Walkthrough: Claude Cowork: plugins, skills & local work

Cowork is where plugins, connectors, hooks and sub-agents run together. Open this tab first, because hooks and sub-agents only run in Cowork, in plain chat they appear grayed out. You should land on the Cowork home, a Let's-knock-something-off-your-list headline above the task composer. If you cannot see the sidebar at all, it is collapsed: the Reference screens gallery at the end of this lesson shows the expanded sidebar with the Cowork switch at the top.

• 1Team/Enterprise only. No Cowork tab? Your plan does not include it. Ask your workspace owner to enable Cowork on a Team or Enterprise plan.
• 2Why start here. Configuring plugins from the Cowork tab lets you use their full power, hooks and sub-agents included.

🎙Voiceover Start in the left sidebar and click Cowork to open the workspace where your tasks, files and tools all live together.

Customize is the single hub for Plugins and Connectors. It sits with your account controls near the bottom of the left sidebar.

• 1Find it lower-left. Customize lives near the bottom of the left sidebar, by your account controls.
• 2One hub, two tabs. Everything you need today is under Customize: the Plugins tab and the Connectors tab.

🎙Voiceover Now click Customize in the sidebar. This is where you manage plugins, connectors and skills that shape how Cowork behaves.

The Plugins tab lists what is already installed, scan it first so you do not duplicate something your org pushed to you. Clicking the + beside Personal plugins opens the small menu shown here: Browse plugins (the route you take next) and Create plugin.

• 1Check what's installed. Scan the current list so you do not duplicate a plugin your org already pushed to you.
• 2Org-managed lock. Organisation-managed plugins cannot be edited and may not be uninstallable, that is expected.

🎙Voiceover Open the Plugins tab, then click the plus beside Personal plugins, choose Browse plugins to see what is available.

"Browse plugins" opens the gallery of available plugins grouped by marketplace. You are still only browsing, nothing is installed yet.

• 1Browse, don't commit. Read each plugin's description and bundled skills before deciding. Install is a deliberate next step.

🎙Voiceover Click Browse plugins to see what's available. Have a quick look through before installing anything, so you pick what actually fits your work.

Anthropic & Partners is the default marketplace and holds general role plugins. Personal plugins are shown separately when your workspace has custom or user-installed plugins.

• 1Pick your domain. Look at the chip row under the search bar. Anthropic & Partners is selected, with Personal beside it. That row is the marketplace switch; everything else on this screen is unchanged from the last step.
• 2Add your own later. Under "Personal plugins" click "+" then "Add marketplace", then choose "Browse Anthropic sources" or "Add from a repository" for a GitHub/git URL.

🎙Voiceover Select the Anthropic and Partners marketplace. It groups role-based plugins built for everyday office tasks like writing, research and review.

Opening a role plugin shows its description, example prompts, bundled skills, and connected services. If it is not installed yet, install it; if it is already installed, use Manage to review it before giving it real work.

• 1One at a time. Install a single plugin first, test it, then add others. Fewer moving parts means easier debugging.
• 2Grayed-out items. Hooks and sub-agents are active here in Cowork. In plain chat they appear grayed out because they only run in Cowork.

🎙Voiceover Open a role plugin that matches what you do. If it is new, install it; if it is already installed, use Manage to inspect what it adds.

The Manage view shows the installed plugin source, version, author, skills and connector set. Customize opens a new Cowork task pre-filled with a prompt asking Claude to tailor the plugin before you give it real work.

• 1It opens a task. "Customize" launches a fresh Cowork task with a ready-made prompt, you steer the customisation in conversation, not in a settings form.

🎙Voiceover If you'd like, open Manage and click Customize in the upper right to tweak the plugin. This step is optional, so skip it if the defaults suit you.

Connectors link Claude to external services like Google Drive. This is the primary, settings-side place to add and manage them.

• 1Two ways in. You can also reach this from chat: click "+" (or type "/"), hover "Connectors", then "Manage connectors".

🎙Voiceover Back in Customize, open the Connectors section. Connectors link Claude to outside tools and accounts like your file storage.

This opens the add menu, choose Browse connectors to open the catalogue and find Google Drive. Adding a connector here makes it available to enable later, per task.

• 1Available, not active. Adding a connector here does not auto-use it. You still toggle it on inside each task.

🎙Voiceover Click the plus next to Connectors to add a new one. You'll choose the service and grant access in the next couple of steps.

