# Batch 3 Claude Cowork — Creation Guide

Use this guide with the personalized studio at **https://mock.netray.co/b3**.

## The four things you can build

### 1. Claude Skill

A Skill is a reusable procedure Claude can load when the request matches. Use a Skill when you repeatedly paste the same checklist, investigation method, output format, or team procedure.

Recommended project structure:

```text
.claude/
└── skills/
    └── your-skill-name/
        ├── SKILL.md
        ├── references/
        ├── examples/
        └── scripts/            # only for deterministic checks
```

Your `SKILL.md` should include:

- precise YAML frontmatter and a high-signal description;
- when the Skill should and should not run;
- `$ARGUMENTS` handling when the user supplies a goal or identifier;
- an inspect-first workflow that discovers available tools and schemas;
- expected output structure and evidence requirements;
- privacy, approval, and failure-recovery rules;
- links to detailed references rather than an oversized main file;
- realistic examples and a repeatable test checklist.

Official guide: https://code.claude.com/docs/en/slash-commands

### 2. Claude Code Plugin

A Plugin is the shareable package. It can bundle Skills, focused agents, hooks, MCP configuration, examples, and team documentation.

Current structure:

```text
your-plugin/
├── .claude-plugin/
│   └── plugin.json
├── skills/
│   └── your-skill/
│       └── SKILL.md
├── agents/                     # only when separate context is useful
├── hooks/                      # only for deterministic automation
├── .mcp.json
├── README.md
├── CHANGELOG.md
└── tests/
```

Only `plugin.json` belongs inside `.claude-plugin/`. Skills, agents, hooks, and `.mcp.json` stay at the Plugin root. Use `${CLAUDE_PLUGIN_ROOT}` for Plugin-relative paths and never embed credentials.

Useful validation flow:

1. Validate the Plugin structure and manifests.
2. Load locally with `claude --plugin-dir ./your-plugin`.
3. Confirm the namespaced Skills appear.
4. Run `/reload-plugins` after changes.
5. Use `/mcp` to confirm bundled MCP servers.
6. Exercise happy, missing-data, conflicting-data, and unsafe-write cases.
7. Document rollout, upgrade, and rollback.

Official guides:

- https://code.claude.com/docs/en/plugins
- https://code.claude.com/docs/en/plugins-reference

### 3. MCP server

An MCP server gives Claude governed tools, resources, and reusable prompts. Build one when Claude needs evidence or actions that are not already exposed safely.

Start with:

- a narrow business outcome;
- atomic, clearly described tools;
- strict input and output schemas;
- stable record identifiers and source provenance;
- structured errors and bounded output;
- resources for README, schema, policy, and examples;
- synthetic data and read-only tools first;
- optional identified grants, safe audit metadata, and rate limits;
- contract, integration, negative, timeout, and partial-outage tests.

Official guide: https://code.claude.com/docs/en/mcp

### 4. Full solution

A complete solution combines the smallest useful Plugin, focused Skills, and the right MCPs. It should make the workflow easier to repeat while keeping evidence, human judgment, and approval visible.

Recommended stages:

1. Discover capabilities and schemas.
2. Classify data, permissions, and boundaries.
3. Collect atomic evidence with source IDs.
4. Detect conflicts and alternative explanations.
5. Prepare a decision with confidence and limitations.
6. Pause for approval before consequential action.
7. Execute only approved sandbox writes.
8. Validate the effect and record privacy-safe audit metadata.

## Workshop MCP collection

- Excelsoft Analytics
- Excelsoft Books
- Excelsoft CRM
- SARAS Test & Assessment
- Excelsoft People
- Support & Exam-Day Ops
- AI-Levate
- SARAS LMS
- Excelsoft HRMS
- Excelsoft ERP

The hosted collection uses synthetic workshop data. Never paste AWS, Cloudflare, database, application, or production credentials into Claude prompts, Skills, Plugin manifests, `.mcp.json`, examples, or email.

## The fastest way to start

1. Open https://mock.netray.co/b3.
2. Choose your name.
3. Choose **Skills**, **Plugins**, **MCP builds**, or **Full solutions**.
4. Start with **Quick win**.
5. Copy the complete prompt into Claude Code.
6. Inspect Claude’s proposed files before approving writes.
7. Run the included tests and validation steps.
8. Shortlist and download the prompts you want to keep.

Managed by Netray · https://netray.co