How to build a plugin from scratch in Claude Cowork
A Cowork plugin is a group of files that can instruct Claude on how to perform a specific task—coding your methodology, workflow, and tool connections.
Anthropic releases pre-built plugins for common roles such as Sales, Finance, and Legal. Customizing them to your company's context can help tailor the plugin to your specific needs. Building from scratch is for cases where your team has internal workflows, processes, or knowledge areas that existing plugins don't cover.
What's inside a plugin?
A plugin can include several types of components. Claude will build the right combination based on what you describe.
Skills – knowledge and work processes – shape how Claude operates.
A skill might be a few lines of optional formatting descriptions, a detailed methodology encompassing your entire assessment framework, or a structured workflow that you activate by typing / in the prompt. Claude reads the description of each skill and uses it when it fits what you're working on—or you call it directly from the slash menu.
Skills can contain supporting references—manuals, legal frameworks, pricing matrices—that Claude loads when relevant. They can accept input, so /close-package North America and /close-package Europe run the same process with different scopes. Recurring skills can run as scheduled tasks with criteria applied each time. Skills can be chained together, where the output of one skill provides data for the next.
Connectors - connect Claude with your tools
A plugin that works without connectors – you can manually upload data or describe the situation – but connected tools allow Claude to retrieve what it needs during the workflow and record the results.
You can instruct Claude to perform a broad search across a connected tool or to search in a specific location, depending on the workflow's needs.
Sub-agents are specialized tools that assist with complex tasks.
One agent might research a company while another pulls data from your CRM, and a third scans Slack for context. They can run in parallel or sequentially, each with its own focus and new context window.
How to create a plugin in Claude Cowork
Understanding the components within a plugin can provide you with a framework for sharing what's happening during the build process: Knowledge, judgment, and repetitive tasks become skills, data sources become connectors. If you have existing documentation or examples of deliverables you're happy with, share them—Claude will map them to the relevant parts of the plugin.
In Cowork, describe the plugin you want to build—even just one sentence will do to get you started.
A simpler starting point is also fine – Claude will proceed with questions about workflows, tools, standards, and how you handle exceptions.
Tôi cần một plugin cho nhóm chăm sóc khách hàng của mình Changes when using plugins
With detailed instructions, Claude can handle many of these workflows. The plugin packages them into skills that anyone on the team can run.
Finance
Here's our P&L. Materiality threshold is $25K. Decompose by price, volume, and mix. Check these GL accounts against the subledger. Format the JE with debits on top. Đây là báo cáo lãi lỗ của chúng tôi. Ngưỡng trọng yếu là 25.000 USD. Phân tích theo giá, khối lượng và hỗn hợp. Kiểm tra các tài khoản GL này so với sổ phụ. Định dạng bút toán ghi nợ ở trên. No plugins required : Claude can handle it, but having to repeat your entire bookkeeping method in each session increases overhead, and the output may vary depending on the Claude instructor.
/monthly-close North America With your plugin: Skills load your account charts, key performance thresholds, closing dates, and narrative standards. Connector pulls data from your warehouse. Skills run the entire process—a new analyst runs it and gets consistency.
Sell
Pull usage data for Meridian Health from our analytics dashboard. Check for support escalations in the last 90 days. Compare their current contract against our pricing tiers. Score the renewal risk and flag expansion opportunities. Lấy dữ liệu sử dụng cho Meridian Health từ bảng điều khiển phân tích của chúng tôi. Kiểm tra các yêu cầu hỗ trợ trong 90 ngày qua. So sánh hợp đồng hiện tại của họ với các mức giá của chúng tôi. Đánh giá rủi ro gia hạn và lưu ý các cơ hội mở rộng. Without using plugins : You are guiding Claude to each source, repeating the scoring criteria and specifying the output format each time.
/renewal-score Meridian Health With the plugin : Skill contains your scoring criteria and competitive position. Connector pulls data from your CRM and analytics dashboard. Claude scores accounts and notes expansion opportunities—consistent quality whether performed by a senior or new employee.
Legal
Review this vendor agreement. Our standard on indemnification is capped at 2x contract value. Flag auto-renewals over 12 months. Use GREEN/YELLOW/RED. Xem xét thỏa thuận nhà cung cấp này. Tiêu chuẩn bồi thường của chúng tôi được giới hạn ở mức gấp 2 lần giá trị hợp đồng. Lưu ý các gia hạn tự động quá 12 tháng. Sử dụng MÀU XANH LÁ/VÀNG/ĐỎ.
No plugins required : This prompt includes several checks on the contract. A full review applies dozens of standard places across all types of clauses, and you will need to reiterate all of them for each new agreement.
/due-diligence With the plugin : Skill contains standard positions on all types of terms and your severity rating. Connector accesses the data room. Claude reviews all documents against your manual and generates a risk-rated summary.
Complete plugin
- Use it in real-world work and refine it gradually . If you run a skill and something goes wrong—a step is missing, criteria need adjusting, output format is incorrect—report it to Claude and it can directly update the plugin files. Reviewing the files after Claude builds them can also help detect shortcomings early.
- Show Claude the product to be delivered . Upload an example or show Claude an example on the connected drive. Claude will directly recognize the structure, center of gravity, and format.
- Keep skills focused . Claude will generate many skills when a quest spans multiple areas, and focused skills with specific descriptions tend to be more reliably triggered than generic ones. If a skill isn't loading when you expect it, its description might be too vague—structure it by function, usage time, and scope.
- Consider using sub-agents for multi-source or lengthy tasks . If the workflow retrieves data from multiple tools simultaneously, or if a task frequently encounters context limitations due to processing too much at once, sub-agents allow Claude to break down the work across separate context windows.
You should read it
- ★ How Claude Cowork helps you automate your familiar workflows.
- ★ How to organize files on your desktop with Claude
- ★ What is Claude Computer Use? What is a ChatGPT agent?
- ★ A guide to building AI workflows with Claude Code (even if you're not technically savvy)
- ★ 4 plugins that help GIMP become a professional tool.
- ★ The most detailed guide on how to change your location on your phone.
- ★ 3 factors that contribute to the successful deployment of AI agents in the real world.
- ★ 10 situations where you shouldn't use ChatGPT
- ★ Instructions on using Copilot Vision to analyze screen content.
- ★ 10 essential skills for Claude (and any programming agent)
- ★ What is ChatGPT Go? Is it worth it?