AI for complete beginners: what to install on your computer and phone, and the difference between chat, Cowork, and Claude Code

Most people today know AI in exactly one form: an open browser tab you occasionally type something into. It works, it sends something back — and that's where it ends. Meanwhile, people talk about apps, about Cowork, about Claude Code, about connectors. The question most beginners are left alone with is entirely practical: what am I actually supposed to install, where, and why is there so much of it?

The key to the whole mess fits into one sentence. You're not learning a new program — you're learning that one and the same model reaches you through several different windows, and each of those windows lets it see something different. Chat sees only your conversation. Cowork sees a folder on your disk. Claude Code sees an entire project and can run commands. Once you understand that, choosing between them stops being a question of technical skill and becomes an ordinary question of “where does the thing I need to work on actually live.”

This guide assumes no technical knowledge — whenever a word like terminal or connector shows up, an explanation of what it is and where to find it sits right next to it. The prompts are ready to copy; just fill in the brackets.

A typical scenario

Marcela does the books for three small companies. She's known AI from the browser for six months: she occasionally has it rephrase an awkward email to a client. She uses one tab among fifteen others and regularly closes it by accident, conversation and all. She has nothing on her phone. She heard about Cowork from her son, who told her “that's the one,” but couldn't explain why. Meanwhile, once a month a client hands her an envelope with fifty receipts she has to transcribe.

Over one evening, this is what changes for her. An app on her computer means conversations don't live in a tab and open with a keyboard shortcut. An app on her phone means she can photograph a scanned contract right there over coffee and ask “pull out the deadlines, penalties, and notice period.” She scans the receipts into one folder, and in Cowork they turn into a spreadsheet in twenty minutes that she now only has to check. And she stops hunting for “where does Claude draw pictures,” because she discovers it doesn't. None of this is complicated. It's just that nowhere lays it all out in one place.

One thing first that makes all of this simpler

Before you install anything, separate two terms that everyday speech blurs together. Without this, the rest of the guide won't make sense.

The model is the thing that thinks. A huge program running on the provider's servers, not on your device. You never download it and never actually see it. Your message flies to a server, the model works out an answer there, and it comes back. It's one and the same model whether you're asking from a browser, a phone, or an app.

The interface is the window you talk to it through. A browser, a mobile app, a desktop program, a window with a command line. The interface doesn't think — it just delivers the message and determines what extra information the model gets and what it's allowed to do with the result.

Picture a meeting with a capable colleague. Meet them in a café and hand over papers in a folder — they'll read it and nothing more. Invite them into your office and show them a shelf of documents — they can leaf through it and write into it. Or hand them the keys to a workshop where they're allowed to switch on the machines too. Same colleague throughout; only what's within reach changes.

That gives you the rule for the rest of this guide: choosing an interface isn't a question of “which is better,” it's “where does the thing I need to work with actually live.” And one more practical point: you only need one account. Create it on the web, sign in with it in the app and on your phone, and you'll see the same conversations everywhere — finish on your computer in the evening a conversation you started on your phone.

Installation: computer

What's the point of an app when you already have the same thing in a browser? It gets its own window that doesn't get lost among thirty tabs, and you open it with a keyboard shortcut — asking something takes two seconds instead of twenty. And most importantly: for Claude, it's a prerequisite for Cowork, meaning work over a folder on your disk. A browser can't reach your files for security reasons, and that's as it should be.

Claude on a computer, step by step

1. Open claude.ai. Type it into the address bar by hand. Don't search for “claude” and click the first sponsored result — popular brands attract lookalikes in ads.
2. Create an account with email or by signing in with Google. With email, a verification code lands in your inbox that you copy into the window; it's only valid for a few minutes.
3. Download the app. The link sits right on claude.ai, and the page usually detects on its own whether you're on a Mac or on Windows. Only download from here.
4. On a Mac, a file with a .dmg extension downloads. Double-click it and, in the window that opens, drag the Claude icon into the Applications folder. You can then delete the downloaded file — it's just packaging.
5. On Windows, run the downloaded installer and follow the wizard. Windows may show a blue system-protection screen — click “More info” and confirm the publisher is Anthropic.
6. Sign in with the same account as on the web, and you'll see the same conversations. Then pin the icon to the Dock or taskbar, and find the keyboard shortcut setting to bring the window up instantly — a small detail that decides whether you use the tool daily or once a week.

