How I actually use Claude Cowork (and where it falls apart)

Most “AI agent” demos are a person typing one prompt, getting a dashboard that looks like a SaaS Series A, and saying “and you don’t even need to code.” Cool. I’ve watched a hundred of those. Almost none of them tell you what it’s like on day forty, when the novelty is gone and you’re just trying to get work done.

So this is the version I wish someone had shown me. What Cowork is, how I set it up, the workflows I actually keep, and the parts that still annoy me.

[KENT: one or two sentences on why you bothered with Cowork at all. What were you doing before it that sucked? The honest hook. Something like “I already live in Claude Code all day, so why did I need another tab?”]

What Cowork even is

Cowork is a local AI agent. Translation: it’s an AI that lives on your actual computer and can do things with your files, your apps, and your connected tools, on its own, instead of just talking back at you in a chat window.

If you’ve used something like OpenClaude (or whatever the open-source flavor of the month is called), think of Cowork as Anthropic’s take. More locked down, built to stay inside the Claude ecosystem, and pointed squarely at people who don’t write code for a living.

The thing that clicks once you use it: a regular chatbot can tell you how to organize your desktop. Cowork opens the folder and does it. That gap is the entire product.

Download it, install it, and you get a local hub with three tabs:

• Chat is the same Claude you already use on the web, history and all.
• Cowork is the local agent. This is the one we care about.
• Claude Code is the coding agent.

[KENT: you live in Claude Code already. Worth a parenthetical on how Cowork and Code feel different in your hands, because most readers won’t have that frame. You’re one of the few people who can actually compare them straight.]

Setup, the boring part that matters

Before doing anything fun, spend five minutes in settings. Capabilities, then turn on what you’re comfortable with. I run [KENT: how much do you actually leave on? You’re more security-paranoid than the average YouTuber, lean into that]. Under visuals, make sure artifacts and inline visualizations are on, or half the cool output won’t render.

Two settings worth calling out:

Global instructions. These apply to every Cowork session. Mine is a short brief: who I am, how to work with me, how to build things, and a list of things to never do. If you’ve ever written a CLAUDE.md file, it’s the same energy.

[KENT: paste a sanitized version of YOUR global instructions here, or at least describe what’s in them. This is gold for readers and nobody else can fake it. Your “never do this” list especially.]

If you don’t have one yet, here’s the prompt I use to generate the first draft. Paste it into any Claude chat. The first line is the whole point: it tells the model to use what it already knows about you instead of making you fill out a form.

I want to create a custom assistant settings document I can paste into my AI
tool (Claude Cowork global instructions, a custom GPT system prompt, or
similar).

First, assess what you already know about me from our history. If you have
enough to write something specific and opinionated, draft it. If there are
gaps that would make the output generic, ask me ONLY the questions you
actually need answered. Don't interview me for things you already know.

Cover: how I like to work, my communication style and tone, what I'm building
or managing, how I want you to handle ambiguity and pushback, what kinds of
decisions or actions need my explicit approval before you proceed, and any
recurring workflows or context you should know upfront.

Output a formatted settings doc I can paste in directly. Specific to me, not
boilerplate. Keep it condensed. This gets loaded on every session, so every
line has to earn its place.

Dispatch. Lets you poke Cowork from your phone and have it do things on your computer remotely. [KENT: do you actually use this, or is it a party trick? Be honest. If you don’t use it, say so, that’s more useful than pretending.]

Then there’s Claude in Chrome, which lets the agent go drive a browser and visit sites. [KENT: your take, one line. Actually useful or mostly a demo?]

Level 1: point it at a folder

The fastest way to get what Cowork is for: give it a messy folder and tell it to clean up.

You choose a folder, allow access, and ask. “Help me organize these files.” It scans, proposes a structure, and asks before it nukes anything. The first time I did this it found [KENT: what did it actually find on your machine? The YouTuber found an exposed API key sitting on his desktop. You almost certainly have your own embarrassing version. Real specifics here, even if mildly humiliating, are the whole point].

This is also the cleanest way to feel the difference between Cowork and chat Claude. The web version cannot touch your machine. Cowork assigns itself a task, shows a progress pane on the right, and goes.

A second workflow I keep coming back to: point it at a folder of statements or exports and ask for an interactive dashboard. [KENT: the transcript used 24 months of credit card statements. Do you have a real version of this? Lunch Money exports, client reporting, SEO data dumps? Pick the one that’s actually yours. Bonus: you can note that chat Claude caps uploads at 20 files, and Cowork doesn’t, because it’s reading off disk.]

One thing the demos gloss over: it does fail sometimes. File permission errors, a delete that doesn’t go through, a download it can’t finish. [KENT: did you hit any of these? “I broke production three times learning this” energy fits perfectly here. Show a real stumble.]

Level 2: skills, connectors, plugins

This is where Cowork stops being a neat trick and starts being a system.

Skills are reusable instructions. Say you build a brand book, or a formatting standard, or a way you always want reports structured. You turn it into a skill once, and from then on you just say “apply [the skill]” and it does. You can keep them local or share them. [KENT: do you have a real skill you’ve built? You’ve got a whole writing voice system, a humanization playbook, SEO brief templates. Any of those would make a far better example than “I made a brand book.” Show one of yours.]

