Sonith ManakkatTeaching Claude something once is good. Teaching it once for the whole team is better. Everything a teammate would otherwise have to explain to Claude again (how we build, how we write, which tools we use) can be written down and shared. This part is a practical playbook, with a cheat sheet picture to choose the right tool. Facts come from Anthropic's Claude Code and support documentation, checked on 2026-10-21; where availability depends on your plan, check your plan.
The idea: write it down where everyone gets it
What to share, and where it lives
| What | Where it lives (as documented) | How the team gets it |
|---|---|---|
| Instructions and rules | ./CLAUDE.md or ./.claude/CLAUDE.md, plus .claude/rules/ | Commit to the repository |
| Reusable know-how | .claude/skills/<name>/SKILL.md | Commit to the repository |
| Specialised helpers | .claude/agents/ (subagents) | Commit to the repository |
| Connections to tools | .mcp.json at the project root | Commit it; Claude Code asks for approval before using project servers |
| Settings, plugins | .claude/settings.json | Commit it; each person still installs plugins on their own machine |
Expert viewDetails behind the table
Anthropic's memory page says project instructions are "shared with your team through version control", while CLAUDE.local.md holds personal project preferences and should be added to .gitignore. For skills, the Claude Code docs say to commit .claude/skills/ to share. Subagent files are Markdown with YAML frontmatter, and the docs say to commit .claude/agents/ so teammates get the same subagents. For MCP, project scope stores servers in a .mcp.json file at the project root; the docs say Claude Code prompts for approval before using project-scoped servers in interactive sessions, and that ${VAR} and ${VAR:-default} expansion lets teams share configuration while keeping machine-specific paths and API keys out of the file.
Takeaway. Commit once, and every teammate's Claude can use it.
Step by step: a shared setup
- Start with CLAUDE.mdAdd a rule when Claude gets something wrong twice
- Add a skillWhen you paste the same playbook into chat a third time
- Connect a toolWhen you keep copying data from a system Claude cannot see
- Package a pluginWhen a second repository needs the same setup
Expert viewPlugins and marketplaces
Anthropic describes a plugin as a directory of skills, agents, hooks and MCP servers that Claude Code installs as one unit, usually with a manifest at .claude-plugin/plugin.json. A marketplace is a catalog (a repository or directory with a .claude-plugin/marketplace.json file) that lists plugins; you add a marketplace once and install plugins by name, using /plugin in a session. Install scopes are user, project (enabled for everyone in the repository through the committed .claude/settings.json, though each collaborator installs it on their own machine) and local. Plugin skills are namespaced, such as /my-plugin:review.
Through managed settings an organisation can allowlist or block marketplaces and force-install plugins. A plugin runs code with your user privileges, so Anthropic's documentation points to its plugin security page for what to review first. An enabled plugin also adds its skill and agent descriptions to context on every turn, so measure what each one costs.
- In .mcp.json to Teammate has it The committed .mcp.json arrives with the project.
- Teammate has it to Approval asked Claude Code prompts before using project-scoped servers in interactive sessions.
- Approval asked to Server in use The teammate has checked the server and trusts it.
- Approval asked to Not used The server is not used.
Takeaway. Sharing a connection never skips the approval step, and secrets stay out via variables.
Sharing in the Claude app: Projects
For people who work in the chat app rather than a code editor, Projects are the shared workspace.
Shared in a project
- Project instructions
- Knowledge files
- Access for members, on Team and Enterprise plans
Not shared
- Chat context between the chats in the project, unless it is in the knowledge files
- Custom skills uploaded in claude.ai, which stay individual to each user
Expert viewRoles and limits
Anthropic's support article says members can have view-only or edit access, and editors can change instructions and knowledge, add or remove members and contribute actively. Chats inside a project do not share context unless the information is in the knowledge base. Anthropic's skills documentation says custom skills in claude.ai are individual and cannot be centrally managed by admins, so for team-wide skills use Claude Code (repository or plugin) or the API (workspace-wide). Availability varies by plan, so check your plan.
Cheat sheet: which tool for which job
| Tool | Use it for | Loads into context | Shared by |
|---|---|---|---|
| CLAUDE.md | "Always do X" rules, commands, conventions | Every session, in full | Committing the file |
| Skill | Reference material and repeatable workflows | Description always, body on demand | Committing .claude/skills/ or a plugin |
| Subagent | Context isolation, parallel or specialised work | Its own separate context | Committing .claude/agents/ or a plugin |
| MCP | Reaching outside services and data | Tool names at start, details on demand | Committing .mcp.json |
| Hook | Something that must happen every time | Nothing, unless it returns output | Settings or a plugin |
| Project (app) | Repeated chat work with the same files | Instructions and knowledge in each chat | Team and Enterprise members |
- CLAUDE.mdThe house rules
- SkillThe playbook binder
- SubagentA specialist in another room
- MCPThe phone line to other systems
- HookThe lock that always engages
- ProjectThe shared folder in the app
Habits that keep a shared setup healthy
- Review it like codeChanges to CLAUDE.md and skills go through pull requests
- Keep it shortMove rarely needed material into skills or rules
- Prune regularlyRemove outdated or conflicting lines; try /doctor prompt-audit
- Onboard with itNew joiners read the same file Claude does
If it is worth teaching Claude twice, it is worth committing once.
Frequently asked questions
Can I share skills with colleagues who use only the chat app?
Not centrally. Anthropic says custom skills in claude.ai are individual to each user, so each person uploads their own. Projects are the shared option there.
Is a private repository safe for secrets?
No. Keep secrets out of the repository and use environment variables or a secrets manager.
How do I stop a teammate's private preferences leaking into the shared file?
Put them in CLAUDE.local.md (kept out of git) or ~/.claude/CLAUDE.md.