Most of the conversation around Claude's desktop experience focuses on skills. That makes sense. Skills are a visible, shareable, installable thing. Anthropic open-sourced the Agent Skills spec, the ecosystem took off, and now there are hundreds of thousands of SKILL.md files floating around GitHub and skill marketplaces, compatible not just with Claude Code but with Codex CLI, Gemini CLI, Cursor, and others.
Skills matter. But the conversation about skills is missing the bigger picture of what actually makes Cowork different from every other AI interface available right now.
What skills actually are
A skill is a markdown file with instructions that gets loaded into context when it's relevant. That's it. When you ask Claude to create a PowerPoint, it reads the pptx SKILL.md file from the filesystem, pulls those instructions into the context window, and follows them. If the skill references other files (a template, a schema, a script), Claude reads those too. The skill itself never sits in memory permanently. It gets called in when needed and ignored when it's not.
This is a good design. It keeps the context window clean. It means you can have dozens of skills available without burning tokens on ones you're not using. But the mechanism underneath, pulling the right information into context at the right time from the local filesystem, is the thing worth paying attention to. Skills are one application of that mechanism. They're not the whole story.
Building context as you work
Here's what changes when you use Cowork the way it's designed to be used. You don't just ask questions and get answers. You create artifacts along the way. Research summaries saved as markdown files. Meeting notes. Decision logs. Project briefs. Competitive analysis. A running document where you're working through a strategy. These files live on your filesystem, in folders you've granted Cowork access to.
The next time you ask Cowork to do something related, it can read those files. Not because you uploaded them into a chat window, but because they're sitting right there in the project folder. You're building a modular, persistent knowledge base as a natural byproduct of doing your work. And Claude can traverse it, reference it, and build on it.
This is fundamentally different from how web-based AI chat works. In a browser-based conversation with Claude, ChatGPT, or Gemini, context lives and dies inside the conversation window. You can upload files, but they're attached to that session. You can use projects to persist some reference documents, but those are static. They don't grow and evolve as you work. The context you're building in one conversation doesn't automatically inform the next one.
In Cowork, it does. Because the context isn't trapped in a chat thread. It's on your filesystem.
The compounding effect
This matters more than it sounds like it should. When you're three weeks into a project and you've been using Cowork throughout, the accumulated context is substantial. There's a research file from week one. A draft strategy doc from week two. Notes from three different conversations. A competitive landscape summary. A set of decision criteria you worked through.
When you ask Cowork to draft a presentation on that project, it doesn't start from zero. It reads the folder. It pulls in what's relevant. The output is informed by everything you've built along the way, not just what you remembered to paste into a prompt.
This is the difference between an AI that answers questions and an AI that works with you over time. The quality gap widens the longer you use it, because the context keeps compounding.
Where the other platforms stand
ChatGPT has a desktop app for macOS and Windows. It's a solid product for chat, and it has connectors to cloud services like Google Drive and SharePoint. But its relationship with your local filesystem is different. When ChatGPT generates a file, it creates it in a cloud sandbox and gives you a download link. There's no persistent workspace folder, no automatic saving to your local file system. You can upload files into conversations, and you can add static files to Projects for persistent reference. But ChatGPT doesn't read and write files in your local folders the way Cowork does. You can't point it at a directory and say "work here."
ChatGPT Projects are the closest analog, and they're useful. But the files in a Project are snapshots. You upload them once and manually update them when they change. They don't grow as you work. There's no equivalent of Cowork creating a research summary, saving it to your project folder, and then referencing it three days later in a different task.
(A technical note: ChatGPT's desktop app does support MCP, so you could set up a filesystem MCP server to bridge some of this gap. But that requires manual configuration, and it's not the native experience the way Cowork's folder access is.)
Gemini doesn't have an official desktop app for knowledge workers. The web interface at gemini.google.com supports file uploads and connects to Google Drive through Workspace extensions. Gemini CLI exists for developers and supports local file access, but that's a terminal tool, not a knowledge work environment.
The pattern Cowork establishes is compelling enough that it's reasonable to expect others will follow. A desktop application with direct filesystem access, where the AI can read and create files as part of its workflow, where skills and connectors and accumulated project context all work together. That's the direction knowledge work with AI is moving.
Why this matters for how you work
The practical implication is straightforward. If you're using AI for real work (not just one-off questions, but sustained projects that unfold over days and weeks), the tool that can maintain and build on context over time will produce better results than the one that starts fresh every session.
Skills get the attention because they're the portable, shareable piece. And they're genuinely useful. But skills are just one type of context that gets loaded from the filesystem. The markdown files you create as you work are another. Your CLAUDE.md project instructions are another. The notes and drafts and research you accumulate are another. All of it feeds the same mechanism: the right information, loaded into context, at the right time.
That's what Cowork actually is. Not a chat interface with skills bolted on. It's a working environment where context accumulates, persists, and compounds. And right now, nothing else works quite the same way.