ChatGPT on a computer

The process is the same; create your account at chatgpt.com. An official desktop app exists for both macOS and Windows; on Windows it installs from the Microsoft Store, the app store built into Windows itself — that's the official and safe route. For macOS, check the system requirements on the download page first: newer versions need a newer OS, so on an older Mac you'll stay with the browser.

When installation doesn't work

It asks for an administrator password. On a work laptop you usually don't have one, and that's intentional — ask IT whether the app is approved. The antivirus warns about an unknown publisher. Check where you downloaded the file from — a warning from the official site is routine, a warning from a link in an email is a reason to delete the file. And if there's no version for your system, no problem: a browser can do everything except work over a folder.

And if you're not sure what a screen is asking of you, just ask right there in the browser — describe what you see and have it explained. Just one caution: if the answer includes a specific download address, don't click it blindly. Compare it against the official site you open by hand. Models sometimes make up addresses, and downloading an installer from a made-up address is exactly the mistake you don't want to make.

Installation: phone

A mobile app installs more easily, because the app store does almost everything for you. The snag lies elsewhere — in the subscription.

1. Open the App Store (iPhone) or Google Play (Android) and search for “Claude” or “ChatGPT.”
2. Check the publisher, not the icon. For Claude it should list Anthropic; for ChatGPT, OpenAI. This is the single most important step of the whole installation: the stores are full of apps with similar names and icons, things like “AI Chat Assistant Pro,” that charge extra money for the same thing. A good signal: official apps have hundreds of thousands to millions of ratings; lookalikes have hundreds.
3. Sign in with the same email you registered with on your computer. You'll see the same conversations.
4. Allow access to the microphone and camera when the app asks — without them, neither dictation nor photographing documents will work, and those are the two things you mainly want the phone app for. Access can be revoked anytime in system settings.
5. Put the icon on the home screen, in the bottom row your thumb reaches. A tool you have to tap through three screens to reach stops getting used within a week.

Subscriptions paid through an app store

This is where beginners get confused most often. When you pay for a subscription inside the mobile app, the payment isn't processed by the AI provider but by Apple or Google. There are three consequences. First: you then cancel the subscription not in the app, but in your Apple ID or Google account's subscription settings — people search in vain for a “cancel” button that isn't there. Second: the receipt and any dispute go through the store. Third, and most painfully: if you accidentally pay for a subscription once on the web and again in the app, you're paying twice, and the payments don't cancel each other out.

Recommendation: pick one place and stick with it. And before you pay, check that you're signed in under the same email as on your computer — the plan is tied to the account, not the device. Anyone who accidentally creates a second account via “sign in with Google” ends up paying on one and using the other; you'll notice because your phone won't show the conversations from your computer.

Why have it on both computer and phone

The point isn't having the same thing twice. Each device can do something the other can't, and together they cover two completely different kinds of situations.

Your phone is your eye and ear. It has a camera and a microphone and is always with you:

• Photograph a document. A receipt, a whiteboard covered in notes after a meeting, a handwritten note, a label, a poster with deadlines. Claude and ChatGPT both read images, including handwriting and scans, turning a photo into text, a table, or a task list. More detail in the tip from photo straight to action.
• Dictate a long brief on the go. On the walk back from a meeting, you can speak in two minutes what would take ten to type.
• Ask a quick question in line or on the train. A short question you'd otherwise forget to ask.

Your computer is your workshop. It has a keyboard, a big screen, and access to files: long texts you can't edit on a phone; work with folders, since Cowork only exists here; two windows side by side; and a more thorough brief, because you only write a ten-line prompt on a keyboard — on a phone, you settle for one sentence, which is the most common reason for weak answers.

