PLAYBOOK · 01 · GETTING STARTED · BY MARC KLEINMANN
An hour from install to your first shipped output. The mental model, the setup, the decisions that matter, and the mistakes I see operators make most. Written for the SMB owner and the ops lead deploying AI inside the business they already run.
01 ·Context
You've probably already tried this once. You opened Chat in a browser, typed a vague request — "help me with our marketing" — got a paragraph of generic advice, decided the AI wasn't ready, and went back to doing the work by hand. Six months later you tried again, got the same result, gave up again. The product didn't get worse. You never learned how to use it.
That failure mode is not a Claude problem. It's a mental-model problem. Claude is not a smarter Google. It is a system with two modes, four context layers, a connector graph, and a governance pattern. You can use it as a chatbot — and a lot of people do — but you'll get chatbot results. The operators who get real work out of it treat it like a colleague who joined three days ago: capable, willing, fast, and completely dependent on you telling them what the company actually does and how things actually work.
The other failure mode is more subtle. People install Claude, find the Chat interface, do useful work there, and never discover that Cowork exists. Cowork is where the deliverables happen. Cowork is where the system reads files on your computer, plans multi-step work, executes autonomously, and produces a .docx or .xlsx or .pdf you can send. If you've only used the browser, you've only used half the product. You don't know what you don't know.
The cost of getting this wrong is real. I've seen operators spend weeks treating Claude like a slot machine — pulling the lever, getting noise, pulling again — when the right setup would have gotten them to working output in an afternoon. This playbook is the afternoon. Sixty minutes if you move fast.
02 ·Architecture
Claude is one account that resolves into two operating modes plus a shared connector graph. The modes share an account, billing, and the tools you connect — but they do not share context with each other.
Chat lives in your browser or in the desktop app's chat surface. You type, it responds. It is conversational, stateless across sessions except for what you explicitly capture, and cloud-based. Chat is the right tool for quick questions, short drafts, brainstorming, anything that ends with a paragraph or a list rather than a file.
Cowork lives only in the desktop app. It reads files on your computer, plans multi-step work, executes across that plan, and saves finished files to a folder you specify. Cowork is the right tool for anything that produces a deliverable: a status report, a campaign brief, a financial summary, an analysis with charts. Think of Chat as asking a colleague a question. Cowork is delegating a project and checking back when it's done.
These two modes have separate identity systems. They do not share what they know about you. Your Chat preferences do not carry to Cowork. Your Cowork identity file does not carry to Chat. If you want both surfaces to know the same things about you, you set both up — twice. Annoying, but it's the cleanest mental model: think of them as two products that happen to share an account.
When you open Chat, four layers of context resolve in order. The first three are global; the project layer turns Chat from a smart search box into a domain specialist who already knows your business.
| Layer | Scope | What it controls |
|---|---|---|
| User Preferences | Account-wide | Identity, constraints, defaults — read at the start of every conversation |
| Style | Per-conversation | Tone and formatting for the conversation you set it on |
| Memory | Account-wide | Patterns Claude has learned about you over time |
| Chat Project | Per-project | Project Instructions and Knowledge Files scoped to that project |
Three files do most of the work. CLAUDE.md sits at the root of any folder Claude is working in and tells it the rules — what to do, what not to do, where to write, who owns what. about-me.md is your personal identity file — how you write, what you don't like, what you assume Claude already knows. brief.md is the per-project mission file — what the project is, what it produces, how it operates. Beyond those three, the folder structure itself carries context:
your-shared-folder/ # rename to whatever your team uses ├── CLAUDE.md # the rules — what to do, what not ├── about-me.md # your personal identity ├── brief.md # per-project mission ├── knowledge/ # reference material ├── memory/ # auto-memories Claude writes about you ├── skills/ # repeatable task playbooks └── outputs/ # finished work lands here
This is the load-bearing difference. Chat's context lives in the cloud and is per-user. Cowork's context lives in files on your filesystem, syncs across the team via your shared storage (Dropbox, iCloud, Google Drive — pick one and stick with it), is version-controlled by whatever sync tool you use, and is auditable. Both work. They serve different jobs.
