Most people start wrong. They open Claude in their browser, use it like a search engine, burn through their allowance, and wonder why they are hitting limits by Wednesday.
The desktop app is where the full feature set lives. Projects are more stable. Cowork is only available here. And you get a permanent sidebar so your work is always one click away.
Go to claude.com/download. Choose Mac or Windows, install, and sign in with your Anthropic account.
Which Claude product is which?
Claude Chat
The web or desktop interface. You type, Claude responds. Fast and flexible — the right tool for 80% of your daily work.
Cowork
Claude working on your actual files and computer while you step away. Requires the desktop app and the Max plan. Hand it a brief; come back to finished work.
Claude Design
New in April 2026. Creates visual assets — graphics, layouts, mockups. Still early for a consulting context, but worth watching.
Claude Code
Terminal-based. Builds and deploys software. Not what this guide covers — but worth knowing it exists if you need cloud-hosted automation.
Do this before you open a single chat. It is the highest-leverage action you can take to reduce wasted tokens and stop Claude asking the same clarifying questions every time.
Global instructions are read at the start of every conversation. They tell Claude who you are, how you communicate, and what it should never do. Without them, Claude starts from scratch every chat — and you pay for the back-and-forth every time.
Where to find them: Settings (bottom-left corner) → Profile → scroll to preferences or custom instructions.
The ACCURACY FIRST block — the most important addition
Claude, like all AI, sometimes states an opinion or best guess as though it were fact. This is one of the biggest frustrations people raise once they start using it seriously. Adding ACCURACY FIRST prompts Claude to flag uncertainty explicitly — so you stop treating confident-sounding wrong answers as reliable information.
If you only add one thing to your global instructions today, make it this.
Optional — sparring partner mode
Spend 15 minutes on this section. It will save you hours across every conversation you ever have with Claude.
Every Claude Pro plan has a usage allowance. Think of it like a mobile data plan. Regular chat is browsing the web: light. Cowork is streaming video in 4K: heavy.
What uses allowance faster
- Extended thinking on every message — Extended thinking is powerful but compute-heavy. It is the difference between asking someone to answer a question vs. asking them to show all their working. Only switch it on for tasks that genuinely need deep reasoning.
- Uploading large documents without summary instructions — When you upload a 40-page PDF, Claude reads the entire thing. If all you need is a summary or three specific data points, say so upfront. Otherwise you are paying for Claude to read and retain content you do not need.
- No global instructions — Without context, Claude guesses. Guessing leads to wrong outputs. Wrong outputs lead to corrections. Corrections cost tokens. Your global instructions from Step 2 eliminate most of this.
- Using Cowork when a regular chat would do — Cowork is compute-intensive. Using it to answer a quick question is like summoning a removal van to carry one cardboard box. Regular chat uses a fraction of the allowance for the same output.
- No "concise" instruction — By default, Claude is thorough. Thorough means long. Long means more tokens. Both your input and Claude's output count toward your allowance. Add "respond concisely" to your global instructions as a permanent default.
Four settings in the chat window most people miss
Inside any Claude chat, look at the toolbar beneath the input box. You can adjust response style (concise vs. detailed), disable web search when you do not need it, turn off extended thinking for straightforward tasks, and control which tools are active. Each of these affects how much allowance gets used per response. Turn off what you are not using.
Match the model to the task
Not every task needs the most powerful model. For straightforward reformatting, short summaries, or quick replies, a lighter model uses significantly fewer tokens. Save the heavier models for tasks that genuinely need deep reasoning. One small habit change that compounds over a month.
Monitor your usage at: claude.ai/settings/usage
A project is a container for related conversations with shared context. It includes project instructions (like global instructions but specific to that context) and a knowledge base (files Claude can reference).
What goes IN project instructions
- The single specific purpose of this project (one sentence max)
- What format, tone, or output type you always want here
- What Claude should never do in this project
- Any recurring context Claude needs (audience, constraints, situation)
What does NOT go
- A copy of your global instructions (those already apply everywhere)
- General Claude preferences (those belong in Settings)
- Vague instructions like "be helpful" or "work efficiently"
Real example — vague vs. specific
Too vague
"Help me with my marketing content."
Specific and usable
"This project is for drafting LinkedIn posts. All posts are written in first person, professional but warm, under 200 words. No hashtags unless explicitly requested. My audience is senior corporate professionals in financial services. Never use bullet points in a post."
The second version means Claude can act correctly on the first attempt. That is the goal.
Projects are your best token-saving tool
Each project has its own context window — its own memory of the ongoing work. When you keep a project focused on one type of task, that context stays relevant and Claude does not have to search through unrelated material to find what matters. One project per type of work is the rule.
Skills are specialist instruction files that live inside Cowork. They tell Cowork exactly how to handle a specific type of task — the steps, the context, the output format. You do not have to explain it every time.
Think of a skill as a Standard Operating Procedure for AI.
The one-skill-per-task rule
Do not create one giant skill that covers everything you do. It will be too long for Cowork to parse reliably, and performance will degrade. One skill per use case. Keep each skill focused on a single type of work.
Update a skill when the task type is the same but your process has changed. Create a new skill when it is a completely different type of work.
The three things that make Cowork different from chat
It works with your actual files
Cowork reads from and writes to real folders on your computer. You give it a brief, it processes your documents, creates outputs, and saves them — all without you staying in the window.
You can walk away
Regular Claude chat requires you to stay in the conversation. Cowork executes multi-step tasks while you do something else. You come back to finished work, not half a conversation.
It connects to your tools
Via plugins and connectors, Cowork can interact with apps like Notion, Google Calendar, Slack, and more. This is where it starts to feel genuinely autonomous.
Your computer needs to stay awake while Cowork is running.
Use Cowork for
- Organising or processing a folder of files
- Generating a formatted Word, Excel, or PowerPoint output from a brief
- Synthesising multiple documents into a single report
- Scheduled recurring tasks (a weekly digest, Monday morning summary)
- Any task that involves reading from and writing to your actual files
Do not use Cowork for
- Quick questions with one-line answers
- Single-step drafts you could write in a regular chat
- Anything that does not involve your files
- Tasks that must run 24/7 regardless of whether your computer is on
The before-state rule
Before you hand a task to Cowork, make sure the starting materials are already in the folder you have connected. Cowork needs something to work with. If the folder is empty, the session is wasted.
How to set up a scheduled task
- Run the task once manually first — Make sure it works correctly as a one-off. Fix any issues before you schedule it.
- Ask Cowork to schedule it — Tell Cowork: "I want to run this task every [Monday morning / day at 9am]. Set up a schedule." It will walk you through the options.
- Keep your computer awake at the scheduled time.
Answer these five questions. Then copy your answers and paste them into a new Claude chat with the prompt below. Claude will recommend the three most relevant projects to create right now, with specific instructions for each one.
The five questions
- What is your job title or main role?
- What are the three tasks at work that eat the most of your time each week?
- Do you regularly work with documents, spreadsheets, or slide decks?
- Do you need Claude to write or communicate in a specific voice or tone on your behalf?
- Are you using Claude on your own, or will others on your team use it too?
Copy this prompt into Claude
Next step
Ready to go further?
You have the setup. Now learn how to get 10 hours a week back. Claude Cowork Workshop — live, 90 minutes, no jargon, no slides full of definitions. Just hands-on workflows you can use the next morning.
See the workshop →