Claude Projects vs Cowork vs Skills: When to Use Each
Pick the right Claude tool in minutes. The decision rule for Projects vs Cowork vs Skills vs Plugins — plus 3 prompts that do the setup for you.
I keep getting some version of this question: “I want to use Claude for X. Should I make a project? A skill? Is this a Cowork thing? Or should I use the global chat?”
If you’ve asked yourself the same question, you're in the right place.
Claude has shipped so many new tools so fast that even people using it every day get tangled.
Projects. Cowork. Cowork Projects. Skills. Plugins. Scheduled tasks. Computer use. Claude Code. Managed Agents. And that's not even the full list.
I already wrote the complete guide to everything Claude can do. (And by the way, I keep it updated every time Anthropic ships something new.)
That one covers what each tool is.
This one helps you decide which one to pick for what you’re trying to do. And it’s a prompt, not a chart.
Here’s what we’ll cover
How to choose between Claude Projects, Cowork, Skills, and Plugins
3 prompts to pick your Claude tool and set up Projects or Cowork Projects
Claude Projects vs Cowork Projects: how to choose and set up both
How to set up Amplifiers (the AI Blew My Mind MCP) in Claude
How to choose between Claude Projects, Cowork, Skills, and Plugins
Claude is built to work best when it's personalized to you. Every tool and every layer of instructions exists so that Claude learns how you work, what you need, and how you like things done.
The more you set up, the less you repeat yourself, and the better the output gets.
The problem is that there are a lot of places where you can set things up. And if you put the wrong instruction in the wrong place, it either gets ignored or it conflicts with something else.
So the real question isn’t just “which tool should I use”. It’s “where does this instruction live so Claude always has it when it needs it?”
That’s what the layer model helps you figure out.
The Claude layer model, explained
Claude isn’t just one tool. It’s a stack of layers. And each layer answers one question: how wide should this reach?
Think of it from the bottom up.
Profile preferences. These apply to every conversation, everywhere. Things like “be concise” or “I prefer code examples in Python”. Set it once, forget it.
Skills. A specific capability you set up once for any recurring task. Skills run everywhere: in chat, in Cowork, in Claude Code. If you find yourself repeating the same instructions across conversations, that’s a skill.
Cowork global instructions. These apply to every Cowork session on your computer. Your general working style, your defaults.
Claude Projects. A persistent workspace inside claude.ai for one body of work. Your newsletter project, your client project, your course project. Each one gets its own instructions and files.
Cowork Projects. Same idea, but inside Cowork, with memory that sticks around across sessions. This one is new. (More on it in a minute.)
Plugins. Bundles of skills and commands that run in Cowork or Claude Code. Think of them as packages you can share or install.
The autonomy layers. Scheduled tasks, computer use. These kick in when you want Claude to work without you watching.
How the layers work together
Two things about this stack matter.
First, layers don’t compete. They combine. Profile preferences load first. A skill kicks in whenever it applies. Project instructions add on top when you open that project. Claude reads all of them. That’s why you never need to repeat “be concise” inside every project. The layer below already handles it.
Second, pick the smallest layer that still covers the need. If a rule applies only to one project, it’s project-level. If a behavior applies everywhere, it’s a skill or part of global instructions. Stuffing universal rules into every project you build is a common mistake I see. And it’s why people feel like Claude “stops following the rules” halfway through a conversation. There were too many to hold.
That’s the whole mental model.
3 prompts to pick your Claude tool and set up Claude Projects or Cowork Projects
Now, nobody wants to hold a layer diagram in their head every time they have a new idea (though you develop this intuition anyway the more you use it).
So I built a set of prompts in Amplifiers (the AI Blew My Mind MCP) to handle this for you.
The first one is a decision prompt. You tell it what you’re trying to do, it asks a few questions, and it tells you which Claude tool fits best. No chart to memorize.
But picking the right tool is only step one. You also need to set up Claude well.
