A primer
The landscape, the concepts worth knowing, and a prompt you can paste to set up your own context.
Most people who try Claude and bounce off it don't have a bad tool — they have a blank context. Claude doesn't know who you are, what you work on, or how you like to be talked to. Every conversation starts from scratch. Twenty minutes of setup changes this, and this primer walks you through it.
There are so many terms… Claude Chat, Code, Cowork, MCPs, connectors, skills, plugins, CLAUDE.md. Most explainers make these sound more complicated than they are. Here’s the map, grouped by when to touch each piece. Start at the beginning, and reach for the “when you’re ready” items as concrete needs arise. Everything else can wait until you have a specific reason to use it.
Where almost everyone should start.
claude.ai or the mobile app. Conversations, file uploads, artifacts, web search, connectors to things like Gmail and Notion. If you’re learning Claude, don’t jump to anything else yet.
The single highest-ROI thing most people never set up.
A block of text in your settings that tells Claude who you are and how you want to be talked to. Applied to every new conversation. This is what this primer is mostly about.
Claude that works directly in your files. Worth a look once chat starts to feel limiting.
It lives in your terminal, reads and writes your real documents, runs commands, and remembers context between sessions. Long treated as developer-only, but increasingly useful for anyone with a folder of work to manage. See “Going further” below if you’re curious.
A desktop app that does multi-step tasks for you.
Claude with hands, in a friendly desktop app — no terminal required. It can operate on files, manage documents, and run multi-step workflows. Worth evaluating when you have a recurring task you’d love to hand off.
A way to group related conversations and share context.
Each project has its own instructions and can hold reference files. Useful once you have an ongoing body of work — a book, a client, a class, a long-running idea.
Permissions that let Claude reach into your other tools.
Gmail, Google Drive, Notion, Slack, GitHub. The “MCP” acronym you’ll see everywhere is just the behind-the-scenes protocol. Add connectors one at a time when you have a specific need, not preemptively.
A plain text file that gives Claude lasting memory of your work. Anyone can use it.
It holds the context you’d otherwise re-explain every session: who you are, what you’re working on, how you like things done. It started in Claude Code, which earned it a developer-only reputation it doesn’t deserve. The file is plain English. If you can write a note for a new assistant, you can write a CLAUDE.md. See “Going further” for how it becomes the backbone of a system that remembers.
Reusable playbooks Claude loads automatically when relevant.
“Here’s how my team writes case studies.” “Here’s how I format client reports.” Worth building when you catch yourself explaining the same process more than twice. Most people don’t need these for months.
User preferences are the closest thing Claude has to “knowing you.” They do two things:
What you do, what projects you’re juggling, what vocabulary is natural for you, what you’re trying to learn. Without this, the tool defaults to an imagined average user, or basically nobody. And you are not nobody.
Length, tone, formality. Whether to push back on your ideas or just execute. Whether to ask clarifying questions before diving in. How to handle uncertainty. A lot of AI frustration stems from answers in the wrong lane. They aren’t bad responses, but they aren’t what you need.
A useful way to split them: contextual preferences (who you are) and behavioral preferences (how you want the tool to respond). The starter prompt at the bottom of this page covers both.
Here's a condensed version of what a mature set of preferences looks like — from someone who uses Claude across several unrelated areas of work. Don't copy this. Your life is different. But it's useful to see the shape of something that works before you build your own.
Workstreams & Context
I work across several distinct, usually unrelated areas. Treat each conversation's context as set by me — don't carry assumptions across workstreams unless I explicitly invite it.
— Client work (strategy and BD)
— Board work with a nonprofit community
— A climate-focused nonprofit where I sit on the board
— A class I'm taking on AI and climate
Before executing, confirm intent
If my request is ambiguous, ask a clarifying question first. I'd rather refine my request than rework a full response.
Why this exists: she'd rather spend 30 seconds clarifying than 20 minutes reading an answer aimed at the wrong problem. This single line changes most interactions — without it, Claude defaults to guessing.Honesty over agreeableness
I prefer honest and direct over agreeable. Don't soften bad news or overvalidate. Push back especially on strategic assumptions, underestimating complexity, and solving the wrong problem.
Why this exists: AI tools default to sycophancy. Without a line like this, you get affirming mush. This is probably the single most valuable behavioral preference you can write.Work modes
Calibrate to the mode I signal: Brainstorming — poke holes. Stress-testing — balanced analysis. Drafting — I've decided; help me execute.
Why this exists: the same request (“help me with this proposal”) needs very different responses depending on whether she's still thinking or already committed. Naming the modes gives her a shorthand to switch registers without re-explaining each time.Notice what's not here: no lists of favorite colors, no hobbies, no life story. Preferences that don't change how the tool responds to you are noise. Keep it lean.
Rather than giving you a template to fill in — which is where most people bounce — the prompt below runs the interview for you. Paste it into a new conversation. The AI will ask you a series of focused questions, one at a time, then draft your preferences for you to edit.
You can come back and revise later. Your first version won't be your last. The goal is to stop starting from scratch.
Claude will output a block of text. Go to your Claude settings — in the web app, click your initials bottom-left, then Settings → Profile — and paste it into the “What personal preferences should Claude consider in responses?” field. Save. You're done. Every new conversation will now start with Claude knowing who you are.
Most people never need to leave the chat app, and this section is safe to skip until you do. The signal that you’ve outgrown it is specific: you catch yourself pasting the same files into every new conversation, or you want Claude to actually run something on your computer instead of just describing how. When that happens, two tools pick up where chat leaves off.
| Cowork | Claude Code | |
|---|---|---|
| Built for | Everyday knowledge work, no terminal | Files, automation, anything on your machine |
| Lives in | A friendly desktop app | Your terminal (also IDE + desktop) |
| Best at | Multi-step tasks, docs, research, ops | Reading and writing real files, running commands, remembering context |
| It’s for you if | You want Claude to do tasks, not just answer | You want Claude to live in your folders and pick up where it left off |
Cowork is the gentler entry to “Claude with hands.” Code is more powerful and more direct, and once carried a reputation for being developer-only. That reputation is fading. If you keep a folder of work you return to, Code earns its place.
npm install -g @anthropic-ai/claude-code, and run claude inside any folder.cd into a project or working directory. Everything Code does is rooted in the folder you launch it from.CLAUDE.md. Run /init and Code drafts one by reading the folder. This file is the persistent-context unlock — the single biggest difference from chat.Stop asking Claude questions. Start handing it tasks in a folder it can see.