Anthropic documents Claude Skills workflows

Anthropic’s current Claude documentation lays out a fuller picture of how the company wants “Agent Skills” to work in practice: not as one-off prompts, but as reusable bundles of instructions, scripts and reference material that Claude can discover and load when a task calls for them. Anthropic says Skills are filesystem-based resources that provide domain-specific workflows, context and best practices, and that Claude only loads them when relevant. The documentation ties that design to enterprise and developer use cases. Anthropic says pre-built Skills cover common document tasks including PowerPoint, Excel, Word and PDF handling, while custom Skills can package organizational knowledge for internal workflows. Anthropic’s product blog says those Skills can be used across Claude apps, Claude Code and the API. (platform.claude.com) ### So what is a Claude Skill, exactly? Anthropic defines Skills as directories that contain instructions, executable code and reference materials. In the overview guide, the company says the format is meant to work more like onboarding material for a new employee than a single prompt pasted into a chat box. Anthropic says the system uses staged loading rather than pulling everything into context at once. (platform.claude.com) Metadata is always loaded, the main `SKILL.md` file is read when the Skill becomes relevant, and additional files are loaded only as needed. Anthropic says that approach is intended to preserve context for the rest of a conversation. ### How does Anthropic say agents discover and run these workflows? (platform.claude.com) Anthropic’s product blog says Claude scans available Skills for relevant matches while working on a task, then loads only the minimum information and files needed. The company says Skills are composable, meaning Claude can identify and coordinate more than one Skill when a request spans multiple tasks. The API guide says Skills integrate through Claude’s code execution environment and can come either from Anthropic or from a customer’s own workspace. (platform.claude.com) Anthropic says the integration shape is the same in both cases, with users specifying Skills in the container parameter and including identifiers such as `skill_id`, type and optional version. ### Why do the docs keep stressing governance and review? (claude.com) Anthropic’s enterprise guide says Skills deployed at scale should go through security review and vetting. The company lists code execution, instruction manipulation, MCP server references, network access patterns and hardcoded credentials among the risk indicators that should be checked before approval. (platform.claude.com) Anthropic says enterprise reviewers should read all content in a Skill directory and assess what tools, scripts and external access it invokes. The guide frames that as a separate question from whether the platform is safe in general, focusing instead on how to vet a specific Skill before deployment. ### What do the docs say about token budgets and context windows? (platform.claude.com) Anthropic’s authoring guide says “the context window is a public good” and warns that Skill content competes with system prompts, conversation history, other Skills’ metadata and the user’s request. The company says only metadata is preloaded at startup, but once a Skill is opened, every token in it matters. (platform.claude.com) Anthropic’s model overview says Claude Opus 4.7 and Claude Sonnet 4.6 support 1 million-token context windows, while Claude Haiku 4.5 supports 200,000 tokens. A separate launch note for Opus 4.7 repeats the 1 million-token figure and says the model is aimed at complex reasoning and agentic coding. ### Where does this leave developers and enterprise teams? Anthropic’s API guide says custom and Anthropic-managed Skills use the same execution environment, but differ in where they come from and how they are managed. (platform.claude.com) The guide also says the feature is not eligible for Zero Data Retention and follows the standard retention policy for the feature. Anthropic’s public documentation now splits the work into separate tracks: overview pages for how Skills function, authoring guidance for writing concise `SKILL.md` files, enterprise governance for vetting deployments, and model pages for the context limits those workflows have to fit inside. (platform.claude.com) Those pages were live on May 24 on Claude’s docs site and product blog. (platform.claude.com 1) (platform.claude.com 2)