Connectors wire Claude into the apps you already use: Gmail, Calendar, Slack, Drive, your CRM, your billing tool, your project tracker, dozens of others. You connect once at the account level and the tool becomes available in both Chat and Cowork. The key point: when you connect, Claude inherits your permissions inside that tool. If you have read-only access to a shared inbox, Claude can read but cannot send. If you have admin rights to your CRM, Claude can update deals. Anything Claude does through your connector shows up in the tool's audit log as your action.
The frame that worked best for me, and that I now hand to every operator I onboard:
If you keep that frame, the rest of this article reads obvious.
03 ·Setup
Six steps. Do them in order; later steps depend on earlier ones. I've timed each one against a normal operator pace — not "I write Python for a living" pace.
Browser-only access gets you Chat. The desktop app gets you Chat plus Cowork. There is no reason to skip the desktop app — it's free, the install is a normal one-click, and you need it for the half of the product that does the deliverable work. Mac and Windows both work; the experience is the same.
Sign in with the email you intend to use as your work account for AI. If you're at a company with multiple people on Claude, use your work email — that's how teammates will eventually share folders with you. If you've already signed up with a personal email and you want to switch, do it now while you have nothing to lose.
Claude offers more than one model. The two you'll use are Sonnet for most things and Opus with Extended Thinking for hard things. Sonnet is fast and handles roughly 80% of daily Chat work — quick questions, short drafts, light analysis. Opus with Extended Thinking is slower but reasons harder; it is the right pick for anything multi-step, anything that requires real judgment, anything you'd ask a smart consultant to do.
You pick the model per conversation, at the top of the chat window. There is no global default. Cowork sessions should always run on Opus with Extended Thinking — Cowork's autonomous planning needs the stronger model.
In Settings → Connectors, you'll see a list of every tool Claude can talk to. Don't connect everything at once. Pick two or three you use every day. For most operators that's: Gmail, Calendar, and one of Slack, Drive, or your CRM.
Each connector takes you through an OAuth flow — the same kind you'd use to log into a third-party app. Authorize the scopes Claude requests. Verify the connection is live. Move on. The connectors persist; you'll only do this once per tool. If you're at a company that uses Microsoft instead of Google, swap Gmail and Calendar for Outlook and Microsoft Calendar. The principle is the same: connect the two or three tools that hold the data Claude will need to be useful to you.
In Settings → Preferences, you'll find a text field for account-wide instructions Claude reads at the start of every Chat conversation. This is where you write the things you'd tell a new assistant once and then never want to repeat. Four kinds of lines, with copy-able starting points:
I run operations at a [mid-size services business]. I'm not technical but I'm comfortable in spreadsheets.Most of what I produce goes to [clients or my CEO]. Default to executive register unless I tell you otherwise.Never write in marketing speak. Never use words like "leverage," "synergy," "robust." If I ask for a list, give me a list — not paragraphs.
I use [HubSpot, QuickBooks, and Slack]. Default to those when you reference where data lives.Style is a separate, per-conversation setting accessible from a dropdown near the model selector. It controls tone and format. You can leave it on Default for most things; switch it to Concise or Formal when the conversation calls for it.
This is the step that turns Cowork on. You need a folder somewhere on your filesystem that Claude can read and write to. Whatever sync tool you use — Dropbox, iCloud, Google Drive, OneDrive — pick one and put your Cowork folder there. The point of the sync is twofold: it's a backup, and if you later add teammates they'll be able to share folders with you the same way they share documents.
Now open the desktop app, start a new Cowork session, mount the folder you just synced as the working folder, and type this in plain English:
Run the about-me quiz.
