Anthropic defines agent 'skills' model

Agent frameworks keep running into the same problem. A giant prompt can tell a model what to do once, but it does a bad job of turning that behavior into a reusable capability. Anthropic’s answer is “Agent Skills” — packaged chunks of expertise that sit somewhere between a prompt, a tool, and a mini app. The company’s Claude API docs now lay out that model pretty clearly, and the timing matters because Anthropic only put Claude Managed Agents into public beta on April 8. (platform.claude.com) ### What is a “skill” here? A Skill is a modular capability for Claude. Anthropic describes it as a filesystem-based resource that can include instructions, metadata, and optional resources like scripts or templates. That sounds simple, but the important part is the packaging. Instead of pasting the same operating rules into every session, developers can define a reusable unit once and let Claude reach for it when the task calls for it. (platform.claude.com) ### Why not just use prompts? Because prompts are blunt instruments. Anthropic draws a clean line here: prompts are conversation-level instructions for one-off tasks, while Skills are reusable capabilities that load on demand. Basically, the model doesn’t need to carry every specialized workflow in context from the start. That saves context window space and makes behavior easier to reuse across products and sessions. (platform.claude.com) ### How does Claude decide when to use one? Anthropic built this around “progressive disclosure.” The metadata for a Skill gets loaded at startup so Claude knows the Skill exists and what it is for. If a user asks for something relevant — say, a PowerPoint deck or spreadsheet analysis — Claude can then load the fuller instructions for that Skill, and finally exe(platform.claude.com)mp. (platform.claude.com) ### What can Skills actually do? Right now, Anthropic highlights pre-built Skills for PowerPoint, Excel, Word, and PDF work. Developers can also upload custom Skills to their organization. In the Messages API, both kinds use the same basic shape — `type`, `skill_id`, and optionally `version` — and both run through code execution. That symmetry is the real produc(platform.claude.com 1)(platform.claude.com 2) ### Where do these show up? Across a lot of Anthropic surfaces. The docs say Skills can be created in Claude Code, uploaded through the API, or added in claude.ai settings. They also plug into Managed Agents, where developers can attach them to agents in a session. There are limits — up to 8 Skills per Messages API request, and up to 20 per Managed Agents session across all agents in that session. (platform.claude.com) ### What’s the catch? The catch is that a Skill is not just text. Anthropic’s enterprise guidance treats bundled scripts, MCP references, network access, hardcoded credentials, and broad filesystem access as concrete risk indicators. In plain English — once “capability” becomes a folder with code and instructions, governance gets harder. You’re no longer reviewing a prompt. You’re reviewing behavior, permissions, and possible data paths. (platform.claude.com) ### Why does this matter now? Because Anthropic is building a more opinionated agent stack. Managed Agents arrived in public beta on April 8, memory followed in beta on April 23, and Skills now read like one of the core abstractions for making those agents specialized without hardcoding everything into one giant system prompt. Turns out this is less about a docs tweak and more about Anthropic defining the unit of reusable agent behavior. (platform.claude.com) ### Bottom line? Anthropic is basically saying the future agent is assembled, not just prompted. Skills are the building blocks — but once you package expertise as code-plus-instructions, you also inherit software-style versioning, review, and security problems. (platform.claude.com)