Claude Cowork Just Got Projects, and It Changes How You Use It

Claude Cowork has had a problem since it launched in January: every task starts from zero. You open a new session, explain what you’re working on, point Claude at the right files, describe your preferences, and then finally get to the actual work. Do it again tomorrow and you’re repeating the whole setup. Projects fix that.

Open post on X

Anthropic rolled out Projects for Cowork on Desktop on March 20, 2026. The feature lets you group tasks, files, instructions, and memory into persistent workspaces tied to local folders on your machine. Instead of treating every Cowork session as a blank slate, you can now pick up where you left off with all your context intact.

What Projects Actually Are

A project in Cowork is a container for related work. Think of it like a workspace folder with a brain attached. Each project holds four things:

A local folder. This is where your files live. Claude can read from and write to this folder during tasks. Your financial reports, research docs, marketing assets, or whatever else you’re working with stay on your computer, not in the cloud.

Instructions. These are standing rules that apply to every task you run inside the project. You might tell Claude to use a formal tone for client deliverables, follow a specific template for weekly reports, or always cite sources when doing research. Set them once and they stick.

Memory. Claude remembers context from previous tasks within a project and applies it to future ones. If you told Claude three weeks ago that your Q1 targets changed, it remembers that the next time you ask for a progress report. Memory is scoped to each project. What Claude learns in your “Marketing” project doesn’t leak into your “Finance” project.

Scheduled tasks. Recurring work that runs on a cadence you set. Daily briefings, weekly report generation, regular file cleanup. These are project-specific, so each project can have its own schedule.

Three Ways to Create a Project

When you click “Work in a project” in Cowork, you get three setup options:

Start from scratch. Creates a new folder with custom instructions and files. Best for kicking off new initiatives where you want full control over the structure.

Import from chat. Pulls instructions and files from an existing Claude chat project into a Cowork project. If you’ve been using Claude’s web-based Projects feature and want to bring that context into Cowork’s agentic workflow, this is how. One note: bulk imports aren’t supported yet, so you bring over projects one at a time.

Use an existing folder. Point Claude at a folder already on your machine. This is the fastest path if you have an organized file structure. Claude indexes the contents and you add instructions on top.

How to Get the Most Out of Projects

After spending time with the feature and reading what early users are reporting, a few patterns stand out.

Build projects around recurring responsibilities, not one-off tasks

The real value of Projects shows up over time. A project for “Q1 Board Deck” that you use once is fine, but a project for “Monthly Financial Reporting” that you use every month for the rest of the year is where the memory and instructions compound. Claude gets better at each cycle because it remembers what worked before.

Good project candidates:

Weekly team status reports
Monthly financial reconciliation
Ongoing content calendars
Client account management
Research pipelines that build on previous findings

Write instructions like you’re onboarding a new hire

The instructions field is the single most high-leverage thing you can set. Treat it like an onboarding doc for a competent colleague who doesn’t know your preferences.

Bad instruction: “Write well.”

Better instruction: “Use AP style for all written content. Keep paragraphs under 4 sentences. When citing data, include the source and date. Format deliverables as Google Docs-compatible documents with headers, not walls of text.”

The more specific your instructions, the less time you spend correcting outputs.

Pair scheduled tasks with project memory

Scheduled tasks and memory work together in a way that isn’t obvious at first. When a scheduled task runs inside a project, it has access to everything Claude has learned from previous tasks in that project.

Say you set up a daily briefing that scans your inbox, calendar, and task manager. On day one, it gives you a generic summary. But over time, as Claude learns which meetings matter most to you, which senders you always want flagged, and what format you prefer, the briefings get sharper without you changing the prompt.

To set up a scheduled task, type /schedule inside any Cowork task, or click “Scheduled” in the left sidebar. You pick the frequency (hourly, daily, weekly, weekdays), the model, and the working folder. One caveat: scheduled tasks only run while your computer is awake and the Claude Desktop app is open. If your machine is asleep at the scheduled time, Cowork skips the run and catches up when you’re back.

