How to set up Claude for your business
A working setup in one sitting: which model to default to, the two memory settings most people never turn on, projects as buckets, and the scheduled tasks that run while your laptop is shut.
Setting up Claude for business work takes four things: turn on memory so it references past chats, paste a profile of your role and goals into the account instructions, split work into projects with one wired to your Drive as a company brain, and connect your tools read-only before scheduling recurring tasks.
Treat it as a thought partner, not an autopilot
The useful split is roughly 10/80/10. You own the first ten percent: the idea, the concept, what the market actually wants. AI takes the eighty in the middle, which is the repetitive work you have already proved you can do and should never do again. Then you own the last ten: look at the result and fix what is wrong.
It is the way a pilot flies. Heavily engaged on takeoff, cruising on autopilot at 35,000 feet, hands back on the controls for the landing. We built the funnel for a digital workshop this way — the concept and the offer were ours, the landing page, the button-moving and the first draft of the copy were not, and then we went back in and changed what needed changing.
Which model to default to
Set Opus 5.5 at medium or high effort and leave it there for day-to-day work. It handles the broad mix well enough that switching models constantly is wasted attention.
Two exceptions are worth knowing:
- Sonnet 5.5 for deterministic, repetitive jobs — "every time this kind of email arrives, draft a reply shaped like this." Little cognitive load, and considerably cheaper to run at volume.
- Fable for the genuinely large jobs. If you are putting two years of sales calls through an analysis, let that one run.
The two settings most people never turn on
Open settings and go to memory. Turn on the ability to search and reference past chats, and the ability to generate memory from chats. Without them, every new chat window starts cold: you spent yesterday explaining a client situation in detail, and today it has never heard of them. With them on, it carries that context across chats and projects.
While you are in settings, check your training setting — what happens to your data differs between plan tiers, and it is worth knowing which one you are on rather than assuming. The rest is ordinary hygiene: ask for sources, check the numbers before you forward them, and never paste passwords or API keys into a chat.
Make it know who you are
Under account there is a box for instructions for Claude, and most people leave it empty. Fill it. The fastest route is to have Claude interview you — what you do, your role, what you want the most help with, how you like answers delivered — then have it write the summary paragraph and paste that into the box.
Three minutes of answering questions, and from then on every chat and every setting opens with your role, your company and your goals already loaded. It is the single highest-leverage thing in the whole setup, and it is the one almost nobody does.
Projects are buckets, and one of them is your company brain
Think of projects as separate buckets for the separate parts of your life. One for health. One for the book you are trying to write. One per client. The point is that the workspace does not collapse into a single undifferentiated mess, and that you go to a specific place to do a specific kind of work. Dictation and voice now work inside each project, and you can add files from your device or pull them straight in from a connector like Drive or Canva.
Then make one project your brain. Point it at your Google Drive for context, and tell it in the project instructions to keep updating its memory as you work. People build this with Obsidian and a dozen other tools; you can get most of the way there inside one project. After that, business questions go to that project, which already knows the answers.
Connect your tools, read-only first
This is where it stops being a chatbot. There are thousands of connectors — Google Workspace or Microsoft, Slack, Notion, Canva, your CRM, your finance and marketing stack. Connect your core tools, starting with email and calendar.
The privacy concern is real and the answer is simple: set them read-only. It can see your inbox, your calendar and your Drive, and it cannot send anything. That alone is most of the value, because the job you actually need done is synthesis — noticing the things that fall through the cracks. Once connected, this kind of request works:
- Look through my email from the past 30 days. What did I miss?
- Find the five Drive documents I have worked on most recently and turn them into a dashboard.
- Look at my calendar. Which of these needs prep, and what do I need to do?
Stop asking and start scheduling
Most people stop at the point where they have connected a few tools and have it do tasks on request. The step past that is work that happens without being asked. Go to scheduled tasks, choose to create one with Claude, and describe it in plain language: every Monday at 8am, go through my Gmail, find the emails that need a follow-up from me, draft the replies, and put anything from last week's invitations that I missed onto my calendar.
It walks you through the setup, and then it runs in the cloud — your computer does not need to be on. A recruiter we know has one that scans job boards every morning for new postings matching his criteria and emails him at 9am with five roles and five people to contact. The test to apply to your own week: what do I do every single morning, and what would I hand to an assistant if I had one?
Artifacts are for the things you hand to other people
Artifacts build small interactive things — web apps, slides, simple sites, polls. The slide deck behind this walkthrough was built with them, which is suitably circular. The better example was a team training: an interactive poll sent to the room, fifteen people answered a few questions live while a song played, and then we pulled up the results on the call — the average comfort score with AI, the ten things people were most afraid of, and what they were most excited about.
That is not a toy. It is a way to get a room to tell you what it actually thinks in ninety seconds, and it took no design work to produce.
The short version
- Default to Opus 5.5 at medium or high effort. Use Sonnet 5.5 for cheap repetitive automation and Fable for very large analysis jobs.
- Turn on both memory settings — reference past chats, and generate memory from chats — or every chat starts from nothing.
- Fill in the account instructions box with your role, company and goals. Three minutes, and it applies to every chat afterwards.
- Use projects as buckets, and wire one to your Drive as a company brain that updates itself as you work.
- Connect your tools read-only, then move from asking to scheduling. Work that runs on Monday at 8am without you is the actual goal.
Questions this raises
Which Claude model should I use for business work?
Opus 5.5 at medium or high effort is the right default for everyday work, and switching constantly costs more attention than it saves. Use Sonnet 5.5 for high-volume deterministic jobs like drafting a standard reply to a recurring email, where it is both capable and much cheaper. Save Fable for genuinely heavy analysis, such as working through two years of call transcripts.
Is it safe to connect my email and calendar?
Connect them read-only to start. The tool can then see your inbox, calendar and Drive but cannot send an email or take any other action on your behalf. Most of the value is in synthesis rather than action — catching what slipped through — so read-only access covers the majority of real use cases. Never paste passwords or API keys into a chat, and check your plan's training setting while you are in the settings menu.
What is a company brain and do I need another tool for it?
It is a single project that holds the context about your business, pointed at your Google Drive and instructed to keep updating its own memory as you work in it. People build the same thing with Obsidian or a dedicated knowledge base, but a project inside your existing subscription gets you most of the way there with no extra tool to maintain.