The Complete Guide to Claude Chat, Cowork & Claude Code Desktop

1. Overview — Three Tools, One Ecosystem

Anthropic ships Claude across three distinct surfaces. Each has a different purpose, a different interface, and a different kind of user in mind. Using the right tool for the task is the first decision that determines how useful your session will be.

These three tools are not interchangeable. Choosing the right surface for the task makes the difference between a productive session and a frustrating one.

2. Claude Chat — claude.ai

Claude Chat is the web interface available at claude.ai. It is conversation-based, requires no installation, and works in any modern browser. It is the best starting point for thinking through problems, drafting content, conducting research, and exploring ideas.

2.1 Choosing a Model

You can switch between Claude models at the top of any conversation, depending on the complexity of your task:

Switch models freely within a session. There is no penalty for switching mid-conversation.

2.2 Extended Thinking

On supported models, Claude can reason through a problem step by step before giving its final answer — making its chain of thought visible. Enable it via the toggle before sending a message.

When to use it:

• Multi-step logic or math problems
• Strategic decisions with competing trade-offs
• Complex analysis where the reasoning matters as much as the conclusion
• Debugging faulty logic in your own thinking

Extended Thinking slows down the response. Reserve it for tasks where depth matters more than speed.

2.3 Artifacts

When Claude generates structured content — code, a document, an HTML page, an SVG diagram — it renders it in a separate side panel called an Artifact. The Artifact is independently editable and can be copied or downloaded without re-reading the conversation.

Supported content types:

• Code — any programming language
• Markdown documents
• HTML pages (rendered live in the panel)
• SVG diagrams
• React components (interactive preview)

Tip: Ask Claude to update a specific section of an Artifact rather than regenerating the whole thing. It is faster and keeps changes scoped.

2.4 File Uploads

Attach files directly to any message. Claude reads and reasons about their contents.

2.5 Web Search

Claude can search the web in real time and cites its sources. It activates automatically when you ask about recent events, current data, or anything beyond its training cutoff.

Useful for:

• News, current events, recent product releases
• Live pricing, statistics, or market data
• Verifying facts that may have changed recently

When Claude searches, it shows the sources alongside the answer so you can verify them independently.

2.6 Voice Mode

Available on the Claude mobile app (iOS and Android). Speak your prompt and receive a spoken response. Useful for hands-free brainstorming, thinking out loud while commuting, or drafting rough ideas away from a keyboard.

2.7 Memory

Claude can remember facts about you across separate conversations — your role, preferences, recurring projects, and working style. You can view and manage what Claude knows via Settings → Memory.

What you can do:

• Ask Claude what it currently remembers about you
• Tell it to remember a specific preference or fact
• Ask it to forget anything you no longer want stored

Memory is rolling out progressively. If it is not yet available in your region, you can replicate it using Projects (Section 3).

2.8 Practical Tips for Claude Chat

• Be specific about format. “Give me a bullet-point summary in under 150 words” produces better output than “summarise this.”
• Iterate in the same conversation. Context accumulates — Claude’s answers improve as you build on earlier messages.
• Edit your last message. Press the ↑ arrow to revise and re-run your last prompt instead of sending a corrective follow-up.
• Paste long text directly if it is under ~100k characters. Upload as a file if it is larger.
• Name your conversations. A clear title in the sidebar makes previous sessions much easier to find and resume.

3. Claude Cowork — Projects & Team Collaboration

Projects is the persistent-context layer built into claude.ai. Instead of starting fresh with each conversation, you give Claude a stable knowledge base that remains active across every session within that project. Think of it as the difference between briefing a contractor from scratch every morning versus having a permanent team member who already knows your work.

Available on Pro, Team, and Enterprise plans.

3.1 Setting Up a Project

To create a Project:

• Click Projects in the left sidebar on claude.ai
• Click New Project
• Add a name and short description
• Write your Project Instructions (see below)
• Upload Project Knowledge files

3.2 Project Instructions

Project Instructions are a persistent system prompt. Claude reads them at the start of every conversation in the project. They are the most powerful lever available — use them to define Claude’s role, tone, constraints, and background knowledge.

What to include:

• Who Claude should be in this project (role, persona, expertise level)
• Tone and style preferences (formal/casual, short/detailed)
• Persistent facts about your product, company, or audience
• Things Claude should never do or say

Example instruction block:

You are a content editor for a B2B SaaS company called Meridian.