The rule of thumb: collect on the phone, process on the computer.

Chat, Cowork, and Claude Code: what's what

This is the heart of the article and the point of the most confusion. These aren't three different AIs, and they aren't three levels of user skill. It's the same thinking operating in three workspaces that differ in what the model has within reach and what it's allowed to do with it.

Chat: the simplest, and plenty for most work

What it is. A window with a conversation. It runs on the web at claude.ai or chatgpt.com, in the mobile app, and in the desktop app — it's the same thing everywhere.

What it sees. Exactly and only what you type or upload into the conversation: text, a document, a spreadsheet, an image. It can't see into your computer; there's no way for it to. And it doesn't change anything — it returns text that you copy yourself.

Who it's for. Everyone, and an estimated ninety percent of everyday work. Anyone who claims chat is only for beginners is wrong.

What it handles right here. A quick answer and a first draft of a text, a summary of an uploaded document, an explanation of something you don't understand. Plus two extras: Projects — persistent context you upload reference material into once (a price list, a communication style, a company description), and the model draws on it across every conversation in that project. And artifacts — output that shows up as a separate document or a small working app, like a simple calculator.

I'm preparing for [a meeting with three suppliers about shipment
delays]. I'm pasting in my messy notes:

[paste your notes as they are — bullet points and typos are fine]

Turn this into:
1. a 30-minute agenda with a time breakdown of the points
2. one sentence per point on what I want to achieve
3. three questions I have to ask even if time runs out
4. a list of information I need to gather beforehand

Keep it short and to the point. Where something's missing from
my notes, ask rather than guess.

It returns a usable brief in about a minute. That last sentence is there on purpose: without it, the model fills gaps in your notes with a plausible-sounding invention.

Cowork: working over a folder on your disk

What it is. A mode in the Claude desktop app where Claude works over a specific folder on your disk. It isn't a separate program you install on its own — it's a way of using the app from the previous section. There's no Cowork in the browser, and there can't be.

What it sees. The files in the folder you point it to, and its subfolders. Nothing else on the disk — no passwords, no email, no other folders. It can read them, edit them, and create new ones. That's the key difference from chat: you don't have to copy the output, it becomes a file right next to your source material.

Who it's for. Anyone whose work is spread across files. It requires no technical skill, just the willingness to organize a folder ahead of time.

What it handles that chat can't. Transcribing fifty scanned receipts into one spreadsheet. Going through a folder of thirty source documents and writing a summary report. Checking twenty invoices against a list of orders. There's no problem in the thinking — the problem is that you couldn't fit it into a single message.

Claude Code: the command line, and the widest reach

First, what a terminal is. A terminal (on Windows it's called Command Prompt or PowerShell) is a window where, instead of clicking, you type commands and get answers back as text. It looks stark, and plenty of people only know it from movies. In reality it's just another way to give the computer an instruction — instead of “click, click, confirm,” you type one line. On a Mac, press Cmd and Space and type “Terminal.” On Windows, press the Windows key and type “Terminal” or “PowerShell.”

What it is. Claude Code is a tool that runs right in that window. It can work with files, run scripts, and work with git (a system that remembers every version of your work and can roll it back). It also has extensions for editors like VS Code or the JetBrains IDEs.

What it sees. The entire project folder you launch it in — and it can also run commands. The widest reach of the three, and the responsibility to match.

Who it's for. It isn't a tool just for programmers, and that's the most common myth putting people off unnecessarily. It's for anyone willing to type commands instead of clicking — that willingness is the only ticket in.

What it handles right here. Tasks that need to repeat, run, and be watched over: renaming three hundred files based on a date inside their content, converting a hundred documents from one format to another, a script that recalculates something every week, versioning work you need to roll back three days on.

Start with a task where nothing can go wrong — just have it describe what it sees:

I'm a complete beginner, this is my first time working with a
command line. In the folder you're running in, I have [three
years of scanned invoices].

