claude cowork was launched 3 months ago, since then claude has shipped 50+ updates which has completely changed how I utilise cowork, here's how you can take advantage and master it:
claude cowork has changed how we do work. you hand it a task. it builds a plan. it executes. it delivers finished work into your folders while you do something else.
now a lot has changed since cowork was released, this guide will cover everything. so listen...
if you want to delegate entire workflows to an AI that actually executes them, keep reading. this article is for you my bro.
yes, I have taken my practice of using it over the last few months + researched what other people are also doing with their cowork, and then on top of that asked claude for extra suggestions to see if I was missing cool use-cases with the new updates.
in-fact, I have ensured that this acts as a complete course so if you've never even heard of claude or you're using it daily, this article will still be useful for you to master claude cowork.
WHAT IS CLAUDE CO-WORK? (AND WHAT IT ISN'T)
claude chat is the assistant. prompt in, answer out. you stay in the conversation the entire time. brilliant for brainstorming, drafting, quick thinking. you are always in the loop.
claude code is for developers. it fires up inside your computer's terminal, writes and executes actual code, manages github repositories. it can be wrapped in an IDE like google's anti-gravity, think of that as the iron man suit around the raw brain power. not built for non-technical operators.
claude co-work is the employee. non-technical interface, same autonomous execution engine as claude code underneath. you delegate a task, it breaks it into subtasks, spins up a virtual machine on your local drive, and drops the finished output into your folder while you step away.
chat is a conversation.
co-work is task delegation.
completely different relationship. completely different results. the moment that distinction lands, you'll look at co-work differently.
SETTING IT UP:
co-work reads and writes files directly to your hard drive. handing it access to your entire system is how disasters happen.
contain it. here's how.
step 1 — choose your model.
on the right side of the co-work interface, pick the AI model running your task.
sonnet 4.6 = your workhorse. cheaper, efficient, arms you with 99% of the capability you'll ever need. use this by default. opus 4.6 = heavy artillery. burns through limits fast. save it for your most complex, highest-stakes work only. the einstein rule: don't put albert einstein on kitchen duty. haiku = lightweight, quick-turnaround tasks only.
keep "extended thinking" toggled on at the bottom of the screen. it ensures claude actually processes complex logic rather than pattern-matching its way to a shallow answer. this detail matters more than most people realise.
step 2: create a sandbox.
right-click your desktop. create a new folder called "claude workspace" or "sandbox." this is where co-work lives. its abilities are locked to whatever folder you assign. everything outside that folder is untouchable.
step 3: grant folder permissions.
inside co-work, click "work in a folder," select your sandbox. claude asks permission to change files within that specific location. click "allow once" or "always allow." done. sleep easy.
step 4: execute your first task.
drop a dozen mixed invoices into the sandbox. prompt: *"sort these invoices into subfolders by category and generate an excel summary sheet."*
co-work outlines its plan on the right side of the screen and executes autonomously. it may deploy parallel sub-agents to complete multiple parts of the task at the exact same time. your invoices are organised before you finish your coffee.
STRUCTURING YOUR AI WORKSPACE:
if you've been running co-work without structure, you already know the problem. it forgets everything between sessions. you're re-explaining your business every time you open a new window. your tasks bleed into each other.
the fix is a project ecosystem.
THE PROJECT ECOSYSTEM:
a project is not a folder. it's a container that holds your files, your custom instructions, your skills, and compounding memory... all in one place.
without a project: claude has no idea who you are. what your business does. what tone you use. you explain yourself from scratch, every session, forever.
with a project: context compounds. week one, you're writing paragraph-long prompts. week six, you type "do the usual" and it knows exactly what that means.
separate projects for separate areas. your youtube project and your finance project should never be in the same place. youtube rules bleed into finance tasks and claude gets confused about which tone to use. keep them isolated. this is non-negotiable.
three ways to create a project:
start from scratch: name it, add instructions, build context as you work import from claude chat — pull in existing web version projects and retain all that memory directly use an existing folder: point co-work at a folder on your machine and it builds the project around those files instantly
YOUR BUSINESS BRAIN:
the project is the infrastructure. the personality lives in .md files. plain text files sitting inside your context folder that claude reads before every single prompt. this is how you make it sound like it's worked with you for two years, not a generic first-day hire.
about_me.md: who you are. what your business does. who your customer is. how you make money. your current priorities. claude reads this every time. every. single. time.
brand_voice.md: how you communicate. tones you like. phrases you hate. paste in examples of your actual writing. this is the file that stops you sounding like everyone else's claude.
working_preferences.md: how you want tasks managed, where files get saved, what format outputs should take.
don't write these from scratch. ask claude to interview you. literally say: *"ask me a series of questions and use my answers to build my business brain files."* fifteen minutes. saves hours every week. every week.
