Anthropic just stopped pretending AI is a tool and started treating it like infrastructure.
The Summary
- Anthropic now runs most internal work through Claude, using it as an "internal operating system" that replaces traditional software workflows
- They've built "Skills," version-controlled prompt packages that turn one-off Claude interactions into repeatable company-wide workflows
- This isn't about using AI to enhance existing systems. It's about replacing the operating logic of work itself.
The Signal
The most telling detail here isn't that Anthropic uses its own AI. It's how they use it. Mike Krieger says work now "begins with a single prompt" and Claude interprets intent, pulls context, and produces outputs that "often bypass the underlying systems entirely." Read that last part again. The systems still exist. The AI just ignores them.
This is the first concrete look at what happens when you stop bolting AI onto your workflow and let it become the workflow. Most companies are still in the "AI copilot" phase, where the model suggests code or drafts emails inside the old interface. Anthropic skipped that. They went straight to prompt-as-primitive.
"Work no longer needs a fixed system to run through."
The "Skills" layer is the real innovation buried in this story. When someone in finance figures out how to use Claude for contract review, that workflow becomes a packaged, reusable asset. It's version-controlled. It spreads across the company. The next person gets the same quality on day one, not after three weeks of trial and error.
This solves the core problem with prompt-driven work: variance. If every employee is writing their own prompts for the same task, you get inconsistent outputs and no way to audit quality. Skills turn tribal knowledge into infrastructure. They're the middleware between human intent and AI execution.
Key mechanics of Anthropic's AI-first workplace:
- Prompt becomes the interface, not an input field inside software
- Claude pulls context across systems without requiring integrations
- Skills package proven workflows into repeatable, version-controlled units
Here's what this really means. Most enterprise software is built on the assumption that humans need structure to stay consistent. You fill out forms. You follow approval chains. You click through menus. The UI enforces the process. Anthropic is betting that AI can enforce the process instead, and the UI can collapse into natural language.
Krieger frames this as making employees "materially better at the work they already do, and capable of doing things they could not reliably do on their own." That second part is the shift. Not faster email. Not better summaries. New capabilities that didn't exist before because the cognitive load was too high or the coordination cost was too steep.
The Implication
If this works at scale, the entire SaaS category has a problem. Why buy workflow software if the workflow is just a prompt? Why integrate ten tools if Claude can query all of them through APIs and synthesize the output? The operating system metaphor isn't marketing. It's a map of where enterprise software is headed: toward invisible infrastructure that surfaces only when you ask for it.
Watch how fast other AI labs copy the Skills concept. Version-controlled prompts are the new APIs. The companies that figure out how to package, share, and govern them will own the next layer of enterprise tooling. The ones still selling seats to point-and-click interfaces are selling the past.