For now, don't change, delete, or rename ANYTHING. Just look around
and write down:
- how many files are here and of what types
- what naming pattern they follow, if any
- which files look like duplicates
- what you'd suggest renaming or reorganizing, and why

At the end, spell out exactly what steps you'd take if I approved
them, and which of those couldn't be undone. Explain it in plain
language, I'm a layperson.

It returns an inventory of the folder and a plan you then approve. That last paragraph is there for a good reason: AI proposes, a human approves. That goes double for deleting and renaming, because some steps can't be undone. Before you run the first task that actually changes something, make a copy of the folder.

A side-by-side comparison

                     CHAT                COWORK              CLAUDE CODE
Where it runs        web, phone,         desktop app         terminal
                     desktop app                             (+ editor)

What it sees          just the           one folder you      the whole
                     conversation and    point it to         project folder
                     uploaded files

Changes files         no                 yes, within         yes, plus can
                                          the folder          run commands

What you need to      nothing            organize a          type commands
know                                     folder

Typical task           email, summary,   50 receipts         rename 300
                       explanation       into a table        files, a script

Risk of a mistake      low                medium              higher
                       (text only)       (changes files)     (runs commands)

How to choose

Three questions, in this order:

1. Does everything the model needs to know fit into one message? Can you paste in text, upload a document, describe the situation? Then use chat — and don't look for anything more complicated.
2. Is it scattered across files on your disk? Ten, fifty, or three hundred files you'd spend half an hour uploading to chat? Then use Cowork.
3. Does it need to repeat, run, or be version-controlled? Then use Claude Code.

And a rule on top of all this: always start at the top. Beginners typically do the opposite — they read that Claude Code is the most powerful and start there. The result is frustration with a tool solving a problem they don't have. Three months in chat is better preparation for Cowork than three hours reading about Cowork. How these levels build on each other is covered in the chapter five levels of working with AI.

How to connect your accounts

What a connector is, in plain language. A permission slip you give the AI to look into another one of your services — email, calendar, drive, notes. Without it, it knows absolutely nothing about your calendar. With it, you ask “what do I have this week” and get an answer pulled from your actual calendar. Technically this kind of connection is called MCP; it's a shared standard so every service doesn't have to invent its own way of connecting, which is why more of them keep appearing and why you can add custom ones too. For Claude, the common ones are Gmail, Calendar, Drive, Notion, Slack, and plenty more from the connector directory; ChatGPT has similar connections to storage and work tools, and they're especially useful in deep research, where it can pull in your internal documents too. An overview is in the tip connectors and tools.

How a connector gets added. In settings, open the connectors section (it might be called Connectors, Connections, or Apps), pick the service from the list, and click connect. A window from that service's own provider opens — Google, Microsoft, Notion — where you sign in with your regular account. That window belongs to them, not the AI; you enter your password to them. Then a permissions screen appears; you confirm, and the connector is active.

What the permissions screen means. A list of phrases like “view your emails,” “manage your calendar events,” “upload and delete files.” Most people skip it and just hit allow — and yet it's the one place where the real decision gets made. The key distinction is between the verbs: view and read means the AI can look; manage, edit, send, and delete means it can act on your behalf.

So allow read-only access for the first two weeks. You'll see whether the connection is even useful to you, and nothing can break. The worst case with read access is a bad answer; with write access, it's a sent email you never approved. Only add write permissions once read access is actually paying off — and even then, the rule stands: AI proposes, a human approves.

You have my calendar and email connected, read-only for now.
Don't send anything, don't create anything, don't cancel anything
— just put together an overview for me.

Go through my calendar for the next seven days and my email from
the last three days:
1. what I have firmly scheduled this week, chronologically, with times
2. which meetings the email suggests I should prepare something for
3. which emails I clearly haven't replied to and someone's waiting on me
4. where two events overlap, or where I have less real time between
   locations than the schedule implies

For each point, note which specific event or email it came from,
so I can verify it.

