Thirteen ideas, one per screen. Each is a thing to point at while you explain. Cowork is Claude doing the work: in your files, your apps, your browser. Take the silences.
Four moves: meet the tool, make it yours, run it everywhere you work, then share what works and keep it safe.
Three modes of Claude on the desktop, built for three shapes of work. The shift that matters: you hand Cowork an outcome, not a question.
Chat is where you go to think out loud. Cowork is for the work you'd normally do yourself: describe an outcome and Claude plans, executes, and delivers the whole thing: context-gathering, analysis, file production, tool use.
Three setup choices: the working folder Claude reads and writes, the connectors that reach your apps, and how closely it asks before acting.
Install the desktop app, sign in, pick Cowork. The folder is the single most important choice: pick the smallest one that holds what the task needs. Cloud connectors are often read-only; your local folder is where Claude builds and iterates.
When a task is multi-step, lives in real files, or spans your tools, give it to Cowork. The patterns tell you what to delegate.
The through-line: work that used to mean juggling steps, files and tools in your head can now be handed off as a single task. Scheduled tasks and Dispatch only run when your computer is on and Claude is running.
A good prompt names the deliverable, the inputs, and the nuance Claude can't guess. Then you answer its questions, steer if it drifts, and review the result.
More of the back-and-forth happens upfront, before Claude starts, instead of across five rounds of "actually, can you also…". The deliverable is the artifact, not the chat. Check that it met the objective and that nothing's made up.
Each one gives Claude a different kind of knowledge to carry forward, so the context you supply once keeps paying off.
You won't start with all four on day one. Set global instructions first, add a project as recurring work surfaces, build a skill when you're tired of re-explaining a process, and bundle a plugin once you have something worth sharing.
Global instructions cover you, in every session. A project covers one stream of work, and quietly builds its own memory as you go.
That last one is the difference projects make. Outside a project each session starts fresh; inside one, every conversation adds to what Claude knows, so the next task opens with the client's situation and last week's decisions already in hand.
Key conceptChat memory builds itself across all chats; Cowork carries context from what you set up (global + project), and project memory is scoped to that project. → teacher notes (incl. why scheduled tasks wait when the computer is asleep).
A folder of files that teaches Claude how to do one kind of work the way you'd want it done. Up to four parts; only the first is required.
Instructions tell Claude what to do; assets give it the raw materials; references show it what "good" looks like; scripts run the repeatable parts the same way every time. Include what the work needs, nothing more.
Where a skill is one playbook, a plugin is several, plus the connectors and subagents they need. It comes in two shapes.
Anthropic publishes plugins for common roles: finance, legal, sales, support. Install one off the shelf, customise it to your templates, or build your own. The expertise travels with the install, not the person. Start with /setup-cowork to find the right one.
Dashboards, vendor portals, web apps behind a login. Claude reads and acts on the page itself, then hands the results back to Cowork.
You stay signed in: Claude works in the session you've already authenticated, it can't log in for you. One conversation, both surfaces. (Claude in Chrome may not be available on some enterprise plans.)
Claude shows up as an add-in in Word, Excel, PowerPoint and Outlook. One conversation carries context from one app to the next.
Reach for Cowork when work pulls from many sources and ends in a deliverable; reach for Claude in M365 when you're editing the Office file in place. The two hand off cleanly, and the file is the file in both.
Cowork always asks before deleting. That's the floor. These are the pre-emptive moves you bring on top.
Most mistakes don't happen because a safeguard failed. They happen because someone clicked through a confirmation that wasn't quite the action they intended. The dialog exists because the action matters; treat it that way.
Before a skill leaves your desk, see it side by side against the baseline. No code, no test scripts, just your judgment.
Built into skill-creator. The bar for shipping isn't perfect evals; it's that the cases you care about pass meaningfully better than the baseline, and you've named the ones you don't yet handle. Change one thing at a time.
A skill that started as one person's way of working becomes something a teammate can install, through your org's private marketplace.
Teammates see it in their Directory, labeled as coming from your company; your updates flow to them automatically. Keep it healthy: one named owner, evals before every publish, name skills specifically, review quarterly.
Cowork goes beyond chat to let you delegate real work. Make it yours with standing context, skills and plugins; run it across your browser and Office; then share what works and keep it safe.
Key conceptReview Cowork's output like a colleague's draft: check the facts (ask which sources it used, then verify) before you send or act. Claude can prepare; you ship. → teacher notes.
tutto.one/praxis · daniel@tutto.one