
A team's first week with Claude Cowork works best as five short days. On Day 1 everyone sets up the app, a working folder, the connectors their work runs on and their global instructions. On Day 2 each person runs one real task with Manual approval on. On Day 3 the team sets up a project for each stream of work and names an owner for each workflow. On Day 4 the work people repeat becomes skills the whole team uses. On Day 5 the team reviews what worked and decides what comes next.
Claude Cowork is Anthropic's agent for knowledge work. It runs on the paid Claude plans (Pro, Max, Ultra, Team and Enterprise), reads and writes only in the folders you connect, and, according to Anthropic's guide to getting started with Claude Cowork, "uses the same agentic architecture that powers Claude Code, with no terminal required". Anthropic's tutorial on scaling Cowork across an organization warns that "IT flipping the switch and sending an email isn't a rollout" and its advice for launch day is "Don't hand people a blank page."
On Team and Enterprise plans, Claude Cowork is on by default, and an organization owner can turn it off for everyone. On an Enterprise plan, owners can also use groups and custom roles to turn it on for specific teams, and organizations that have enabled HIPAA start with it off. Anthropic's guide to Cowork on Team and Enterprise plans puts plugins under the same admin switch and lets owners create plugin marketplaces that hand curated plugins to the whole organization.
For an Enterprise rollout, Anthropic's admin guide sets the order: complete your security review, start in a sandbox organization or with a pilot group through a role, and review your network egress, connector, sharing and remote control settings before you turn it on. It then advises you to "Monitor usage and gather feedback for two to four weeks, then expand in waves." The first week opens that window. Pick a pilot group that mixes keen and doubtful people, since Anthropic's tutorial warns that "an all-enthusiastic pilot gives false positives".
Each person installs the Claude Desktop app, signs in and selects Cowork in the message box. If Cowork is missing, an organization owner has turned it off, the organization has enabled HIPAA and nobody has turned Cowork on, or, on an Enterprise plan, the person's group has not been given it. Claude Cowork sees no files until you connect a folder, so connect one dedicated working folder rather than granting broad access, and keep files with sensitive information, such as financial documents, out of it, as Anthropic's guide to using Claude Cowork safely advises.
Next, authenticate the connectors the team uses most, such as Slack, Gmail, Drive, Asana or GitHub, and only connectors from sources you trust. The day-one checklist for a department in Anthropic's rollout tutorial is "authenticate 3 connectors, install the department plugin, run its flagship skill".
Then each person writes global instructions, the standing brief Claude reads in every session: who you are and what you do, the shorthand and acronyms your team uses, and how you like output delivered (format, length and tone). Anthropic's Claude Academy lesson on standing context adds that "Your global instructions don't need to be complete on day one."
Last, agree how much Claude may do without asking. Cowork has three permission modes, and you can change the mode at any time from the mode selector in the chat box. Anthropic's advice is to match your oversight to the stakes.
| Manual | Auto | Skip | |
|---|---|---|---|
| What Claude does | pauses and asks for approval for actions, and you choose Allow or Deny | keeps working without stopping to ask about every step, and blocks anything it determines to be unsafe | doesn't pause to ask, and nothing checks its actions automatically |
| Default | the default on Enterprise plans | rolling out as the default on Pro, Max, Ultra and Team plans | the default on no plan |
| Use it when | work with real consequences, such as money, messages sent as you or important files | you want the speed of letting Claude work without interruptions, knowing it can miss risky actions | you completely trust every action, connector, file and app involved |
For a new team we recommend Manual for the whole first week, which on a Team plan means switching from the default on purpose. Anthropic's safety guidance says to "Start with lower-stakes tasks and build trust gradually, like you would with a new colleague." In every mode, Claude asks for your explicit permission before it permanently deletes any file. Write the Day 1 choices down where the team can find them: the working folder, the connectors allowed, the files that stay out of reach and the mode for the week.
On Day 2 each person runs one real task. Anthropic's Claude Cowork product guide recommends "one real task you were about to do yourself" over a demo, such as a one-page brief from a folder of PDFs, a status update from one Slack channel, or a cleaned spreadsheet with a chart.
Write the request the way the guide's best practices describe. Name the folder, the channel and the time window, and say what the output is: its format, length, audience and tone. "A good Claude Cowork prompt front-loads context; a bad one makes Claude Cowork guess." When the first draft is close, tell Claude what to change in the same task, since it remembers the conversation and edits faster than it starts again. Read the result before you send, publish or act on it, especially anything with numbers, names or citations.
Day 3 moves from one person's tasks to the team's streams of work. A project in Cowork is a workspace tied to one stream of work, with its own files, links, instructions and memory, and every conversation in the project can reach its folders and links. Set up one project for each stream the team runs, such as a client account or a monthly report, and write its project instructions: the scope, the output formats and when Claude should stop and hand the task back to a person. According to Anthropic's guide to projects in Claude Cowork, Team and Enterprise plans let you share a project with people in your organization, with "Can view" or "Can edit" access, and Claude can remember context from the tasks run in a project and apply it to later tasks in the same project.