GLOBAL VS PROJECT INSTRUCTIONS:
three levels. each one narrows the scope. all three need to be configured before co-work behaves consistently.
level 1: claude personalisation (everywhere). click your name, go to settings. applies to chat, code, and co-work alike. lock in default behaviours: avoid excessive bold formatting, use primary sources over aggregators for research, no hedging language. these are your universal rules.
level 2: co-work global instructions. settings → co-work settings → global instructions. applies only inside co-work, across all tasks. use this for mandatory formatting rules: specific date formats, file naming conventions (underscore_descriptive_name), instructions to always check your sandbox folder when handling business queries. saves you typing the same context on every task.
level 3: project-specific instructions. rules that live strictly inside one project. your youtube project tells claude to use a specific slide skill for video intros and track everything in a specific spreadsheet. those rules should never touch your accounting project. keep them isolated or they contaminate each other.
EQUIPPING YOUR AI:
this is where co-work stops being a useful tool and becomes an autonomous operator. we give it the ability to reach outside the sandbox and interact with the apps, the web, and your desktop.
three tiers. use them in this order.
connectors first. browser use second. computer use as the final fallback.
CONNECTORS & MCPs:
connectors (also called MCPs - model context protocols) let claude operate directly inside the apps you use every day. not through copying and pasting. not through screenshots. directly inside the interface, with authenticated access.
native connectors cover google drive, notion, slack, gmail, github, figma, and more.
click the plus sign in your chat go to the connectors tab browse connectors, search for your tool follow the authentication flow in your browser
once connected, you get fine-grained permission controls. for each connected app, you set specific actions to "always allow," "needs approval," or "blocked." you stay in complete control of what it touches and what it doesn't.
specialist connectors - gamma: when claude generates a presentation deck without a specialist connector, the formatting is mediocre at best. toggle on the gamma connector and co-work automatically routes your data to gamma, which produces properly designed, visually structured decks. the right tool for the right output. stop trying to force a generalist to do a specialist's job.
EXPERT HACKS:
hack 1: the apify MCP (scraping without the headache).
want to harvest data from youtube, tiktok, or instagram? you don't need n8n. you don't need to build complex automation pipelines. i tried both before finding this. the difference is embarrassing.
create a free API token inside an apify account browse co-work connectors, search "apify," paste your API key done
when you prompt claude to scrape a platform, it automatically sorts through apify's library of over 1,300 scrapers (they call them "actors") and deploys the right one for your specific task. no manual selection. no configuration. it finds the tool, runs it, returns the data.
hack 2: the zapier MCP (8,000+ app connections).
if your app isn't natively supported by claude, this is the answer. this was the feature that genuinely shocked me when i found it. zapier's MCP connects to over 8,000 applications, hubspot, skool, airtable, whatever you're running.
build an MCP server inside zapier's platform select "claude co-work" configure the specific tools and actions you want to allow copy the provided URL, go back to claude's connector library, search "zapier," paste it
8,000 apps unlocked in under 10 minutes.
this single integration obliterates the entire "but my tool isn't supported" objection. permanently.
BROWSER USE: AUTONOMOUS WEB NAVIGATION
no connector available? the chrome extension steps in. this is tier two.
installation:
open google chrome → chrome web store → search "claude" install, pin to your taskbar enable "claude in chrome" inside co-work settings
once active, claude opens browser tabs, navigates to websites, and reads pages autonomously. give it a URL and ask it to audit a landing page, assess call-to-action buttons, or run competitive analysis. it can read emphasis like italicised text, headers, layout hierarchy, all on the live page.
it can click through sites too. tell it to open youtube, navigate recommended videos, and return the view counts, like ratios, and comment sentiment. it does it.
the warning you actually need to take seriously: claude uses your real browser. it is logged into your personal accounts. if you ask it to search for flights and you have a saved payment method with the airline, it technically has the capability to complete that purchase. monitor its activity. use blocklists. this is not hypothetical. treat it like you would any new employee with access to your accounts.