So there are two more Amplifiers that help with that:
A project instructions prompt that helps you write clear, effective instructions for any Claude Project.
A Cowork project setup prompt that walks you through configuring a Cowork Project from scratch (since this is brand new and there’s not much documentation for it yet).
Let me walk you through each one.
How to decide which Claude tool to use (with one prompt)
Once you have Amplifiers set up (takes two minutes), open a new Claude chat and ask Claude to search for “Which Claude tool should I use?” prompt. Then tell it what task you’re trying to figure out.
You describe what you're trying to do in plain language. "I want Claude to proofread my newsletter drafts". "Summarize my sales calls every Monday". "Write a one-time client proposal in my voice".
What you get back:
The tool. Which Claude surface fits, and why that one over the alternative you probably assumed.
The wrong assumption. Most people arrive with a guess. The prompt tells you when that guess is off, and why.
The setup steps. Three to five specific things to do right now, not vague advice.
The model. Sonnet for most things, Opus with Extended Thinking for complex or strategic work.
One high-leverage tip tailored to your specific use case.
A ready-made prompt from the Amplifiers library if one already exists for your task, so you don’t start from scratch.
If you’re getting started with Claude, or you just switched from ChatGPT and the number of options feels overwhelming, this is the prompt I’d run first.
It cuts through the “what does all this stuff do” phase and gets you to the right setup for your actual work. Instead of spending an afternoon experimenting with the wrong tool, you get a clear answer in two minutes.
And if you get an answer like I did, where the recommendation is to set up a project, that’s exactly what the next two Amplifiers help you with: writing the instructions for a Claude project, or setting up a Cowork project from scratch.
Claude Projects vs Cowork Projects: how to choose and set up both
Most of the time, the decision prompt’s answer lands in one of two places. A Claude project or a Cowork project.
They look similar. The structure is almost the same: both have instructions, both hold files, both persist across conversations. But they’re built for different kinds of work.
What a Claude Project is best for: A Claude project lives in claude.ai. You work with Claude in conversation, going back and forth, iterating together. It can create files too, but you’re there guiding it. A newsletter editorial system. A research initiative. A client engagement.
What a Cowork Project is best for: A Cowork project lives in Claude Desktop. You assign a task and come back to it done. It works with your local files, and it’s where you unlock the autonomy layers: scheduled tasks, Dispatch from your phone, and computer use. A folder of client proposals. Recurring monthly reports. Your tax prep.
Creating the project takes a minute. Setting it up so Claude actually does what you need takes a bit longer.
That's what these two Amplifiers are for:
Claude / ChatGPT Project Instructions Builder
Interviews you about your workflow, then writes copy-paste-ready instructions and tells you what to upload to the knowledge base. It also catches when a project is scoped too broadly before you waste time building it.
Cowork Project Setup
Handles the three things most people rush or skip: project instructions, folder architecture (what files to create and how to name them so Claude finds them), and scheduled task prompts that work when nobody’s watching.
Why Cowork projects are worth setting up now
Cowork projects are new. And they fixed Cowork’s biggest limitation.
Until recently, Cowork had a frustrating quirk. Memory reset every chat. You’d teach it your file structure, your preferences, what “the report” meant in your world. And the next morning it was a stranger again. Every session started from zero.
That’s gone. With Cowork projects you now get:
Persistent memory across chats. Every 24 hours, Claude updates the project’s memory based on your conversations. The useful bits, not everything you said.
Project-specific instructions separate from your global Cowork instructions.
A dedicated folder Claude can access across every session.
Scheduled tasks scoped to the project.
You can start a project from scratch, import an existing Claude project you want to bring to Cowork, or point it at a folder. The Cowork Project Setup Amplifier walks you through all of it.
And once you've got more than one, you can organize them into a full system: a Cowork operating system with a root, workspaces, and projects that all share memory, so Claude reads your whole setup automatically.
