Introduction to Claude Cowork: teaching notes

"Thirteen ideas, one screen each. The whole course turns on one move: you stop typing questions and start handing over work. Chat is where you think out loud. Cowork is where you give Claude a job and it goes and does it, in your files, your apps, your browser. I'll show you, then I'll stop and ask where it fits your week."

"Here's the whole thing on one screen. We meet Cowork and hand it a first task. We make it yours so it stops needing re-explaining. We run it in your browser and your Office apps. Then we share what works across the team, safely. If you lose the thread, this is the map back."

"Three shapes of the same Claude. Chat is turn-by-turn, where you go to think out loud and draft. Cowork is different in kind, not degree: you describe a finished thing and Claude gathers the context, does the steps, uses your tools, and saves the deliverable back. Code is for developers building software. The instinct from chatbots is one question, one answer, then you stitch the rest together yourself. Cowork takes the whole stitch off your plate." (The deck's caption: Chat is to think out loud, Cowork is for the work you'd normally do yourself.)

"You install the desktop app, sign in, pick Cowork, then make one decision that matters more than any other: which folder. That folder is the boundary, the room Claude works in. Pick the smallest one that holds the task, not your whole Documents drive. Connectors reach into your email, Drive and the rest, often read-only. And you choose how it asks: ask-before-acting is the default, and it always asks before deleting, no exceptions." (The deck's caption: cloud connectors are often read-only; your local folder is where Claude builds and iterates.)

"How do you know a task is right for Cowork? Three patterns. It takes several steps, gather, compare, research, draft, format. It lives in real files and ends in a real file. Or it spans the tools you already use. Hit one or more and it's a handoff. And you don't have to be at your desk: you can schedule it to run weekly, or dispatch it from your phone, as long as your machine is on and Claude is running." (The deck's caption: the through-line is work you used to juggle in your head, handed off as a single task.)

"Brief it the way you'd brief a capable colleague who's about to disappear into a task and come back with it done. Three things: name the deliverable, a one-page brief, a slide, a ranked list; name the inputs, which folder, which channel, which date range; and name the nuance only you would know. It'll ask a question or two to close the gaps, then go. If it drifts, jump in, the cost of a redirect is low. Then you review the file, not the chat." (framings.md: treat Claude like a junior on their first day; spell it out, the mistake is being too economical with words.)

"Right now you re-explain yourself every time. These four stop that. Global instructions: who you are and how you work, set once. Projects: one stream of work, so Claude acts like it's already on the team. Skills: how a specific process should be done, your way. Plugins: your whole role's expertise, bundled. The more it knows, the more it takes off your plate, and the context you supply once keeps paying off." (Previews framings.md: the folder is the memory; the context stops living in a chat you lose.)

"Two kinds of standing context. Global instructions are a brief about you, set once in Settings: who you are, your shorthand, how you like output. They apply to every session. A project is scoped to one stream of work, a client, a recurring report, a launch, and it holds instructions, the folders it works from, and scheduled runs. The difference projects make is the last part: memory. Inside a project, every conversation adds to what Claude knows. Outside one, each session starts fresh." (framings.md: context window; here the project is the memory, not a chat that forgets the beginning.)

"A skill is a playbook. A folder where you write down how one kind of work should be done, once, so Claude does it your way every time. The brief, SKILL.md, is the only required part: what it does, when, and how, written like a runbook for a new colleague. Add templates as assets, your best past work as references, and a script for the part that must come out identical every time. Claude loads it on its own when the task matches. And you build it by talking: 'I want a skill for the thing I keep re-explaining, walk me through what you need.'" (framings.md: skill file as institutional memory: the rules that lived in your best person's head, pulled out onto disk so a new joiner produces the same work.)

"A plugin is a package built around a job. Two shapes. A pipeline: several skills that run in sequence, like pull the actuals, build the variance, draft the board memo. Or a toolkit: the skills a function reaches for most, bundled so a new teammate installs one thing and has the whole set. The point is the expertise travels with the install, not the person. Anthropic publishes them for finance, legal, sales, support; start with /setup-cowork and it finds the one that fits your work." (framings.md: skill file as institutional memory, scaled: a plugin is your team's way of working, installable.)

"Some of your tools have no connector: the BI dashboard, the procurement portal, the CRM behind single sign-on. Claude in Chrome is the bridge. It reads and clicks the page in the session you're already signed into, pulls what you need, and hands it back to Cowork to build the deliverable. One conversation across both. The line that matters: it can't log in for you, you stay signed in, and it asks before sensitive actions." (The deck's caption: you stay signed in; Claude in Chrome may not be available on some enterprise plans.)

"The same Claude turns up inside the Office apps as an add-in. In Excel it writes formulas and variance commentary and debugs your #REF errors. In PowerPoint it reads your slide master and builds native, editable charts. In Word it drafts and reformats in place with tracked changes. In Outlook it triages and drafts replies with the context of the thread. And one conversation carries the context across, the Excel model becomes the deck becomes the email. Reach for Cowork to build from many sources; reach for M365 when you're editing the file itself." (The deck's caption: the two hand off cleanly, and the file is the file in both.)

"An autonomous tool with file and app access needs guardrails, and most of them are yours to bring. Three layers. Set up: a dedicated folder, not a catch-all; back up anything irreplaceable; test new tasks on copies. Write the prompt: name destructive verbs precisely, name the bounds, 'only Q3 files', keep it draft-only until you trust it. And in the moment: read the plan before it runs, watch for scope creep, and when a confirmation appears, read it. Most mistakes aren't a safeguard failing, they're someone clicking through a dialog that wasn't quite the action they meant." (The deck's caption: the dialog exists because the action matters, treat it that way.)

"An eval is just a try-out. Don't let the word scare you. skill-creator runs the same realistic request two ways, once with your skill, once without, and shows you the pair. You read both and answer two questions: would I send the skill version, and if not, what's off? You say it in plain English, 'too formal', 'skipped the summary', and Claude revises. It's a loop, not a one-time gate. You ship when the cases you care about beat the baseline and you've named the edge cases you don't handle yet." (The deck's caption: change one thing at a time; the bar is meaningfully better than baseline, not perfect.)

"The payoff: what started on your laptop becomes something the whole team installs. Inside a company, that runs through a private marketplace an admin manages. You hand your plugin to whoever owns it, and they publish it at a level, available to opt into, installed by default, or required for things everyone must run the same way. Teammates see it in their Directory, labelled as coming from your company, and your updates flow to them automatically. Keep it alive with four habits: one named owner, evals before every publish, specific names, and a quarterly review to retire what nobody runs." (framings.md: skill file as institutional memory; this is the moment one person's know-how becomes the firm's.)

"If you keep one thing: hand off the work, don't just chat about it. Teach Cowork your way once, with global instructions, a project, a skill, and the next deliverable comes back better and faster. Trust it to do the doing, but review the result like a colleague's draft, because you own what goes out. Everything we saw today, the browser, the Office apps, the marketplace, is a way to do that in more places."