Keep project scope tight

It’s tempting to make a mega-project that covers everything you do. Resist that. Memory is scoped to each project, so a project that mixes client work with internal ops will have Claude applying context from one domain to the other. That gets confusing fast.

Better to have five focused projects than one sprawling one. You can always create new projects cheaply, and the sidebar keeps them organized.

What This Looks Like in Practice

Here are a few concrete workflows people are running with Projects.

Financial reconciliation. Drop bank statements and invoices into the project folder monthly. Claude reads them, categorizes expenses, matches invoices to transactions, flags discrepancies, and outputs a formatted spreadsheet. The instructions specify your chart of accounts and preferred formatting. Memory means Claude remembers your vendor naming conventions from last month.

Content production pipeline. A project folder with your brand guidelines, tone of voice doc, and content calendar. Claude drafts blog posts, social copy, and email newsletters following your rules. Scheduled tasks pull trending topics daily. Over time, Claude’s outputs match your voice more closely because memory retains feedback from previous rounds.

Research synthesis. Point a project at a folder where you dump PDFs, articles, and notes. Claude reads everything, identifies themes, and produces a structured summary. Instructions tell it what angle you care about. Each week’s research builds on the last because memory carries forward.

Meeting prep. A project connected to your calendar and email. A scheduled task runs 30 minutes before each meeting, pulling relevant documents, summarizing recent email threads with attendees, and creating a one-page brief. Instructions specify how you like the brief formatted.

What Projects Don’t Do (Yet)

A few limitations worth knowing:

No cloud sync. Projects are desktop-only and stored locally on your machine. If you use Claude on two computers, your projects don’t follow you. Anthropic has mentioned cross-device synchronization as a future possibility but hasn’t committed to a timeline.

No sharing. Team and Enterprise users can’t share Cowork projects with colleagues. Each person’s projects are private to their machine.

No Claude Code integration. Projects exist only in Cowork for now. If you use Claude Code for development work, those are separate workflows. Anthropic says Claude Code support is planned.

Desktop app required. You can’t access Cowork projects from claude.ai in a browser. The Desktop app has to be installed and running.

Archiving is metadata-only. When you archive a project, the sidebar entry disappears but your local files stay untouched. The project’s configuration (instructions, memory, scheduled tasks) is what gets removed from the UI.

Why This Matters

Cowork launched in January 2026 as a research preview, positioned as “Claude Code for the rest of your work.” The pitch was that knowledge workers could delegate multi-step tasks the way developers delegate coding tasks. The problem was that every session was ephemeral. You couldn’t build on previous work without re-explaining everything.

Projects change Cowork from a tool you use into a tool that knows your work. The combination of persistent files, standing instructions, scoped memory, and recurring schedules means Claude’s effectiveness improves over time within each project. That’s a fundamentally different value proposition than a chat window that forgets you exist between sessions.

The timing is interesting too. This lands one day before Anthropic shipped cloud-based scheduled tasks for Claude Code, which removes the need for your local machine to be running. The two features serve different audiences (knowledge workers vs. developers) but follow the same direction: making Claude a persistent collaborator rather than a stateless tool you prompt from scratch each time.

Requirements

Claude Desktop for macOS or Windows (x64 only; Windows ARM64 not supported)
Paid subscription: Pro ($20/mo), Max ($100/mo or $200/mo), Team ($30/user/mo), or Enterprise
Latest version of the Claude Desktop app

Cowork is still labeled a research preview, so expect rough edges. Anthropic notes that agent safety is still in development and that Cowork conversations aren’t captured in Audit Logs or the Compliance API, which matters for regulated industries.

Getting Started

Open Claude Desktop, switch to the Cowork tab, and click the “Work in a project” dropdown next to the chat input. Pick one of the three creation methods, name your project, add instructions, and start assigning tasks. The whole setup takes about two minutes.

If you’ve been using Cowork for one-off tasks and wondering why it doesn’t feel sticky, Projects are the missing piece. The tool was always capable. Now it can remember.

Sources: