Cowork Features Tour

Agenda

01

Skills: from /command to reusable workflow

The Skills section opens with the same light question from the deck: what workflow have you repeated twice this month that deserves a name? A Skill is a named workflow Claude can load when the request matches, or the user can call directly with /. The slash command is the doorway, not the point. The deeper lesson is Skill anatomy and limits: name max 64 characters, description max 1,024 characters, always-visible metadata, a SKILL.md body that loads when relevant, and optional bundled resources such as references, scripts, templates, and assets. Good defaults keep metadata light, triggered bodies under 5K tokens, SKILL.md under 500 lines, and long references loaded on demand. Create a Skill by interviewing the workflow first: what should it enable, when should it trigger, what output should it produce, and what edge cases matter? Add test cases when the output is objective. Use examples and review rubrics when the output is judgment-heavy.

Room prompt

Pick one recurring GGP workflow and answer four questions: what should this Skill enable, when should it trigger, what should the output look like, and should it have test cases?

02

Agents

The Agents section keeps the splash question first: where would you trust Cowork to choose the next step, and where would you give exact steps? Use a Skill when the workflow path is known. Use an Agent when Cowork needs to plan, break complex work into subtasks, inspect evidence, run code in an isolated VM, or split work into parallel streams. Cowork can coordinate sub-agents for research, file review, drafting, and quality checks. The value is coverage, not autopilot. The human still owns objective, risk tolerance, source selection, permission mode, and approval before file changes or external actions.

Room prompt

Name one deal or portco workflow where parallel streams would help: diligence, customer synthesis, memo prep, or board follow-up. Keep the action review-only in Module 1.

03

Plugins

The Plugins section starts with the admin-review question: which plugin category would be worth reviewing first? A Plugin packages capability into one install. Official Cowork docs describe Plugins as bundles of Skills, connectors, and sub-agents; installed Plugin Skills can appear through / or the plus button. Current marketplace limits worth teaching: plugin ZIP max 50 MB, 100 manual Plugins per marketplace, and 500 GitHub-synced Plugins per marketplace. For GGP, the path is governed: personal need to pilot Skill to reviewed Plugin to Protis and Jennifer governance to firm marketplace. Official financial-services plugin examples include financial analysis, investment banking, equity research, private equity, and wealth management add-ons, but those are starting points requiring GGP review, data-provider approval, and professional review before business use.

Room prompt

Name one workflow that belongs as a private pilot before it becomes a firm-supported Plugin. Do not install user-imported Plugins during the session.

04

MCP Connectors

The MCP section keeps the practical first question: which M365 surface should Cowork read first to save the most time? MCP connectors let Cowork work where context already lives. The connector does not make Claude omniscient; it gives Claude an approved path to retrieve from a system, and where supported, act in that system. Enterprise use has three checks for M365: a Claude Owner enables the connector, a Microsoft Entra admin grants tenant consent, and each user connects their own account. M365 content remains in the tenant and is retrieved on demand; the current M365 connector is read-oriented. For GGP today, start with Microsoft 365: Outlook, Teams, OneDrive, and SharePoint. Other connectors and custom remote MCP servers route through Protis and Jennifer.

Room prompt

Pick the M365 surface that would save the most time if Cowork could use it well. Name the source, the boundary, and the output you would want.

05

Scheduled Tasks

The Scheduled Tasks section starts with the runtime question: which weekly task needs an off-laptop runtime? Cowork scheduled tasks are saved Cowork prompts that can run on demand or on a cadence, created through /schedule or the Scheduled page. Each scheduled task runs as its own Cowork session and can use configured connectors, Skills, and installed Plugins. Runtime matters: Cowork scheduled tasks run only while the computer is awake and Claude Desktop is open. If the computer is asleep or the app is closed, Cowork skips the scheduled run, then runs it when Desktop is back and shows a notification. If work must run while the laptop is closed, call it platform automation, Claude Code /schedule, or another non-Cowork runtime, not a Cowork scheduled task.

Room prompt

Pick a weekly task that currently lives in your head. Name the runtime before naming the workflow: Cowork scheduled task, Claude Code /schedule, platform automation, or /loop.

06

Instructions and Memory

The Instructions and Memory section opens with the scope question: what belongs firm-wide, what belongs to you, and what belongs in one Project? The point is scope control. Organization instructions are firm-wide, admin-owned, and capped at 3,000 characters. Global instructions are personal defaults. Project instructions add Project-specific context. Project memory helps Cowork remember useful context inside a Project, but it is scoped to that Project and stored locally with the desktop experience. Memory is supported within Projects, not retained across standalone Cowork sessions. It is not firm truth, not team-shared, and not a place for Tier 3 data. Enterprise caveat: OpenTelemetry can monitor Cowork activity, but the Compliance API does not capture it today.

Room prompt

Put one example in the right layer: firm rule, personal style, engagement context, or durable company truth.