Our audience is operations managers at mid-size companies.

Always write at a reading level accessible to non-technical professionals.

Prefer short paragraphs. Use headers for anything over 200 words.

Never use the phrase “game-changer” or “revolutionary.”

Our product URL is meridian.io. Do not recommend competitor products.

3.3 Project Knowledge

Upload files that Claude should treat as permanent background knowledge. It draws on these in every conversation without you needing to re-paste content.

Keep Knowledge files updated. Claude reads them as truth — outdated documents lead to outdated answers.

3.4 Team Collaboration

On Team and Enterprise plans, Projects become shared workspaces. Every team member accesses the same Instructions and Knowledge, so everyone works from identical context.

What teams can do:

• Share a Project with specific teammates or the entire organisation
• Set permissions — who can edit Instructions and Knowledge vs. who can only use the Project
• Share individual conversations with teammates for review or handoff
• Build a living knowledge base that improves over time

Practical team use cases:

3.5 The Right Mental Model

Project = the permanent briefing. Conversation = the individual work session.

Every conversation inside a Project inherits the Instructions and Knowledge automatically. You bring the task; Claude already knows the context.

4. Claude Code Desktop

Claude Code is Anthropic’s engineering-focused tool. It runs as a CLI (command-line interface), a native Desktop application on Mac and Windows, and as extensions inside VS Code and JetBrains IDEs. Unlike Claude Chat, it has direct access to your filesystem, terminal, and development environment.

It is purpose-built for: writing code, editing files, running terminal commands, managing git history, and executing multi-step engineering workflows autonomously.

4.1 Installation

CLI (terminal-first):

npm install -g @anthropic-ai/claude-code

claude

Desktop app: Download from claude.ai/code — wraps the CLI in a native window with a visual interface.

IDE Extensions: Available for VS Code and JetBrains. Claude Code runs inside your editor without switching windows.

4.2 File Operations

Claude Code reads, writes, and edits files on your machine with precision:

• Read — reads any file you point it to before making changes
• Edit — surgical edits that replace specific strings, not whole files
• Write — creates new files when needed
• Search — scans every file in a directory for a pattern or keyword

Claude Code will never overwrite a file without reading it first. This is enforced by design.

4.3 Terminal & Bash Execution

Claude Code executes shell commands on your behalf — installing packages, running tests, querying databases, managing processes. It will request confirmation before running commands that are destructive or hard to reverse.

Examples of what it runs autonomously:

• npm install / pip install
• Test runners (npm test, pytest, go test)
• Build scripts (npm run build, make)
• Linters and formatters

Examples of what it asks permission for first:

• git push / git reset –hard
• rm -rf or any deletion commands
• Database migrations
• Anything that affects shared infrastructure

4.4 Git Integration

Claude Code treats git as a first-class tool. It reads your history to understand the codebase before acting, and it writes commit messages that match your project’s existing style.

• Reads git log and git diff before making changes — understands what has changed and why
• Stages specific files (never git add -A blindly)
• Writes commit messages by analysing your recent commit style
• Creates and switches branches
• Opens pull requests via the gh CLI

Always review Claude’s staged files before committing. Claude is deliberate — but you are responsible for what ships.

4.5 Multi-File & Codebase-Wide Tasks

Claude Code does not work file by file in isolation. It can search, trace, and modify code across an entire codebase:

• Grep — find any function, variable, or string across all files
• Glob — find files by name pattern (e.g., all *.test.ts files)
• Codebase-wide refactors — rename a variable or function everywhere it appears
• Feature tracing — follow how a feature works end-to-end across multiple modules

4.6 Plan Mode

For large or complex tasks, Claude Code enters a planning phase before touching any file. It maps out what it will change and why — you review and approve before execution begins.

How to trigger it:

• Type /plan before your request
• It also triggers automatically when Claude detects a complex multi-step task

Use Plan Mode for: large refactors, feature implementations spanning multiple files, anything you cannot easily undo. It prevents surprises.

4.7 CLAUDE.md — Project Instructions

Every project can have a CLAUDE.md file at its root. Claude Code reads this at the start of every session — it is the equivalent of Project Instructions but for engineering projects.

