wanna make it as transportable as possible.
01:05But if you have instructions inside of , it's gonna make it very hard to move from Cowork to another tool. So what do I use projects for?
01:17Well, I only use them for one thing, which is to organize chats. So inside of Claude Cowork, every time you click on new task, is Cowork's name for a new chat, and you start a task from here, it will appear in the recent task section here on the bottom left.
01:34But this isn't organized, so it makes it difficult to find important tasks that you wanna locate later. So I've set up projects that I use as folders. For example, here's one I've called presentations.
01:48Anytime I wanna create a new presentation, instead of clicking on new tasks at the top, I instead click on the presentations project. I tell it what I want it to create for me, and then that task will appear down here inside the project, and any outputs it creates like the slide deck will appear in this output section.
02:11That way it's all neatly organized in one place. So projects for me are basically a folder system for tasks. But if I don't use instructions in projects, how do I tell Claude Cowork how I want it to work?
02:25For that, I suggest you utilize trick number two. As you already know, Claude Cowork works directly with files and folders on your computer. Opening up my finder, this is the main folder that I have Claude Cowork work inside of.
02:40I've called it Frank Bot. Inside of this folder is a file called Claude dot m d. This is just a text file that gives instructions to Claude Cowork.
02:52Anytime it goes into this folder to do work, it will read this file and it will give it all the context it needs to help me and the general instructions for working in this folder. Another way of putting it is instead of putting all of your instructions inside of the project instructions, you take everything you would put in there and you just put it in a text file called claud dot m d.
03:18That's it. And then within each of these subfolders, which I basically use as different projects, there's another Claude dot m d file with more specific instructions for that subfolder.
03:31With this setup, anytime you switch to a new tool like Codex, you can just tell it to find all the Claude dot m d files and to change it to the name that it uses for its instruction files, and then it'll be ready to go. I've made videos about this system before and have received a ton of great questions about it. So I'm actually planning to build a school community where I'll share with members a step by step guide on how to set this up along with a lot of other AI tutorials that I don't share here on YouTube, live q and a's, and a member's forum.
04:01If you wanna find out more about that or get on the wait list, all the details are in the description down below. Tip number three. Did you know that when you start a new task and choose which folder on your computer you wanna work inside of, you can actually choose multiple folders.
04:16All you have to do is select a new folder, and you'll see which ones are selected with this check mark on the right. I find this really helpful when I want the agent to have access to information in two different folders, but I don't want it to have access to all the folders because then it will have too much information that may not be relevant to the task.
04:35So for instance, if I have it do some research for me, I might have it pull that research into the research folder, but then I might ask it to create a presentation for me about what that research was. So then I would give it access to both the research folder and the presentation folder.
04:50And while I have this drop down open, tip number four is to set a default folder if you find yourself using one folder more than all the others. So I've set my default as the parent folder that contains all the other folders that I have my agent work on. That one is called Frank Bot.
05:08So if I start a new task, it will always select that one by default. But if I want to, I can always change it. Tip number five may also be a bit of a controversial one, but you need to understand the limits of dispatch.
05:21Dispatch is a feature that's supposed to give you mobile access to Claude CoWork, but there are two issues with it. One, in order for it to work, your computer needs to be on, CoWork needs to be open, and then you need to link up Dispatch with it, and then it will work, which is impossible for everyone.
05:40They can't just leave their computer on all the time. And the second problem is that it honestly doesn't work that well. Just like one continuous conversation that you have with co work, so you can't access all of your chats.
05:54It's hard to have multiple projects going at the same time. Like for simple tasks, it works fine, but whenever I try to use it for multiple things, it just ends up frustrating me and getting confused. Tip number six I have for you is to use voice dictation the right way.
06:11Inside of CoWork, instead of typing in here, you can use a voice dictation button. It's right here. But you can also access it just by pressing command d, and it will start listening to what you're saying.
06:25But there's a secret hidden menu here. On this drop down, you can select whether you want to hold to record, which is like push to talk, or you can toggle it. So if I turn this off and now I press command d, it is now toggled on and I don't have to hold it and can just talk out loud.
06:42And then once I'm finished talking, then press command d again and it will toggle off dictation. Dictation is a very important tool when it comes to using AI agents like Claude Cowork Because a lot of times our best ideas come to us as we're talking out loud about them and just sort of rambling.
07:01And the AI agents are really good at taking our unstructured brain dumps and turning them into structured output. So don't be afraid to turn on dictation and just brain dump a bunch of ideas and let Claude do the work of sorting through all of it.
07:17Tip number seven is a quick one, but it has saved me a lot of money. When you schedule tasks, and you should be using these because this is a great way to automate your work, you wanna make sure you schedule them for off peak hours. If you schedule them for peak hours, you'll see a warning like this.
07:36And opening it up, you'll see that it tells you that you've scheduled this for peak hours, which are between 5AM and 11AM Pacific Standard Time, and that will consume your usage limits faster. So make sure if it's possible to schedule these for off peak hours. Another tip along the same lines is if you have a task that you need to run, but you know you don't have enough limits for it, just set it up as a scheduled task to run later in the day when you know your limits are going to reset.
08:10Even if you're gonna be away from your computer, then it will just automatically start. And when you get back to your computer, it will be done. Tip number eight is to use lots and lots of skills.