It returns an overview that replaces fifteen minutes of clicking around. That last paragraph matters: without it, you get claims with no source, and you can't tell whether the model imagined an item.

Connecting anything is always a decision about trust. For a work account, ask IT before you enable anything. Sensitive data (clients' personal information, health data, payroll) belongs only in a paid account with a data protection agreement, and even there, anonymized is better. And knowing how to turn it off is part of turning it on: a connector is removed wherever you added it, and you can also revoke access at the provider — for Google, in account settings under security. The chapter AI, ethically and safely goes deeper into this.

How to point Claude at a folder

Working over a folder is the single most useful thing you won't get in a browser. In practice, it means picking one folder on your disk in the app, and that folder becomes your desk.

What it sees. The files in the chosen folder and its subfolders — names and content. It can read documents, spreadsheets, PDFs including scans, and images. What it doesn't see: the rest of the disk, other folders, passwords saved in your browser, your email (unless you've added that connector). The scope is exactly what you define.

How to prepare a folder. It takes ten minutes and it's the entire secret to a successful Cowork session:

• One folder per task. Not your whole “Documents” folder, not the desktop. A model rummaging through a thousand unrelated files answers worse — and sees things it doesn't need to.
• Work on copies, not originals. At least at first.
• Clear file names. invoice-2026-03-vodafone.pdf is far better for the model, and for you, than scan0043.pdf.
• Separate inputs and outputs. A source/ subfolder for what you supply, and output/ for what should get created.
• Write the brief into a file. A short brief.md explaining what's in the folder and what should come out of it. You won't have to re-explain it in every new conversation.

A warning worth reading twice. Don't put anything into a folder you show it that you don't want it to see. The most common mistake goes the other way: for convenience, someone points it at the whole desktop or Documents folder, and along with the source material, it sends along a tax return, payroll records, and private photos. Other people's personal data, health records, and login credentials don't belong there — not even in a paid account.

Work in this project's folder, but read-only for now.
Don't edit, don't create, don't rename anything.

The folder has [scanned receipts and invoices from the last
quarter]. Go through all of it and tell me:
1. how many files there are, of what types, and covering what period
2. which files are unreadable, blurry, or cut off and will need
   to be rescanned
3. which look like duplicates of the same document
4. what in the folder, in your view, doesn't belong to this task at all

Then suggest how to turn this into one spreadsheet: which columns
it should have and why those specifically.
Don't build the spreadsheet yet — wait for my approval.

It returns an inventory and a proposed structure before anything changes. Point 2 is the most valuable one — unreadable scans are the most common reason a resulting spreadsheet contains figures nobody actually checked. Once you've approved it and had the spreadsheet built, check it against the originals; for amounts and dates, that step isn't optional.

Voice control

There are two different things, and both get called voice control. Dictation converts speech to text: you press the microphone, speak, text appears in the field, you read it, and you send it. The model gets an ordinary written message, and you have the last word, because you see the text before it goes out. Voice conversation is a spoken conversation — you talk and hear the answers. It's typically a paid-tier feature, and in some modes not everything from written chat works (typically file work or image generation); the tip voice mode covers it in detail.

Dictation is enough to start with. Where to find it: in the mobile app, a microphone icon sits right by the text field; if it doesn't work, use your phone's system dictation keyboard. On a computer it works too: on Windows, press the Windows key and H; on a Mac, dictation is turned on in keyboard settings.

When it's faster. You speak roughly three times faster than you type. It pays off on the walk from a meeting while the context is still fresh in your head; for a long brief where your patience would run out after the third sentence; and at the moment you don't know how to start — speaking out loud comes easier than facing a blank field. And with no exceptions: never while driving.

Why a dictated prompt tends to be better. When you type, you economize on words — you type “summarize this” and hope. When you speak, you speak naturally: you'll say on your own who it's for, what happens with it next, and what's bothering you about it. A dictated brief typically runs eighty words where a typed one would run twelve — and it's exactly that extra context that makes the difference between a generic answer and a usable one. Why context matters so much is covered in the chapter prompting as a skill.