Then ask each person to name two or three tasks they repeat every week that pull files, data or documents from several tools, and give each workflow one owner. An owner keeps each workflow running after the first week, and the list becomes Day 4's starting point.
Some of those tasks belong elsewhere. Anthropic's best-practice guide says Claude Cowork "is best for file-and-context work across tools", that production code across a repository belongs in Claude Code, and that chat is "lighter weight" for quick factual questions; Cowork itself offers to launch a Claude Code session when it detects a coding task. A task that runs the same steps every time on a trigger suits an n8n workflow better, which is how we build that kind of automation.
On Day 4 the team turns what it repeated into standing context. Look back over the week for the corrections each person kept making, which the Claude Academy lesson calls "global-instruction candidates", and add them to global instructions. Then pick the deliverable the team makes again and again and build a skill for it. The product guide describes skills as "pre-written playbooks for common deliverables and team workflows" that "materially improve output quality", and advises starting with one skill for the team's most repetitive task. Anthropic's rollout tutorial adds that "Skills should never contain credentials."
A plugin comes later, when one skill is no longer enough, such as for a task that moves through several tools in sequence or a workflow you re-explain every session. According to the guide to extending Claude Cowork with plugins, a plugin bundles skills, connectors and sub-agents into one package, and on Team and Enterprise plans owners can hand it to the whole organization through a plugin marketplace. A one-off question needs neither, so run it as a plain task.
Day 5 is a team review. Each person shows the task that worked best and the instructions or skill behind it, and the team agrees which skills become shared. Cowork sessions can't be shared with others, so the review is where one person's working setup reaches the rest of the team. Anthropic's tutorial suggests a habit that carries on after the week: "Team posts one skill in the team channel with a short video of it running on real work."
Then decide what to measure. The same tutorial advises moving measurement from activity, such as weekly active users and sessions, to outcomes, such as time saved and the return on each use case. Owners on Team and Enterprise plans can stream Cowork events to their own monitoring tools through OpenTelemetry, which Anthropic says doesn't replace audit logging for compliance purposes. If anyone set up a scheduled task during the week, review its output after each run and pause it while nobody uses it. Then, according to Anthropic's admin guide, keep the pilot group going for two to four weeks before you expand in waves.
We build on Claude Cowork for teams that want it working on their own systems. We connect it to the systems your teams already use, such as Jira and your CRM, and build the connectors your systems need when none exist yet. For knowledge work we build a pipeline that gives Cowork the right documents from your internal repositories. We set which actions wait for approval and when Cowork hands a task to a person, and every team is trained on its workflow. An AI pilot takes 2 to 6 weeks, and scoping sets the timeline for a full build.
Claude Cowork is Anthropic's agent for knowledge work. It runs on the paid Claude plans (Pro, Max, Ultra, Team and Enterprise), works in the cloud and through the Claude Desktop app in the folders you connect, and uses the connectors, skills and plugins you add. Anthropic's Help Center says it "uses the same agentic architecture that powers Claude Code, with no terminal required".
No. On Team and Enterprise plans Cowork is on by default, and an organization owner can turn it off for everyone. On an Enterprise plan, owners can also use groups and custom roles to turn it on for specific teams. Organizations that have enabled HIPAA start with Cowork off, so an owner turns it on there.
We recommend Manual for the first week, so Claude pauses and asks for approval for actions and each person chooses Allow or Deny. Auto is rolling out as the default on Pro, Max, Ultra and Team plans, so on those plans someone switches modes on purpose. Anthropic advises starting with lower-stakes tasks and building trust gradually, like you would with a new colleague.
Global instructions are the standing brief Claude reads in every session: your role, your team's shorthand and how you like output delivered. A project is a workspace for one stream of work, with its own files, links, instructions and memory. A skill is a pre-written playbook for a deliverable or workflow the team repeats, and a plugin bundles skills, connectors and sub-agents into one package.
Claude Code, for production code. Anthropic's product guide says Cowork is best for file-and-context work across tools and that production code across a repository belongs in Claude Code, and Cowork itself offers to launch a Claude Code session when it detects a coding task. Chat stays the lighter choice for quick factual questions.
Two to four weeks, according to Anthropic's admin guide for Enterprise organizations. It recommends starting in a sandbox organization or with a pilot group through a role, monitoring usage and gathering feedback for two to four weeks, and then expanding in waves. Use the first week to set the habits and the weeks after it to see whether the team keeps them.
Rolling Claude Cowork out to a team?
Book a discovery call. A senior partner runs the call and tells you which of your workflows suit Cowork before anything is scoped.