What to put in CLAUDE.md:

• Project architecture overview
• Tech stack and key dependencies
• Coding conventions (naming, file structure, patterns to follow/avoid)
• Commands to run tests, lint, and build
• Folders or files Claude should never modify

Example CLAUDE.md:

# My Project

## Stack

Node.js + PostgreSQL + React

## Conventions

– camelCase for variables and functions

– No console.log in production code

– All DB queries go in /src/db/

– Tests must be written for every new function

## Commands

– Test: npm test

– Build: npm run build

– Lint: npm run lint

## Do Not Modify

– /migrations/ (handled by DBA team)

– /config/secrets.js

4.8 MCP — Model Context Protocol

MCP allows Claude Code to connect directly to external tools and data sources — not just your local filesystem. Once configured, Claude can query databases, call APIs, and interact with third-party services in any session.

Supported integrations (examples):

• Databases — PostgreSQL, SQLite, MySQL (run queries directly)
• Dev tools — GitHub, Linear, Jira (read issues, create PRs)
• Productivity — Slack, Notion, Google Drive
• Custom tools — any internal API you expose via an MCP server

Configure MCP servers in ~/.claude/settings.json under the mcpServers key.

4.9 Hooks — Automated Behaviours

Hooks are shell commands that fire automatically in response to Claude Code’s actions. They enforce consistent behaviour without you needing to remind Claude each time.

Configure hooks in ~/.claude/settings.json.

4.10 Slash Commands

4.11 Background Agents

Claude Code can spawn parallel subagents to run independent tasks simultaneously. You continue working in the main session while background tasks execute, and you are notified when they complete.

Example: a large feature request might spawn three agents in parallel:

• Agent 1 — explores the codebase to understand the existing structure
• Agent 2 — writes tests for the new feature
• Agent 3 — checks documentation for related functions

Background agents are most valuable for research and exploration tasks that do not depend on each other. Claude manages their coordination automatically.

4.12 Worktree Isolation

For risky or experimental changes, Claude Code can work in an isolated git worktree — a separate copy of your repository. Changes are contained and cannot affect your main branch until you explicitly approve and merge them. The worktree is automatically cleaned up if no changes are made.

Use worktree isolation for: large refactors, experimental features, any change you are not confident about before reviewing.

4.13 Settings — ~/.claude/settings.json

The global settings file controls Claude Code’s behaviour across all projects.

Project-level settings.json overrides the global config for that project only — useful for project-specific permission rules.

5. How the Three Tools Work Together

The real power of Claude’s ecosystem comes from combining all three surfaces into a single workflow. Each tool handles a distinct phase of the work.

The Core Workflow Pattern

Claude Chat → Think, explore, draft, decide

Claude Cowork → Store context, align the team, maintain knowledge

Claude Code → Build, edit, test, ship

Workflow Example 1 — Building & Documenting a Feature

• Claude Chat — Brainstorm the feature design. Explore trade-offs. Write the technical spec.
• Claude Projects — Upload the spec + coding standards as Project Knowledge. All future engineering sessions reference it automatically.
• Claude Code Desktop — Implement the feature. Claude reads the spec from CLAUDE.md, writes the code, runs tests, commits with a proper message.

Workflow Example 2 — Content Team Publishing Pipeline

• Claude Projects — Brand guidelines, tone-of-voice doc, audience profiles, and top-performing past pieces live here permanently.
• Claude Chat — Writers open a conversation inside the Project. Claude already knows the brand, audience, and style — no re-briefing needed.
• Claude Code Desktop — If content lives in a static site or headless CMS with an API, Claude Code can push drafts directly via the command line.

Workflow Example 3 — Onboarding a New Team Member

• Claude Projects — Company handbook, role description, tech stack overview, and process guides are uploaded as Knowledge.
• Claude Chat — New hire chats inside the Project. Asks questions, gets answers grounded in real company documentation.
• Claude Code Desktop — New hire sets up their dev environment. Claude reads the CLAUDE.md, installs dependencies, runs the project, explains the codebase.

6. Quick-Reference Cheat Sheet

Claude Chat

Claude Cowork (Projects)

Claude Code Desktop

Closing Principle

Claude works best when it has context. The more deliberately you provide that context — through Projects, CLAUDE.md files, clear conversation framing, and well-configured settings — the less you need to repeat yourself, and the more every session compounds on the last.

The three tools are not separate products. They are three phases of the same workflow: think, align, build. Used together, they replace a significant portion of the coordination overhead that slows teams and individuals down.

Start small: pick one tool, configure its context layer (Instructions or CLAUDE.md), run one real task through it, and observe the difference. Then expand from there.