Claude will interview you for fifteen to twenty minutes. The questions cover writing style, anti-patterns, formatting preferences, behavior rules, and the context you want carried into every Cowork session. Answer in your own voice. Don't corporate it up. The file Claude writes at the end of the quiz — about-me.md — is the Cowork equivalent of your Chat preferences, and it's the difference between Cowork that sounds like a stranger and Cowork that sounds like an extension of you.
Don't end your first session without producing a real output. Pick something small. Three good starting prompts:
Write a one-page weekly status update for me, based on the structure I'd normally use, with placeholders where you don't have the data yet.
Draft three subject lines for a customer follow-up about a delayed shipment.
Pull my last week of calendar events and tell me where my time went.
The point of the first prompt is not to be useful. The point is to confirm the system works end-to-end and to start training your own judgment about how Claude responds. You will learn more about your setup from one real prompt than from ten more pages of explainer.
04 ·Decision points
Five decisions matter. None of them are settings you flip once; they recur every time you work.
The clean rule:
| If the output is… | Use… |
|---|---|
| A file (.docx, .xlsx, .pdf) | Cowork |
| A paragraph, list, or conversation | Chat |
| Quick scoping or brainstorming | Chat |
| Not sure | Cowork — it can produce conversational output; Chat cannot produce a .docx |
The messier truth: a lot of work starts in Chat (brainstorming, scoping, deciding what the deliverable should be) and finishes in Cowork (producing the deliverable). It's normal to switch mid-project. Just don't try to ferry context between them by hand — copy the Chat conversation's conclusion into a brief.md or a prompt for the Cowork session and start fresh.
| Task | Model |
|---|---|
| Quick questions, short drafts, light reformatting | Sonnet |
| Pulling a list out of a connector | Sonnet |
| Multi-step strategy, analysis, multi-document synthesis | Opus with Extended Thinking |
| Anything inside a Cowork session | Opus with Extended Thinking |
The cost difference matters more than people think. Sonnet is fast and cheap and good enough for the majority of daily work. Defaulting to Opus for everything is the AI equivalent of using a consultant for tasks an admin could do — it works, but it's slow and it teaches you bad habits about how to scope work.
When you write something into your Chat Preferences or your about-me.md, you're putting it into your personal context. It follows you into every conversation. When you write into a Chat Project's instructions or a Cowork project's CLAUDE.md, you're putting it into shared context — anyone working in that project sees it.
| Example line | Where it belongs |
|---|---|
| "I work in healthcare and HIPAA matters." | Personal — Preferences or about-me.md |
| "I write in short sentences." | Personal |
| "Our CRM is HubSpot." | Shared — Chat Project or CLAUDE.md |
| "We always confirm before sending external communication." | Shared |
| "Our company tone of voice is direct and unflashy." | Shared |
The most common mistake is putting shared knowledge into personal context. If your whole team needs to know the company's tone of voice, that doesn't belong in your personal preferences — it belongs in a Chat Project that everyone uses or in the team's shared CLAUDE.md. Otherwise every teammate has to rebuild that context from scratch.
Two questions for every potential connector. First: do you trust the tool enough to give Claude your-level access to it? If the tool holds sensitive data and Claude shouldn't touch it, skip the connector — uploading specific files into a Chat conversation is a more contained alternative. Second: would Claude actually need this tool to be useful? If you've never asked Claude about that tool's data, you don't need the connector yet. Add it later.
A starter set of three to five connectors covers most operators. Email, calendar, one communication tool (Slack or Teams), one document store (Drive, Dropbox, OneDrive), one business system (your CRM or your billing tool). Beyond that, add as the use case appears.
The first time you do a task in Cowork, you write a prompt. The second time, you write the same prompt again — slightly different, probably worse. The third time, you turn it into a skill. A skill is a markdown file in your Cowork project's skills/ folder that describes the task in detail — what to do, what files to read, what format to produce, what to confirm before writing. Once it exists, you can trigger it by name or by slash command.
05 ·Runbook
Daily and weekly rhythms that keep the setup honest.
Chat or Cowork. Decide before you start typing. Five seconds. Saves you the worse experience of asking Chat for a file or asking Cowork for a quick gut check.