Connectors plug Cowork into third-party software. Gmail, Drive, Calendar, the usual suspects, plus a long tail of everything else, and you can wire up your own through MCP if it’s not on the list. [KENT: which connectors do you actually run, and for what? You’ve got Ahrefs, ClickUp, Fantastical, Gmail, calendar stuff. Name the two or three that earn their keep. Anonymize client specifics per your rules.]

Plugins are bundles of skills and connectors aimed at a job. There’s a finance one Anthropic ships that turns a folder of statements into actual financial statements with a slash command. For anyone running a business, that’s the kind of thing that quietly saves a weekend. [KENT: GiggTech is a real business. Is there a plugin you’d actually use for it, or one you wish existed? Your honest reaction matters more than the feature list.]

Scheduled tasks, the quiet workhorse

A lot of work is just the same work on a timer. Cowork has a scheduled tab for exactly this.

The example everyone reaches for is a morning brief: every day at 7am, pull my calendar, skim my email, and drop a short brief somewhere I’ll see it. [KENT: you have an entire daily planning system already, /morning and /end-of-day and the rest. Does Cowork scheduling replace any of that, complement it, or is it redundant with what you built in Claude Code? This is a comparison only you can make, and it’s the interesting part. Don’t skip it.]

[KENT: where do you send your outputs? The transcript used Apple Notes so it shows up on the phone. You might use something different. Say what and why.]

Projects: the part that’s actually worth your time

Everything above is useful. Projects are where it gets good.

A project is a persistent workspace wrapped around a folder, with its own instructions, memory, files, and connections. The key word is persistent. Cowork remembers context across every session tied to that project, so you’re not re-explaining yourself every morning like it has amnesia.

Here’s how you set one up. Make a folder. In Cowork, new project, point it at that folder. Write the instructions, who the agent is to you and how it should behave. Then create.

The instruction prompt is where personality lives. Something like “you are my [role]. Be thorough, double-check numbers, push back when something doesn’t match how I actually work, be direct.” [KENT: what do your real project instructions sound like? You’re opinionated about pushback and bluntness. Quote yourself.]

Then the move that makes the whole thing sing: install the productivity plugin and run its start command, even if you don’t think you need task management. The reason isn’t the tasks. It’s that it sets up a memory system for you, CLAUDE.md plus a memory directory, instead of you hand-maintaining markdown files like a caveman. You drop in context, it writes it to memory, and from then on it remembers. Type the update command and it fills in whatever gaps it’s noticed.

[KENT: this is your home turf. You ALREADY hand-maintain a markdown brain (the entire kb repo). So you’re the perfect person to say whether the productivity plugin’s memory system is better than rolling your own, or whether you trust your own setup more. Be critical. This is the most credible thing you could possibly write in this post.]

[KENT: describe one real project you’d run, or already run. The transcript guy had four projects spanning his whole life. You have Fishbowl, GiggTech clients, writing, life. Pick one, show what’s in the folder, show what the agent does with it. Anonymize the day job and client specifics per your sensitivity rules.]

Once memory and instructions are in place, you build whatever you want on top: dashboards, digests, research workflows, all pulling from the connectors and skills you already set up. A “mission control” page per project, basically, so you can see everything in one spot.

[KENT: if you want a concrete one to hand readers, the companion prompt builds an SEO mission control dashboard with this exact three-layer setup, link it: grab the build prompt here. Swap the link target once you know the slug.] [KENT: you literally built Arc for this exact reason. Link it: see how I built Arc. Then the honest question, does Cowork’s project dashboard make Arc redundant, or is your hand-built thing still better? Don’t pull the punch.]

The combo nobody talks about: Cowork plus Claude Code

Here’s the part that puts you in the weird-power-user tier, and the part that needs a caveat: yes, this one assumes you can code a little.

Cowork is great at workflows and dashboards. But it is not a specialized coding agent, and the second your project gets real, you feel it. That’s what Claude Code is for. Switch to the code tab, point it at the same project folder, and now you’ve got the better coding tool working on the thing Cowork set up.

[KENT: you are squarely in the 0.1% here. You run Claude Code daily, you build in Astro and Next and Vue, you deploy to Vercel. The transcript hand-waves this section because the YouTuber can’t actually do it. You can. This is your single biggest opportunity to add value the source never could. Tell a real story of Cowork handing off to Code, or Code building on a Cowork project.]

So is it worth it?

[KENT: your verdict. The whole post earns its keep here. Some honest angles to pick from:

• Who is this actually for? (You’d argue non-coders get more net-new value than you do, since you already had Code.)
• What does it replace, what does it duplicate, what does it not touch?
• The one workflow you’d keep if you could only keep one.
• The thing that still annoys you. Every tool has one. Name it.
• Skeptical-of-hype but enthusiastic-about-what-works is your default register. Land there.]

[KENT: close with a real opinion, not a summary. Your sign-off energy, not “in conclusion.” Maybe what you’re going to build next with it.]