What to watch for. Dictation reliably mangles three things: proper names of people, companies, and places; technical and specialized terms from your field; and numbers, addresses, and reference codes, where one mis-heard digit changes everything. Manually fix key names and numbers after dictating, and read the text before you send it — the transcription occasionally produces a sentence with the opposite meaning. A dictated prompt then looks different from a typed one: longer and more conversational, and that's fine.

This is dictated, expect typos and names that might be transcribed
wrong. Where you're not sure, ask rather than guess.

So. I'm coming back from a meeting with a client, their name is
[name], they do [field], and they'd like us to [job]. About five
things came up that I need to remember. First, they want it done
by [deadline], which feels tight to me. Second, they asked about
price and I said I'd send a quote by the end of the week. Third,
they had reservations about [item], I'll need to address that in
the contract. And they also mentioned something about [another
item], but I didn't quite catch it.

Turn this into meeting notes with a task list of who does what by
when, list separately what I promised and by when, and write three
questions I should have asked and didn't.

It returns notes, a list of promises, and questions for next time — that last point often turns out more valuable than the notes themselves. Double-check names and dates afterward; transcription is least reliable there.

Images and graphics: what each service can do

This is where honesty matters, because it's where nearly every beginner trips up.

Claude does not generate images. It won't draw you a poster, produce a logo, or create an illustration for a presentation. Ask it to, and it will either say so or offer to write a brief for a tool that can. It isn't hidden in settings, and it isn't limited to the paid tier — it simply can't do it. Anyone hunting for a “generate image” button is hunting in vain.

It reads images excellently, though — and people underrate this far more than they overrate generation. It reads a photo, a scan, a screenshot, and handwriting: a receipt turned into a spreadsheet row, a photo of a whiteboard turned into a task list, handwritten lecture notes, an error message explained in plain language, a photo of a form or a label — anything that's faster to photograph than to retype.

I'm attaching a photo of [a whiteboard after a meeting / handwritten
notes]. First, transcribe exactly what you see in the image,
including anything hard to read — for uncertain spots, write
[illegible].

Then turn it into:
1. a task list in the form: who — what — by when
2. decisions that, according to the notes, were made
3. points that were left open

Don't guess at names or deadlines that aren't in the image. Where
a responsible person or date is missing, write NOT SPECIFIED.

The two-step approach — a literal transcription first, then processing — is deliberate. You compare the transcription against the photo and immediately see whether the model read it correctly. If it jumped straight to a task list, you wouldn't be able to catch an error in a misread name or date.

ChatGPT can generate images. It creates them right in the conversation — either you ask for one in a sentence, or you pick it from a menu. It can add text to an image following instructions, fill in details, or make the background transparent, and it can also edit images: you upload an existing one and describe the change, or mark a specific part of it and say what to do with it. Available on the web and in the mobile apps.

Other tools. Gemini from Google generates and edits images too, and is handy if you're already in the Google ecosystem. Canva is a different case — a graphic design editor with templates where you drop the image straight into a poster or presentation at the right size; for someone who needs a finished visual rather than just an image, it's the most practical route. Adobe Firefly is a generator connected to the rest of the Adobe apps.

How to put it all together. Claude works out the brief and the text, another tool renders it. Image models simply aren't good at thinking about what should be on the poster — they're good at drawing.

I need [a poster for an open house] for [audience]. Don't make an
image — I know you can't. Prepare a brief for an image generator
and the text that will appear in the visual.

I want:
1. three visual concepts, one sentence each
   (what's in the image, what mood, what colors)
2. for the chosen concept, a detailed brief for the generator: what's
   in the foreground, what's in the background, style, color palette,
   format, where text should fit
3. the text separately: a headline under 5 words, a subhead under
   12 words, practical details ([date, location, time])
4. one sentence per concept on why it might not work

Write the generator brief in English, the text in [language].

It returns three concepts and a ready-to-paste brief for the generator. It's usually better to leave text out of the generator and add it in a graphics editor afterward — generators still mangle non-English text and diacritics badly. And never generate likenesses of real people or someone else's logo.