08:19You can think of skills as prompts that you load into Claude CoWork once so you never have to type them out every time you wanna use them. Instead of using skills though, a lot of people will create a very long set of instructions either in their claw dot m d file or worse inside their project instructions.
08:38It may say something like, every time I tell you to triage my email inbox, I want you to do x y z. And then every time I tell you to update my CRM, I want you to a do a b c. The problem with that approach is that every time you use the agent, it will read through all of those instructions, which is very inefficient.
08:56Skills fix that because your agent will only load the skills that it needs or the ones that you tell it to load. So it's incredibly efficient. And because of that, you should be using less instructions and more skills.
09:11For example, if we return back to the folders that I've set up for my agent, I have a subfolder project folder here called presentations. And inside of that, I have a claw dot m d file, which again are the instructions for the agent. But if you look at this, literally, all it says is this project folder is used when I want you to create on brand slide decks for me.
09:33And when I tell you to do that, use the slide deck skill. That's literally all it says to do. And then inside of Claude, we have the slide deck skill in here.
09:42So it's only going to load that when it needs to use it, which is a lot more efficient. And tip number nine is to use lots and lots of connectors. If you find yourself jumping between lots of apps to get your work done, see if there's a connector for it.
09:57What I'm trying to build for my school community is a system where people can log into one tool like Claude CoWork and be able to control everything from there like one central command without jumping between all these different apps. Tip number 10 is another one that will hopefully also save you some money.
10:15Take advantage of Claude design. This is included with your Claude subscription that you use to access Claude CoWork.
10:22And with this, you can create slide decks. You can create websites. You can create motion graphics.
10:27So instead of setting those up inside of Claude CoWork, which is going to use those usage limits, instead use the usage limits that are for Claude design. They're separate.
10:38So you might as well burn through those ones first before you start burning through the ones inside of Claude CoWork, which you may wanna use for other tasks. Tip number 11 I have for you is to download and use the app Obsidian. It's free to use and this will be a viewer that you can use to view the files that Claude Cowork makes for you.
10:59Like I was saying before, everything it makes for you lives inside of folders on your computer. So this isn't the most user friendly way to view what it's made.
11:10You can open up each folder and then you have to use a different app to actually view the different outputs, whether that's slide decks or websites it's made for you or even just text documents.
11:23So if you download this app Obsidian, it's free. It's available for iOS, Android, Windows, Mac, Linux, then you get an experience like this.
11:32This is the same folder, my Frank bot folder. You can see all the subfolders in here, and it's a lot easier to view what was created. So here's a text file, and I can come in here and add anything to it that I want.
11:46And even if it makes something for you that's HTML, like this slide deck it made for me, you can view that in Obsidian. Obsidian can get very complicated very quickly.
11:56You don't need to know 95% of what it can do. You literally just download it, point it to your folder, and use it like this, and don't worry about all the extra complicated stuff.
12:06And tip number 12 I have for you might seem like a contradiction, but it's really not. So I'd say about once a month, you wanna come into customize, look at all the skills that you have turned on, and just turn off the ones that you no longer use.
12:21If you have too many skills, especially ones that you're not even using, it can degrade the performance and push you closer to your limits limits faster, your token limits. Same thing with connectors.
12:34If you have too many connectors that you're not even using, then that can confuse Claude Cowork, and it may try to load a connector that you don't even want it to use. So just disconnect anything you're not using or turn off any skills that you're no longer using.
12:49So there you go. 12 tricks for leveling up Claude Cowork. Hopefully, you learned something new today.
12:54If you did, let me know by giving this video a thumbs up. Then if you wanna get on the wait list for my new school community, make sure to do that at the link in the description down below. Otherwise, thanks so much for watching the video, and I'll see you in the next one.
13:07Bye for
The Hook
The bait, then the rug-pull.
After a wave of updates turned Claude Cowork into one of the most capable agentic tools for non-coders, most users are still treating it like a chatbot. Paul Lipsky argues the gap between casual user and power user comes down to twelve decisions — and most of them take under a minute to make.
Named ideas worth stealing.
02:28model
CLAUDE.md folder hierarchy
- Root folder claude.md = global context
- Subfolder claude.md = scoped instructions
- Keep instructions minimal, offload to skills
Portable instruction architecture: store all agent rules in text files at the folder level so they survive tool switches.
Steal forClaude Code project setup, any CLAUDE.md-driven workflow
08:10concept
Skills over Instructions ratio
- CLAUDE.md should be lean
- Skills load on demand only
- One skill per repeatable workflow
Load instructions lazily via skills instead of eagerly via a bloated CLAUDE.md. Reduces token burn and keeps context clean.
Steal forJoeFlow skill architecture, MCN agent system design
07:12concept
Off-peak scheduling system
- Peak = 5AM-11AM Pacific
- Off-peak = afternoon or evening
- Schedule depleted-limit tasks to auto-start at reset time
Treat usage limits like compute credits: batch heavy tasks during cheap hours.
Steal forJoeFlow scheduled batch sessions, MCN automation scheduling
How they asked for the click.
12:11newsletter
“If you wanna get on the wait list for my new school community, make sure to do that at the link in the description down below.”
Soft community waitlist pitch repeated twice: mid-video at ~3:31 and again at outro. No hard sell, description link only.
Visual structure at a glance.