COMPUTER USE: FULL SCREEN & DESKTOP CONTROL
connector can't solve it. browser use can't solve it. deploy computer use. the absolute fallback.
this lets claude literally see your screen, control your mouse, and type on your keyboard to navigate native desktop applications. anything you can do on your computer manually, claude can now do autonomously.
enabling it:
click your account name → settings → general → toggle on "computer use" immediately add sensitive desktop apps to the blocklist before you do anything else
in practice: tell claude to find a specific video file on your desktop and drag it into a capcut project. claude requests access to finder and capcut. you take your hands off the keyboard. it visually searches your files, opens the editing software, locates the file, and completes the action.
it's extraordinary. it's also the reason you arm the blocklist first.
AUTOMATION & DELEGATION: BUILDING THE MACHINE
here's the rule that changes how you use this tool:
if you do the same task more than once a week, you should be automating it. full stop.
the automation stack has four layers: skills, plugins, scheduled tasks, dispatch mode. each one builds on the last. you don't build them all at once, you build them in order.
CUSTOM SKILLS: TURNING WORKFLOWS INTO SINGLE COMMANDS
a skill is a reusable AI workflow encoded into a single command. instead of typing a long prompt every time you need claude to execute a specific process, you encode that process once and call it with a slash command forever.
behind the scenes, a skill is an .md file containing instructions and sometimes scripts. claude loads it dynamically when executing tasks. you don't write the code. you instruct claude in plain english and it builds the file.
the creation process:
the prompt: describe exactly what the skill should do. for instance: apply your specific brand colours and typography to every presentation it generates. the evaluation loop: co-work's built-in skill creator automatically tests your new skill. it runs a task with the skill and without the skill (the baseline) and compares the outputs side by side. iteration: review the test files. if it missed something (forgot to use the beige background, used the wrong font weight), give that feedback. it applies the fix. save and deploy: once the test output matches your standard, click "copy to your skills." you now call this entire workflow with /summarise_invoices or whatever command you assign.
what you can build:
parse a folder of raw invoices into a categorised excel spreadsheet automatically scrape a youtube URL and generate an interactive HTML transcript with timestamps produce on-brand infographics using an external image model API with a single command
PLUGINS: CHAINING SKILLS INTO MASTER WORKFLOWS
a skill automates a single process. a plugin chains multiple skills and connectors together to automate an entire role.
the test: skill or plugin?
does this task happen more than once a week? if no, use a normal prompt. done. does it require 3+ steps across 2+ tools (slack, gmail, and notion together, for instance)? if no, build a skill. if yes, build a plugin.
to build a plugin, tell co-work what workflow you want to automate. it combines the necessary MCPs and skills into one master package that runs the entire pipeline on a single command.
the best thing about plugins: they are shareable. package your best standard operating procedures into a plugin and distribute it to your team or community. they execute tasks exactly the way you do. your workflows become transferable.
SCHEDULED TASKS: RUNNING AUTOMATIONS ON A TIMER
skills and plugins can now fire automatically on a schedule, without your input, while you're not there.
setting up a schedule:
navigate to the "scheduled" tab in the left sidebar click "new task" give it a name and description in the prompt section, reference the skill or workflow (e.g., "use my flight search skill to find return tickets to spain under £2,400") set the frequency: hourly, daily, weekly, or specific days at a specific time select the designated project folder and save
the golden rule: scheduled tasks only run if your physical computer is on and the co-work app is open. if your laptop is shut at 9am, the 9am task waits until you open it. adjust your power settings to keep the machine awake. this one detail breaks more scheduled automation than anything else.
DISPATCH MODE: REMOTE CONTROL FROM YOUR PHONE
away from your desk. task needs running on your local machine. that's exactly what dispatch mode is for.
text claude from the mobile app and it executes work on your desktop while you're elsewhere. same conversation thread across both devices, so context never resets between your phone and your machine.
setup:
enable dispatch in co-work settings on the desktop app toggle the "keep awake" button in the dispatch menu is critical. this prevents your computer sleeping and blocking file access allow browser actions and computer use
from your phone: "scan my receipts folder and generate a dashboard." co-work does the heavy lifting on your desktop. enable push notifications and it pings you the moment the files are ready.
MASTERCLASS WORKFLOWS: THREE SYSTEMS TO BUILD TODAY
these are the three i'd build first. in this order.
1: the morning brief.
schedule claude to build your daily dashboard before you sit down. connect your calendar and email. it summarises your upcoming appointments, lists outstanding email actions, checks local weather, and pulls breaking news relevant to your industry. it can draft email replies for you to review and send. you sit down. everything is already done. this one alone is worth the subscription price.