The first seven days

One thing a day, fifteen to thirty minutes. The order has a reason.

Day seven has a concrete output — your own custom instructions, meaning a short description of you and your work that the tool remembers so you don't have to repeat it in every conversation:

Over the past week we worked through these things together:
[list in bullet points what you tried, including what didn't work]

I work as a [profession] at a [type of organization], my job is
mainly [what you focus on], and the biggest time sink is
[time-waster].

Based on this, suggest text for my custom instructions — permanent
information about me you should remember for future conversations:
- who I am and what context I work in
- how I want you to write (length, tone, structure of answers)
- what to always include and always flag for me
- what to avoid

Write it as finished text ready to paste in, no more than 200 words,
in the first person as if I wrote it. Then, separately, explain why
you included each part, so I can edit it myself.

It returns text you paste into your custom instructions settings, plus an explanation that lets you edit it yourself. Come back to it after a month — it'll look different, because by then you'll know what you actually want from the tool. From here, continue to your first conversation with AI, which focuses purely on the quality of your briefs and checking the answers.

Common mistakes

• Downloading an installer from a search result or an ad. Type the official site's address into the address bar by hand, or install from an app store.
• Installing an app with a similar name. Mobile stores are full of clones like “AI Chat Assistant.” Check the publisher (Anthropic, OpenAI) and the number of ratings, not the icon.
• Paying twice. Once on the web, once in the app through a store. The payments don't cancel each other out, and you cancel in your Apple ID or Google account settings, not in the app.
• Starting with Claude Code because it's “the most powerful.” The most powerful tool for a problem you don't have is just extra frustration.
• Pointing Cowork at your entire desktop or Documents folder. Along with your source material, you send everything else too. One folder per task, ideally a copy.
• Turning on a connector with write and send permissions right away. Read-only for the first two weeks. Sending, deleting, and payments are always confirmed by a human.
• Trusting that dictation got names and numbers right. Transcription mangles proper names and digits most reliably of all.
• Looking for image generation in Claude. It doesn't exist. Claude reads images, it doesn't draw them.

The best tools

• Claude, desktop app — the only route to Cowork, meaning work over a folder of files, and the most comfortable way to write day to day.
• Claude, mobile app — a camera and a microphone: photographing documents, whiteboards, and handwriting, and dictating long briefs on the go.
• Claude Code — tasks that need to repeat, run, or be version-controlled. Requires the willingness to type commands, not a programming degree.
• ChatGPT — a second opinion, and above all image creation and editing, which Claude doesn't have.
• Canva — when you don't need an image but a finished visual at the right size: a poster, a post, a presentation.

What you get out of it

• Time: the first week you'll likely lose a few dozen minutes rather than save them — installing, experimenting, searching. From the second week on, it flips: for someone who writes emails and processes documents daily, it's routinely several hours a week.
• Peace of mind: something you capture the moment it happens — a photographed receipt, a dictated meeting note — no longer sits in your head.
• Quality: a dictated brief with context returns better answers than two typed sentences, and custom instructions mean you stop repeating that context over and over.
• Confidence: the next time you hear about Cowork, connectors, or Claude Code, you'll know what it is and whether it applies to you. Most often you'll find it doesn't apply to you yet — and that's a useful answer too.

Pro tip

The best investment of the first week doesn't even take a minute: set a keyboard shortcut that brings up the window from anywhere. It sounds like a cosmetic detail, but it decides everything else. A tool you have to click through a browser and three tabs to reach gets used when you remember it exists. A tool that pops up with one keypress gets used the moment you actually need it.

And the closing rule that outlasts every change of tool: never hand off a task whose result you can't check. Not because AI is often wrong, but because when it is wrong, it sounds every bit as convincing as when it's right. A document summary is worth something once you've verified three claims against the original. A spreadsheet built from receipts is worth something once you've eyeballed the amounts. Checking costs you minutes; an undetected mistake can cost you trust.