Putting AI to work

Before you start

This picks up after setup: your privacy settings, your first project, the files that teach Claude who you are. If you haven't done that yet, start with Before you start: settings and plans. Nothing in here works well on an empty brain.

This one is about the work. How to ask for things so you don't get generic answers back. How to let Claude build you tools instead of just paragraphs. How to connect it to the apps you already pay for. And how to take something that worked once and make it work the same way every time.

We want to be straight with you about one thing up front. There is a whole world of Claude that involves code, GitHub, deploying websites and a terminal window. It's real and it's powerful and it is not in this guide. Nobody needs it to get value out of this, and the people who chase it first usually quit before they get anywhere.

Part 1: Four settings worth your time

You already turned off model training. These four are the ones we'd change next. All of them live under your name in the bottom left, then Settings.

1. Turn on memory

Settings > Capabilities > "Generate memories from chats"

This lets Claude remember things about you across conversations, not just inside one project. It's the difference between a new temp every morning and someone who's worked for you a while.

If you're coming from ChatGPT, there's a second thing on this page worth doing. Import memory hands you a prompt. Paste that prompt into ChatGPT, let it write out everything it has learned about you, copy the result, and paste it back into Claude. Two minutes, and you don't start from zero.

2. Leave tools on "load when needed"

Settings > Capabilities > Tool access

There are two options. Load tools when needed, or load all tools up front. Keep it on "when needed."

Everything Claude is holding in its head during a conversation, your files, your instructions, the tools it has available, competes for the same limited space. Loading every tool up front fills that space with things you probably aren't using, and your conversations run out of room sooner. Claude is good at figuring out which tool it needs. Let it.

3. Turn on notifications for finished responses

Settings > Notifications > Response completions

Small thing, changes your day. When you ask for something that takes real time, like research across twenty websites or reading a long document, you close the tab and go do something else. Claude tells you when it's done.

4. Turn on deeper chat search

Click the search icon in your sidebar. If your results look thin, there's a link that says go to settings. Turn on search and reference chats.

Without this, search only looks at conversation titles. Which means the thing you and Claude worked out three weeks ago is functionally lost. With it on, Claude searches inside the conversations and can pull from them.

One we'd leave alone: location. It's useful. It means you can ask about something local without typing "in Orlando" every single time.

Part 2: The prompt structure that fixes generic answers

The most common complaint we hear is "it gave me something generic." Almost every time, the prompt was generic first.

There's a four-part structure worth memorizing. Goal, context, action, output.

Goal. What you're actually trying to achieve. The outcome, not the task.

Context. Who you are, what's attached, what it needs to know. This is where your files and links go.

Action. The specific work you want done. Say "use web search" here if the answer depends on anything current.

Output. What the finished thing should look like. Bullet list? Table? A one-page document? An email you can send as-is? Say it.

Here is the difference.

Generic: Help me get more customers.

Structured: My goal is to fill the four open slots I have every week in the fall, which is usually my slowest stretch. For context, I run a residential cleaning business in Winter Park, I've been doing it eleven years, and my best customers are almost all referrals from three neighborhoods. I've attached my customer list with start dates and the neighborhoods they're in. I want you to look at where my customers actually come from, then use web search to find what's happening in those specific neighborhoods this fall, HOAs, schools, community events, and suggest five ways I could get in front of more people there. Your output should be a numbered list, each one with what it would cost me, how long it would take, and why you think it would work for my particular situation.

The second one takes ninety seconds longer to write. It comes back with something you can actually do on Tuesday.

The part almost nobody does

Put that structure into your project instructions instead of typing it every time.

Open your project, click into the instructions, and write your goal, your context, your standing action and your preferred output format there. Once. Then every conversation inside that project starts with all four already loaded.

Two lines worth stealing for your project instructions:

Match my energy and match the length of what I give you. If I send you something short, answer short. If I send you a long detailed prompt, give me the detailed answer.

Before you produce anything, ask me the questions you need answered to do this well.

That second one is the highest-value sentence in this entire guide.

Part 3: Artifacts, or letting it build you things

Most people use Claude to write words. The thing they never try is letting it build them a working tool.

An artifact is anything Claude makes and hands back as its own object rather than as text in the chat. A document. A checklist. A calculator. A small web page you can send someone a link to. It lives in your Artifacts library in the sidebar and you can open it again any time.

The one to try first

Think of one number you calculate by hand over and over. A quote. A square footage estimate. A payment plan. How much each sibling owes toward Mom's care this month. Then ask for it as a tool.

Build me a simple calculator I can use on my phone before I give someone a quote. I charge by square footage, with a different rate for deep cleans and recurring cleans, a minimum of $150, and a $40 add-on for inside the fridge and inside the oven. I want to type in the square footage, pick the service type, check the add-ons, and see the price. Make it look clean and not like a spreadsheet.

You'll have it in about a minute.

The trick that gets you out of AI-looking output

The first version of anything visual will look like AI made it. Purple gradients, emoji, the same centered layout everyone gets. That's the starting point, not the finish.

So ask for options:

Give me ten completely different styling variations of this, each one with a name, in a tab switcher so I can click through them. I want this to look like a person designed it. Make each one genuinely different from the others, not small variations on the same idea.

Then pick. This works for anything visual.

Publish and share

Any artifact can be published to a link you send someone. Nobody needs a Claude account to open it. It's how you send a checklist, a calculator or a one-page summary without attaching a file they have to download.

Part 4: Claude on your actual computer

Everything so far happens in your browser. The Claude desktop app can also work with the actual files on your computer.

You download it from the lower left of claude.ai.

The one setting to change immediately

In settings, turn on the quick access shortcut. It pulls up a small Claude bar anywhere on your computer, over whatever you're doing.