Claude skills: the layer that makes everything else smarter
Skills are what got me hooked on Claude. I’ve written about them more than anything else. My article on Claude Skills examples organized by work domains has over 85,000 views, and my writing skills and brand skill are the most appreciated ones by anyone who tried them.
All this to say: skills aren’t just a cool feature. They solve the thing that makes every AI output feel generic: Claude doesn’t know how you like your work done.
A skill is just instructions that Claude reads every time before it does a task. And it applies automatically across chat, projects, Cowork, and Claude Code. You write it once, it works everywhere.
That’s the compounding effect of the layer model: one skill improves every conversation Claude has with you, on every surface.
In Amplifiers, there’s a set of skill builders for the foundations. These are prompts that interview you step by step, then build the skill for you:
Voice DNA. Teaches Claude your writing voice: rhythm, vocabulary, banned words, how you open, how you close. My most popular skill builder.
Business Profile. Your full business context, so you stop re-explaining it in every chat.
Audience Profile. Your audience research, packaged as a reusable skill.
ICP. Your ideal customer profile. Works for B2B and B2C.
Brand Design. Your visual identity, so every PDF, deck, and doc comes out on brand.
Substack Notes. Learns from your best and worst-performing notes, generates new ones in your proven style.
Client Proposals. Tailored proposals in your voice and offer structure.
And those are just the skill builders.
But Amplifiers has 89 prompts and growing, across writing, marketing, business, operations, thinking, image generation, and more. There are prompts for lead magnets, elevator pitches, sales pages, email sequences, SOPs, onboarding documents, decision-making frameworks.
You can ask Claude anytime to search for what you need.
And any Amplifier you use often enough can become a skill. Run the prompt once, then ask Claude to turn the output into something reusable. That’s how you build the layers that make Claude feel like it was set up for exactly your work.
If you want to see how the prompts work together end to end, I put together a 20-prompt roadmap that takes you from a vague idea to a live business, all inside Amplifiers.
How to set up Amplifiers (the AI Blew My Mind MCP) in Claude
If you’ve read this far and want to try any of the prompts I mentioned, here’s how to get Amplifiers running in your Claude.
Open Claude in the desktop app: Go to Customize from the left side panel. Click Connectors, then the + icon. Select Add custom connector. Name it AI blew my mind and paste this URL: https://mcp.aiblewmymind.com. Click Add.
Now connect it: Go back to the Connectors section. Scroll down to Not Connected. Find the AI blew my mind MCP and click on it. Hit Connect. This takes you to auth.aiblewmymind.com where you’ll create your account. (Make sure you use the same email you subscribed to AI Blew My Mind with). Then click “Always allow“ for all tool permissions.
And just like that, your AI assistant is now amplified.
You’ll find a lot of free prompts inside Amplifiers. The starred ones (the skill builders and some of the more advanced prompts like Cowork Project Setup) are reserved for paid subscribers.
Your turn
Pick one task you’ve been putting off.
Open Claude. Set up Amplifiers. Run Which Claude Tool Should I Use? on it. See what it picks. Then run the builder it points you to.
That’s it. No more scrolling through articles and guessing.
What would you want to see next in Amplifiers? Leave a comment and tell me. Every new amplifier starts with someone saying “I wish this existed”.
If this helped you figure out something you’ve been stuck on, share it with someone who’s still in the “I just use chat for everything” phase. They’ll thank you.
Let’s consider a scenario in my organization: we have Claude Enterprise, but access is restricted to chat-only (no CoWork or code features). When creating a project, I can upload files directly—but we already maintain a large document repository in SharePoint. Instead of duplicating files, can a Claude project connect directly to a SharePoint library? If yes, how can this be set up?
like Ras Mic said, majority of people shouldn't have long CLAUDE.md (global, cowork) instructions. precisely because they are added to context every time you try to do something. while skills can have the same context and only be activated when needed
you need to avoid context bloat, it makes models stupid, definitely not smarter