Anthropic publishes 'Agent Skills' framework to standardize agent instructions and metadata
- Anthropic published Agent Skills docs for Claude, formalizing reusable capability bundles that package instructions, metadata, scripts, and templates for automatic use in workflows. - The docs say Skills load on demand, support up to 20 skills per session, and are not eligible for Zero Data Retention. - The release lands as Anthropic pushes managed agents and pricier coding usage into enterprise governance work. (anthropic.com)
Anthropic has published Agent Skills documentation that turns reusable agent instructions into a formal Claude feature for the API and managed agents. (platform.claude.com) A Skill is a filesystem-based bundle of instructions, metadata, and optional resources like scripts or templates that Claude can invoke automatically when a task matches. Anthropic says Skills are meant to provide domain-specific workflows and organizational context without re-pasting the same prompt every time. (platform.claude.com 1) (platform.claude.com 2) Anthropic’s docs say Claude first loads a Skill’s metadata, then pulls in the full instructions only when needed. The quickstart lists four Anthropic-managed document Skills at launch: PowerPoint, Excel, Word, and PDF. (platform.claude.com) The idea is to treat agent behavior less like a one-off chat prompt and more like a checked-in software component. Anthropic’s engineering post said it has now published Agent Skills “as an open standard for cross-platform portability” after first introducing the concept in October 2025. (anthropic.com) That framing fits Anthropic’s newer managed-agents product, which runs Claude in an Anthropic-hosted environment with file access, code execution, web browsing, and persistent agent configuration. Skills are one layer in that stack, alongside the runtime that decides when tools and code should be used. (platform.claude.com 1) (platform.claude.com 2) The enterprise guidance reads like software governance, not prompt writing. Anthropic tells admins to review, evaluate, deploy, and manage Skills at organizational scale, and warns that a malicious Skill can direct Claude to invoke tools or run code in ways that do not match the Skill’s stated purpose. (platform.claude.com 1) (platform.claude.com 2) The docs also put hard edges around the feature. Managed Agents supports a maximum of 20 Skills per session across all agents in that session, and the Skills pages say the feature is not eligible for Zero Data Retention. (platform.claude.com) (platform.claude.com) The timing matters because Anthropic is simultaneously tuning the economics of coding agents. On April 16, Anthropic released Claude Opus 4.7 as a general-availability model focused on harder software-engineering tasks, with pricing unchanged from Opus 4.6 at $5 per million input tokens and $25 per million output tokens. (anthropic.com) But Business Insider reported on April 29 that Anthropic raised its public estimate for Claude Code usage to about $13 per developer per active day, up from $6 earlier in April. The same report said Anthropic now tells customers that 90% of users stay below $30 per active day. (businessinsider.com) Put together, the message is that agent building is moving away from handcrafted prompts and toward packaged capabilities, hosted runtimes, and budget controls. Anthropic’s own docs now describe the agent as software that needs versioning, review, permissions, and cost limits before it goes to work. (platform.claude.com) (platform.claude.com)