On some connectors the button reads "Install" instead of "Connect". Either way, follow the authentication prompts to sign in to Google and approve the requested scopes. Grant the narrowest access that still does the job. After you approve in the Google window it closes, and Google Drive shows as connected back in the Connectors list, that is your success check before moving on. (If the card shows a settings cog rather than +, it is already connected on this workspace.)

• 1Read the scopes. The Google consent screen lists exactly what Claude can see, read it before clicking Allow.
• 2Authenticate as yourself. On Team/Enterprise, each person authenticates individually. Sign in with the account that owns the Drive files you intend to work on.

🎙Voiceover Pick Google Drive, click Connect, then approve the request in the Google window so Claude can reach your files when you allow it.

Connectors are enabled per conversation, so you need an actual task open to switch Google Drive on. This keeps each task's access explicit and auditable. The Work in a project control in this composer picks the folder Cowork works in, the Reference screens gallery shows its menu.

• 1Scope is per task. A connector enabled in one task is not automatically on in the next, that is by design.
• 2Ask mode is the safety net. Next to the composer sits the permission mode control set to Ask. Claude pauses for your approval before each action. The Reference screens gallery shows the menu behind it; leave Act without asking alone until a workflow is proven.

🎙Voiceover Head back and start a new task in Cowork. Each task is its own conversation, with its own tools and its own settings.

Toggling it on tells Claude it may use Google Drive for this conversation only, each row in this submenu has a toggle on its right edge (just off-frame in this capture; scroll the submenu if Drive is not visible at first, services list alphabetically). An active toggle shows filled. Leaving it off keeps the task untouched by Drive.

• 1On means usable. Claude only reaches for a connector that is toggled on for this specific conversation.
• 2Turn off when done. For sensitive work, toggle the connector off once the task is finished.

🎙Voiceover Click the plus, hover over Connectors, and toggle Google Drive on. Remember, connectors are per-conversation, so you set this for each task.

Go to Customize > Connectors and click the connector's name to open its Tool permissions. (We show the Figma connector because it has all three tool categories. Google Drive exposes the same Always allow / Needs approval / Blocked controls.) Keep read-only tools on Always allow only when appropriate, and keep interactive or write/delete tools on Needs approval for approve-first work. You should see read-only tools grouped separately from interactive and write/delete tools, each group with its own setting. The Reference screens gallery below includes a second connector (Airtable) where a single read-only tool has been bumped to Needs approval, you can override per tool, not just per category.

• 1Approve before acting. Needs approval keeps a human checkpoint in front of interactive, write, upload, send, or delete actions.
• 2Block when unsure. If a tool can touch sensitive records or you do not understand its effect, leave it on Needs approval or block it.

🎙Voiceover Open Tool permissions and check the connector's defaults. Keep risky tools on Needs approval so Claude stops for your confirmation before acting.

Typing "/" (or clicking "+") lists the skills your role plugin added. Selecting one tells Claude exactly which capability to apply instead of leaving it to guess. The list you see contains every skill available to you, personal skills you have added (as in this capture) alongside the skills your installed plugin contributed, each named like /design-critique. Pick one of the plugin's skills; seeing it inserted into the composer is your success check.

• 1Slash to discover. The "/" menu shows every skill from your installed plugins in one place.
• 2Be specific. Choosing the exact skill gives a tighter, more predictable result than a vague free-text ask.

🎙Voiceover Type a forward slash to bring up your skills, then pick the one you want. This invokes a skill from the plugin you installed earlier.

Hand Claude a real local file and a clear instruction, then confirm the tool calls one by one. Because access is per task and tool calls are approved, you stay in control the whole way. While Claude works, the right-hand panel shows three things worth watching: Progress (step ticks for longer tasks), Working folder (where outputs land), and Context (which connectors this task can reach, yours should list Google Drive; this capture shows a task with different connectors loaded).

• 1Watch the approvals. With risky tools set to Needs approval, confirm each tool call so nothing touches Drive or your file unexpectedly.
• 2One file, one goal. Keep the first run narrow, a single file and a single objective, so the result is easy to verify.

🎙Voiceover Attach a local file, describe the task plainly, and approve each tool call as it comes up so you stay in control of every action.

Compare what Claude produced against your original file and the skill you invoked. Treat the first output as a draft to verify, not a finished deliverable, before saving or sharing it.

• 1Verify before trust. Spot-check the output against the source, confirm numbers, names and claims before relying on it.
• 2Refine in place. If something is off, ask Claude to adjust in the same task so the context and connector stay loaded.

🎙Voiceover Open the result and check it against the source. Treat this first output as a draft, then refine it until it's genuinely right.