Sonnet for fast, Opus for thoughtful. Default to Opus inside Cowork.
The single most reliable lever you have on output quality. Vague prompts get vague answers; this is information theory, not a Claude limitation.
A useful prompt names what you want, who it's for, what format it takes, how long it should be, and what the constraints are. The contrast is steep:
Draft a 200-word client follow-up email for a project that's running two weeks late, in our normal direct tone, with a clear path to the next milestone.
Write a client email.
Cowork's default behavior before any significant action is to show its plan — here's what I'm going to do, here's where I'm going to write, here are the files I'm going to read — and wait for your approval. Plans are where you catch the misunderstanding before it costs you twenty minutes of wrong output. Plans are also where you redirect without having to start over.
When you write a prompt that gets you exactly the output you wanted, save it somewhere you can find it again. The simplest version is a prompts.md file inside your Cowork project. The smarter version is to turn the prompt into a skill the next time you'd otherwise write it again.
Each conversation has a context that builds up over time. If you switch from drafting an email to analyzing a spreadsheet inside the same conversation, Claude's response quality degrades — earlier context muddies the new task. Start a new conversation when the topic shifts. It is free, fast, and produces better answers.
Once a week, do two small things. First, glance at the auto-memory files Cowork has written about you — they live in your memory/ folder — and delete or correct anything that's gotten stale. Memory is editable; it is not a black box. Second, look at the prompts you wrote during the week and identify the ones you've now done three times. Promote those to skills.
The instructions you wrote on day one were a best guess. After a month of real use, you know more about what you actually want. Update your Chat Preferences and your about-me.md. Treat them as living documents.
06 ·Gotchas
The six failure modes that catch every operator at least once.
The most common failure by an order of magnitude. Help me with our marketing. Improve this. What do you think? These get you garbage answers that look like advice. The fix is not to ask Claude to "try harder" — it's to rewrite the prompt with the specificity from Section 5. A better input always beats a retry.
Every operator gets caught by this once: Chat and Cowork do not share context. If you spend an hour in Chat training Claude on your business and then open Cowork, Cowork knows none of it. The fix is structural: put shared context in files (about-me.md, brief.md, CLAUDE.md). Chat needs its own setup (Preferences + Chat Projects).
Using Opus for what time is the meeting burns time. Using Sonnet for a multi-step strategy analysis produces shallow work. Match the model to the cognitive load of the task. You'll get this calibrated in about a week of normal use.
Cloud-only files break Cowork. If you sync your Cowork folder with Dropbox or iCloud or any other tool that defaults to keeping files in the cloud, you have to explicitly mark the folder available offline. The symptom is Cowork starting a session and reporting that files don't exist when you can see them in your file explorer. The fix is in your sync tool's settings; once you've done it for the folder, it stays done.
Connectors can write — they can send email through Gmail, update deals in your CRM, post to Slack. The first time you ask Claude to follow up with the client it might genuinely send the email, depending on how you scoped the request. The protection is to be specific: Draft the follow-up; do not send it yet. And add a guardrail to your Preferences or about-me:
If you start adding plugins or installing community skills, you can end up with two skills that respond to the same trigger — and Claude picks one with no way for you to know which. The protection is governance: keep an inventory of what skills exist in your Cowork projects, name them clearly, and prefer team-owned skills over third-party installs for anything load-bearing. Same reason a real ops team doesn't let everyone install whatever Slack apps they want.
07 ·What's next
You've shipped the first hour. Three natural next moves, in roughly the order most operators take them.
Most operators get through this playbook on their own. The reason to bring in outside help — a Praxis Starter engagement, an internal AI lead, a peer who's further along — is when the setup needs to absorb a team rather than just you, when the skills need to encode workflows that matter to the business's ability to function, or when you've hit a governance question (data sensitivity, access control, audit) that you don't want to answer alone. None of that is required to get value from Claude. All of it becomes worth it once Claude is doing real work that the business depends on.