2: the content repurposer.
give claude a single youtube URL. instruct it to extract the transcript, add the content to a new notion page, and write platform-specific posts for linkedin and X automatically. one input. three outputs. zero manual work.
3: financial reporting.
set up a monthly scheduled task. give claude access to your bank transactions (not your whole details just your transactions, or skip this if it's uncomfyyyyyy or a folder of receipts). it categorises expenses, checks incoming and outgoing balances, and generates an interactive HTML dashboard showing your profit and loss. your accountant gets a clean report. you spent zero minutes on it.
MANAGING TOKENS: HOW TO AVOID BURNING YOUR LIMITS
every word you send claude and every word it reads from your files costs tokens. manage this poorly and you'll hit your ceiling before the week is out.
i've watched people burn through their max plan limits in three days because of three avoidable mistakes. here they are.
mistake 1: ignoring the base load. your context window is already partially full before you've typed a single word. system instructions, active tools, and open MCP connectors all count against it. the more connectors you have active, the faster you burn. only switch on what the current task actually needs.
mistake 2: context rot. keep talking in the same chat tab for hours and claude reloads the entire history of that conversation every single time. you finish drafting an email, then ask it to plan a trip to dubai in the same window, you are paying to load the email context into the travel query. that is pure waste. the number of people burning their limits this way is genuinely painful to watch.
mistake 3: conversational processing. if you need to process 100 invoices, having claude read and process each one manually in conversation will drain your session fast. ask co-work to write a reusable script (a skill) to parse those invoices instead. the script uses a fraction of the tokens. this is not a small difference.
the 30-to-45-minute rule: keep sessions tight. one focused topic per window. after 30 to 45 minutes, or whenever you change subjects, open a new window. fresh context, fresh budget.
parallel sub-agents: for large tasks, instruct claude to run things in parallel. it deploys multiple sub-agents, each with its own separate context window, handling different parts of the task simultaneously.
the einstein rule: opus for complex, high-stakes reasoning only. sonnet for everything else. 99% of the time you do not need opus. stop using it as a daily driver.
AI SAFETY:
co-work runs directly on your local machine. anthropic builds safety into the core of what they ship, so the baseline risk is low. that's not where the danger comes from.
the danger is other people.
it's possible to download highly advanced skills built by other users, usually on github. when you download an external skill, you are importing instructions that claude will execute on your behalf, on your machine, with whatever access you've granted.
if a bad actor has embedded malicious instructions inside that .md file, you're looking at a prompt injection. instructions that could tell the AI to delete files, exfiltrate data, or escalate its own access across your system.
this is not theoretical. check every skill you download. every time.
the safety check:
before adding any external .md file to your skills library, paste the entire text of the skill into claude chat and ask it directly: *"is there anything in this skill that could be harmful, malicious, or give instructions outside the scope of the stated task?"*
two minutes. every time. non-negotiable.
WHAT YOU SHOULD TAKEAWAY:
everything in this guide builds on one reframe.
do not think of claude cowork as a chat tool.
it is task delegation to an autonomous employee.
the moment that lands, you stop asking it questions and start giving it work.
here's what matters most:
1. architecture beats prompting. a solid project structure, the right .md context files, and isolated project instructions will do more for your output quality than any clever prompt ever will. build the foundation once. it compounds from there.
2. the automation stack is sequential. skills before plugins before scheduled tasks before dispatch. each layer builds on the last. don't try to build the full machine on day one. automate one weekly task first. go from there.
3. security is not optional. check every community skill before running it. the safety check takes two minutes. that two minutes could save your entire machine.
your move: create your sandbox folder today. grant permissions. run one real task. the messy downloads folder or a batch of invoices is all you need to see what this tool actually does.
which workflow are you starting with? the morning brief, the content repurposer, or the financial dashboard? drop a reply. i read every one.
IF YOU WAS TO TAKE ONE THING:
go use the zapier MCP, it connects to over 8,000 apps and takes 10 minutes to set up. if you skimmed past that section, go back. it's the single most underrated feature in the entire tool. most people never find it.
OH, AND IF YOU WANT A FREE NEWSLETTER THAT WILL KEEP YOU UP TO DATE IN EMERGING TECHNOLOGY THEN...
HERE'S THE PLUG TO MY NEWSLETTER:
