A thinking partner, not just a chatbot
Claude is an AI assistant guided by Constitutional AI — trained to be helpful, honest, and harmless, aligned with human values, steerable, and collaborative. It's available across web, desktop, and mobile, with conversations, projects, memory, and preferences syncing when you're signed in (Free, Pro, Max, Team, Enterprise).
Key concept · What Claude excels at
Writing & content (iterate on tone and structure), research & analysis (a large context window — 200K+ tokens, up to 1M on paid plans with supported models), coding (a core strength), problem-solving (with Thinking, it can reason step by step before answering), and learning (Learning mode guides your reasoning rather than handing over answers). Think partnership, not just Q&A.
Prompt, iterate, personalize
Talk to Claude like a coworker — naturally and conversationally. A good prompt does three things: set the stage (your role, objective, context), define the task (the action you want), and specify rules (style, tone, examples). Uploading relevant files gives Claude the context to work from.
Key concept · The iteration mindset (and evals)
Your first prompt is the start of a conversation, not a one-shot. Treat drafts as starting points, give specific feedback, and start fresh when a chat goes off track. Common fixes: too generic → add context; wrong length → be explicit; wrong format → show an example; confidently wrong → verify and enable web search; wrong tone → describe it plainly. To learn where Claude fits your work, run lightweight evals: gather 5–10 real examples, write prompts, compare outputs, refine. Memory saves context across chats; Styles customize how Claude writes.
Three shapes on the desktop — and how to give Claude structure
The desktop app sorts work into three shapes; recognizing which one you're in is the whole skill.
| Shape | What it is | Lives in |
|---|---|---|
| Working turn by turn | You and Claude go back and forth; the value is in the exchange. | Chat |
| Handing work off | You describe an outcome; Claude plans, does it across your tools, and returns a finished deliverable. | Cowork |
| Building software | Claude works in a codebase — reading, writing, testing, running commands. | The Code tab |
Key concept · Projects, Artifacts, and Skills
Projects are self-contained workspaces with their own knowledge base, instructions, and memory; they scale via RAG when knowledge grows, and can be shared with permission levels (view/edit/owner). Artifacts are standalone, interactive outputs (documents, code, HTML pages, SVGs, Mermaid diagrams, React components) in a window beside the chat — shareable and publishable. Skills are folders of instructions/scripts Claude loads to perform specialized, repeatable tasks (they power .xlsx/.pptx/.docx/.pdf creation). Rule of thumb: Projects store knowledge; Skills perform tasks.
Bringing your tools and the web into the conversation
Three features extend Claude beyond the chat box, and a family of products meets you where you work.
Key concept · Connectors, Enterprise Search, and Research
Connectors link Claude to your tools (Google Drive, Slack, Notion, Asana, and more) so it can read and act on your behalf — powered by the Model Context Protocol (MCP), "USB-C for AI." Claude only sees what you can see, and access is scoped and revocable. Enterprise Search (Team/Enterprise) adds an "Ask [Your Org]" project that synthesizes answers across your company's tools with citations. Research turns Claude into an agentic investigator: it plans, runs many searches that build on each other, synthesizes across sources (web + your integrations), and cites everything — minutes of work, not seconds.
Beyond claude.ai, the same intelligence appears in Claude Code (agentic coding), Claude Tag (in Slack), Claude Design (UI prototypes), Claude for Microsoft 365 (Excel, PowerPoint, Word, Outlook), and Claude in Chrome (browser sidebar). Each meets a specific workflow where it already happens.
Chapter 4 quiz
Sixteen questions on Claude, prompting, the three shapes, and extending reach. Answer first, then expand Show answer.
Part A · Multiple choice
What training approach guides Claude to be helpful, honest, and harmless?
- A Constitutional AI
- B Reinforcement-only learning
- C Keyword filtering
- D Manual moderation
Show answer
A — Constitutional AI.
What three things does a good prompt do?
- A Set the stage, define the task, specify rules
- B Greet, thank, close
- C Shorten, bold, bullet
- D Upload, wait, copy
Show answer
A. Context/objective, the action, and style/tone/examples.
Which capability lets Claude reason step by step before answering?
- A Thinking
- B Styles
- C Memory
- D Artifacts
Show answer
A — Thinking.
Handing off a multi-step task that ends in a finished deliverable happens in:
- A Cowork
- B Chat
- C The Code tab
- D Settings
Show answer
A — Cowork. Chat is turn-by-turn; Code is for codebases.
The rule of thumb for Projects vs. Skills is:
- A Projects store knowledge; Skills perform tasks
- B Projects perform tasks; Skills store knowledge
- C They're identical
- D Skills replace Projects
Show answer
A. The project provides the what; the skill provides the how.
What open standard powers connectors?
- A The Model Context Protocol (MCP)
- B SMTP
- C FTP
- D OAuth only
Show answer
A — MCP, described as "USB-C for AI."
When should you use Research rather than a quick web search?
- A When you need a comprehensive, multi-source report with citations
- B When you need one quick fact
- C When speed matters most
- D When the answer needs one source
Show answer
A. Research is agentic and thorough; a quick fact is a web search.
Which is an Artifact type?
- A A Mermaid diagram / React component
- B A Slack channel
- C A connector
- D A billing plan
Show answer
A. Also documents, code, HTML pages, SVGs.
Part B · True or false
Claude can only be used through the web app.
Show answer
False. Web, desktop, and mobile — plus Code, Slack (Tag), Design, M365, and Chrome.
Through connectors, Claude can access data you don't have permission to see.
Show answer
False. Claude only sees what you see; access is scoped and revocable.
Projects use RAG to scale when the knowledge base grows large.
Show answer
True. Claude searches and retrieves only what's relevant, expanding capacity ~10x.
Enterprise Search cites its sources.
Show answer
True. And only shows what you already have permission to access.
A first prompt is best treated as a one-shot request.
Show answer
False. Treat it as the start of a conversation and iterate.
Part C · Short answer
Name the three shapes of desktop work and where each lives.
Show answer
Turn-by-turn (Chat), handing off (Cowork), building software (the Code tab).
Name a lightweight way to test whether Claude fits one of your recurring tasks.
Show answer
Run a simple eval: gather 5–10 real examples, write prompts, compare Claude's outputs, and refine.
Name the three reach-extending features and what each is for.
Show answer
Connectors (link your tools via MCP), Enterprise Search (synthesize org knowledge with citations), Research (agentic multi-source investigation).