One Role Is Just a Folder of Markdown: Anthropic Is Releasing Org Charts One Role at a Time
11 open-source plugins covering sales, legal, finance, customer support, and product management. Installation is a two-command affair; what is really worth looking at is their internal structure: a bunch of Markdown plus a few tool calls, with one role corresponding to one folder.
Anthropic has open-sourced 11 plugins, covering sales, marketing, legal, finance, data analysis, and product management, plus customer support, enterprise search, biological research, productivity, and plugin management. The repository is called knowledge-work-plugins, built for Claude Cowork and also compatible with Claude Code.
Each plugin has the same structure and consists of four parts:
- skills: domain knowledge and operating procedures that Claude automatically invokes in relevant scenarios
- commands: explicit slash commands, such as
/sales:call-prepand/finance:reconciliation - connectors: connect Claude to CRMs, project management tools, data warehouses, and design tools through MCP
- sub-agents: sub-agents bundled with the plugin
Installation is not complicated. In Cowork, install directly from claude.com/plugins. For Claude Code, use the command line:
claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install sales@knowledge-work-plugins
After installation, it activates automatically; skills are triggered by context, and slash commands are directly available in the session.
It is easy to read this news as 'Claude can now do sales.' But after going through the repository, you find something smaller and stranger: a role has been broken down into a folder of Markdown instructions plus a few tool calls. Sales, legal, and finance each correspond to a pile of files, with no code and no build step.
Alex Zu put this precisely in a reply: Anthropic is releasing org charts one role at a time.
So how long can these plugins hold up against real data? Someone in the comments asked directly. Ofek Shaked's question was: among the 11 plugins, which ones will still be usable once real customer data comes in? Yarad's answer was that these are only starting points; the value depends on which company data is connected, what permissions are configured, and how much the processes are changed. A sales plugin is only meaningful when connected to a well-maintained CRM, and a finance plugin needs human verification.
Kevin Lin added a more specific point: customer data, marketing content, and code have completely different risk levels; when using plugins across domains, the applicable scope and human review points must be clearly defined. The same level of autonomy cannot be applied to every process.
These two judgments are more important than the plugin list itself.
Every plugin in the repository can be modified. To swap connectors, edit .mcp.json to point to your own tool stack; to add company context, write terminology, org structure, and processes into skill files; to adjust processes, directly change the steps written in the skill. The entire repository is Markdown plus JSON; fork it, make changes, and submit a PR. There is also a cowork-plugin-management plugin specifically for creating new plugins or modifying existing ones.
The 11 plugins cover quite a few roles, but treating them as 'out-of-the-box experts' will lead to pitfalls. Anthropic states clearly in the README that these are generic starting points, and they become truly useful only after being customized to a company's actual situation.
Repository: anthropics/knowledge-work-plugins
Source material: Yarad and the comments below
发布时间: 2026-11-15 05:28