Most people open Claude, ask one random question, don't love the answer, and quietly decide it isn't for them. The problem was never the model. It's that they treated it like a search box instead of setting it up around their actual work.
If you came from the video, that's the idea: five steps, one afternoon. Here's the part that doesn't fit in sixty seconds. The order I'd do them in, one real business set up end to end, and the things that quietly go wrong the first time. (The desktop UI shifts a little by rollout, so use the labels you see in your app. Updated 27 Sep 2026 for the Chat and Cowork merge.)
The five steps, in order
Do them in this order on purpose. Each one gives the next something to work with, so by step five Claude is finishing a real task instead of guessing at one. All of this happens in the Claude desktop app. As of September 16, Claude folded Cowork (the part that can open your files and use your connected apps) into every conversation, so you no longer pick a mode first.
Open the desktop app and point it at one folder
Download the desktop app and start a conversation. Pick the single folder that holds the documents for the job you want help with, not your whole drive. Installing the app doesn't hand Claude your computer; it can see the folder you point it at and nothing else. Keeping the scope small also makes it obvious later what it is and isn't allowed to touch.
Write global instructions like you're briefing a new hire
Go to Settings, then Cowork, then Global instructions. This is what you'd otherwise retype in every chat: what you do, who you serve, and how you want it to behave. "Give me the recommendation first. Tell me if my numbers look wrong. Check my files before you guess. Ask before you send anything." One line per rule beats a paragraph. The copyable version is below.
Connect your apps, then test the connection
Open Customize and connect what your business runs on: Google Drive, Gmail, your calendar. Don't take the green checkmark on faith. Ask it to find one specific file or the last email from one specific person. If it pulls the right thing, the connection is real. If it can't, fix that now, before you build a task on top of it.
Make a project for one slice of your work
A project is a workspace with its own files and instructions, so facts about one part of your business don't leak into everything else. Pick one slice, add the files that matter to it, and in the project instructions write the specifics: who it's for, what you're producing, and any dates or numbers that can't be wrong. Global instructions are how you work; project instructions are the facts of this job.
Run one real task, then save the steps as a skill
Now use it for something you'd have done by hand today. Watch the first run, edit the result, and tell it why you changed things. Once you like the output, ask it to save those steps as a skill: a named set of instructions it can rerun. The point of day one isn't one good answer. It's a repeatable process you can point at the next job with new inputs.
A real one, set up end to end
Abstract steps are easy to nod along to and hard to act on, so here's the whole thing for one specific person: a wedding planner on her first day with Claude, two weddings on the books and a folder full of client documents.
- The folder is her "Clients" folder: one subfolder per couple, holding signed contracts, the venue rules PDF, vendor quotes, and a timeline doc. That's what she points Claude at.
- Global instructions are her working style: "I plan weddings for couples in [city]. Show me the total budget impact before you suggest a vendor swap. Never email a client or vendor without showing me the draft first. If a date or a deposit amount isn't in the files, say so."
- Connected apps are Gmail (to read the vendor thread), Calendar (to see what's booked), and Drive (where the folders live). She tests it by asking for the last email from the florist, and it finds the real one.
- The project is one wedding: the Patel wedding in November. She loads that couple's contract, the venue rules, and three catering quotes, and writes the guest count, the ceremony date, and the hard budget cap into the project instructions.
- The real task is turning those three quotes into a comparison with the budget math shown. She corrects one thing (the corkage fee wasn't in the quote, so she adds it), then saves the steps as a "compare vendor quotes" skill she reuses for every couple after this.
By the end of the afternoon she doesn't just have one answer. She has a folder Claude understands, a rule that stops it emailing anyone behind her back, and a reusable process for the part of the job she does over and over.
Copyable starter instructions
Paste this into your global instructions and fill in the brackets. It sets your working style plus the guardrail that matters most on day one: it has to hand the risky actions back to you.
I run [business] for [customer]. My current priorities are [priorities].
Give me the recommendation first, then the reasoning I need to act.
Use the files and connected sources I identify. Link the source behind
factual recommendations. Push back when a claim or calculation doesn't
follow from the evidence. Ask before sending, publishing, buying, or
changing a live customer record. General preferences belong here;
specific project facts belong in the relevant project.Then, inside the project you set up in step four, give it the first real task. This is written for a launch, but swap in your own slice, whether that's comparing quotes, prepping a client update, or drafting a proposal.
Use the files in this project to create a launch brief. Include the
customer problem, the offer, positioning, launch channels, budget
allocation, and next actions. Reference the research behind the choices.
Flag missing inputs rather than inventing them. Give me a draft I can edit.And here's the checklist I run before I call day one done. It's short on purpose. If all four are true, you have a setup you'll actually keep using, not one you configured once and forgot.
[ ] Claude is pointed at one real folder, and I confirmed it can read a file in it
[ ] Global instructions say how I work and what needs my sign-off before it happens
[ ] At least one app is connected and I tested it by finding a specific file or email
[ ] I ran one real task, edited the result, and saved the steps as a reusable skillWhat goes wrong the first time
None of these are dealbreakers, and each has a small fix. Knowing them ahead of time is the difference between a smooth afternoon and a frustrating one.
It confidently states something that isn't in your files. It didn't have the document, so it filled the gap. This is why step three matters: test that it can actually read the folder before you trust an answer. When you catch it, add "if it isn't in the files, say so" to your instructions and reload the missing document rather than arguing with the output.
You skip global instructions to "just try it," and every chat starts from zero. This is the most common way day one stalls. You re-explain your business in each conversation and it feels like the model is dense. It isn't; you never told it who you are. Ten minutes on step two pays for itself by lunch.
You pile every fact about your whole business into the global instructions, and now the November wedding's budget cap shows up when you ask about a different couple. Standing rules go global; the facts of a specific job go in that job's project. Keeping them separate stops one client's details bleeding into another's.
You let it act before you've watched it work. It connects Gmail, drafts a reply, and offers to send. On day one, keep sending, buying, and booking as things it drafts and hands back, not things it does. This is the same lesson from building your first agent: the block belongs in the app's permissions, not a polite line in the prompt.
Do all of this while you watch, starting with the files and apps you already use, and you'll end the day with something that earns its place in your work. Rather use a different assistant? The same five steps map onto setting up Grok in a day. And once one slice is running, the next question is which parts of your business should AI handle at all.