We stumbled into something interesting tonight while doing routine work inside the Claude ecosystem. The kind of thing that seems small at first, then reframes how you think about the whole platform.
Here's the setup. Anthropic now has three surfaces where you can interact with Claude as an agent: Claude.ai (the web app), Claude Cowork (the desktop environment for knowledge workers), and Claude Code (the CLI for developers). All three are powered by the same underlying model and the same Agent SDK. They share the same account. And if you connect an external service, like Slack or Gmail or Notion, through the MCP (Model Context Protocol) connector system, that connection is available everywhere. Web, Cowork, Code. You configure it once, and it follows you.
We discovered this by accident. We were working in Claude Code on a completely unrelated task, needed to send a Slack message, and realized the Slack connector we'd set up in Claude.ai was already available. No extra configuration. No copy-paste of API keys. It just worked.
That was the good news. The gap showed up immediately after.
Skills Are Still Siloed
Skills are reusable prompts that extend what Claude can do. Think of them like plugins with instructions. You write a SKILL.md file that tells Claude how to handle a specific workflow: triaging email, creating PowerPoint decks, managing an AI tools catalog, writing in a specific brand voice. When a skill is loaded, Claude operates with that specialized knowledge as if it were built in.
We have about 20 of these skills in Cowork. They cover everything from investment research to social media posting to email triage across three Gmail inboxes. They've become essential to how we work.
The problem: skills created in Cowork don't appear in Claude Code. And skills created in Claude Code don't appear in Cowork. Each surface maintains its own local skill directory. There's no sync, no shared registry, no way to say "these are my skills" and have them follow you the way connectors do.
Bridging the Gap Manually
So we bridged it ourselves. The process was straightforward but manual.
First, we exported all 20 skills from Cowork. Each skill is a directory containing a SKILL.md file with YAML frontmatter (name, description) and the full prompt body. We committed these to a GitHub repo so they'd be version-controlled and shareable.
Then we converted them into Claude Code's format. Claude Code stores global commands as .md files in ~/.claude/commands/. The filename becomes the slash command name. So email-triage/SKILL.md in Cowork became ~/.claude/commands/email-triage.md in Claude Code. We stripped the frontmatter, kept the content, and wrote all 20 files programmatically.
The result: every skill we use in Cowork is now available in Claude Code as a global slash command. /email-triage, /gadoci-consulting-writing-style, /ai-tools-admin, all of it. Same prompts, same behavior, different surface.
It took about five minutes. But it shouldn't have been necessary at all.
The Convergence Pattern
This is not unique to skills. Watch the pattern.
Scheduled tasks launched in Cowork first. Then they came to Claude Code. MCP connectors started in Claude.ai, then became available across all three surfaces. Skills exist in all three, but they haven't been unified yet. Anthropic has publicly stated they're working on cross-device sync for Cowork, and the recently announced plugin system (bundles of skills, connectors, and slash commands packaged for specific roles) suggests the architecture for unification is already taking shape.
Every major capability starts in one surface, proves itself, then migrates to the others. The trajectory is clear: full convergence. One set of skills, one set of connectors, one set of scheduled tasks, one set of plugins. All available everywhere you interact with Claude.
What This Actually Means
Step back and look at what's being assembled. You have an agent that can read and write files, execute code, connect to external services (email, calendar, Slack, CRM, databases), run on a schedule, operate with specialized knowledge through skills, and collaborate with humans through natural language. It runs on your desktop, in your terminal, and in your browser.
That's not a chatbot. That's not even an assistant. That's the early shape of an operating system for knowledge work.
The comparison to Microsoft isn't hyperbolic. Microsoft built the operating system that defined how people interacted with computers for decades. What Anthropic is building defines how people interact with AI agents. The connectors are the device drivers. The skills are the applications. The plugin system is the app store. The agent itself is the kernel.
We're maybe 18 months into this, and the pieces are already recognizable.
The Ten-Year Education Problem
Here's the part that doesn't get discussed enough. When Microsoft launched Windows, it took a decade of training, documentation, certification programs, and institutional learning before the average knowledge worker was proficient. And that was for software with buttons you could see and menus you could click.
This is harder. The interface is natural language. The capabilities are dynamic. The skills can be customized per organization, per role, per individual. There's no fixed menu of options. The ceiling for what you can do keeps rising, and most people don't know where to start.
That's the real opportunity. Not building the technology. Anthropic is handling that. The opportunity is in helping organizations understand what's possible, build the skills and workflows that matter for their context, and develop the institutional knowledge to use this effectively.
The gap we found tonight, skills not syncing across surfaces, is a small inconvenience with a five-minute workaround. But it's a signal. It tells you the platform is still being assembled, the pieces are converging faster than most people realize, and the organizations that start building their skill libraries and workflows now will have a significant advantage when everything clicks into place.
We wrote 20 skills. Most organizations haven't written one. That gap is going to matter.