While it's open, you can drag to select any part of your screen and it drops straight in as a screenshot. Meaning you can point at a confusing error message, a form, a line on a statement, anything, and ask "what am I looking at" without saving a file or switching windows.

What it's actually for

The honest answer is the boring stuff you've been putting off.

Read my Downloads folder and tell me what's in there. Then move everything into the right subfolder inside Documents. There are already folders for photos, PDFs, invoices and contracts. If something doesn't fit any of them, leave it where it is and tell me what you skipped.

Claude asks permission as it goes, shows you a running checklist, and you can watch it work.

A few things to know:

It asks permission for each new thing it touches. "Allow for this session" is the reasonable middle setting.

Projects here can point at a real folder on your computer. Files it creates land in that folder, on your hard drive, where you can open them normally.

Scheduled tasks mostly run in the cloud now, so your computer can be asleep. The exception is a task that needs files or apps on your computer.

A real business example

1. Make a project. Instructions: use web search, and save what you find as spreadsheets and documents in this folder.
2. First prompt: Find me fifteen businesses within twenty miles that look like they need what I sell. Save them to a spreadsheet with a column explaining why each one is a fit and where you found them.
3. Claude stopped and asked clarifying questions before searching. What area exactly, what kinds of businesses, and what specifically counts as "needs what I sell." Good answers there made everything downstream better.
4. Second prompt: Now take a screenshot of each of their websites into a subfolder, and write me a short, specific cold email for each one. Reference something real about their business. Find the owner's email where you can, and where you can't, give me the phone number instead.

Fifteen prospects, researched, with drafted outreach, in the time it takes to eat lunch.

Part 5: Connectors, or wiring in the tools you already use

A connector lets Claude read and write to another app. Gmail, Google Calendar, Google Drive, QuickBooks, Notion, Slack and a long list of others. Add them from the plus button in the chat bar, then Connectors.

You'll see the term MCP thrown around. It stands for Model Context Protocol and it means exactly one thing: the standard way outside apps plug into Claude. You don't need to know more.

Set the permissions, they're worth the two minutes

After you connect something, click into its settings and look at the list of things it can do. For each one you can choose always allow, ask me first, or off entirely.

We connect Google Calendar with creating events allowed, and updating and deleting events turned off. We connect Gmail with drafting allowed and sending turned off. The app's first permission screen asks for everything. You narrow it afterward, inside Claude.

Where this gets good

Two connected apps and one prompt is where the time actually comes back:

Look through my email from the last ninety days and find everyone I promised to follow up with and never did. For each one, draft a reply in Gmail that references what we actually talked about. Do not send them. Drafts only. I'll read every one before it goes.

You read them, you fix the two that are off, you send.

Say "draft only, don't send" out loud in the prompt. Every time. You want the habit, because the day you're tired is the day it matters.

Part 6: Skills, or making a good result repeatable

Here's the frustration that skills solve. You have a great session with Claude. It produces exactly the right thing. Two weeks later you try to do it again and you can't remember how you got there, and what comes back is different.

A skill is that process, written down, so it runs the same way every time. Same steps, same rules, same format.

Three ways to make one, in order of how much we recommend them

1. Reverse-engineer something that already worked. Finish a session you're happy with and say:

Make a skill that does exactly what we just did, so I can run it the same way every time.

You already know what good looks like, because you just made it.

2. Describe what you want. Tell Claude the job, let it ask you questions, let it build it. Fine for new ideas.

3. Upload someone else's. Fastest, and the least likely to fit your life.

What goes into a good one

The instructions. What the job is, the steps in order, and the rules. What it must always do and never do.

The examples. Attach the two or three past outputs that were right. This does more for quality than any amount of describing.

The hard parts. If some step involves a calculation, a format or a lookup that Claude tends to fumble, say so and let it build a small helper for that step.

Don't build one on day one

Do the job by hand four or five times first. You need to know what good looks like before you can write down how to get there.

Teach it once goes much deeper on skills, including how to tell whether one actually fired.

Part 7: Scheduled tasks

Once a skill works, you can put it on a clock. Scheduled tasks need a paid plan. Sidebar, Scheduled, New task. Name it, write out the full prompt, pick a frequency and a first run time.

Rules we'd hold you to:

Write the prompt as if Claude has never met you. A scheduled run starts fresh with no memory of the conversation where you set it up.

Tell it what not to do. The best example we've seen was a daily task that researched and wrote three blog drafts, then emailed the links. The instruction that made it safe: never publish. Just draft and send me the links.

Turn off every connector it doesn't need. If the task needs Gmail, give it Gmail and nothing else.

Run it manually once before you trust it. There's a run-now button. Use it and watch what happens.

One honest warning. Anything running unattended occasionally touches something you didn't intend. Start with something low-stakes, like a weekly summary emailed to you, before you let anything reorganize your files at two in the morning.

Things we want you to hear

The first version is about half right, and that's normal. The whole difference between people who get value out of this and people who quit is what happens after the first draft.

It will be confidently wrong sometimes. Check any number, date, price, name or legal claim. Treat it like a fast, willing assistant who has never been told when to stop and ask.

Your expertise is the ingredient. If you don't know what a good result looks like, Claude will find a generic definition of one and give you that.

A few hours, not fifteen minutes. Anyone promising you a finished anything in fifteen minutes is selling something.

Training it is like training a new hire, except it follows the instructions better than a person does. The first few rounds won't be what you want. You correct it, you tighten the instructions, and eventually it runs the process the way you would.

What we'd do with the next thirty days

Week 1. Change the four settings. Put your goal, context, action and output into your project instructions. Add "ask me clarifying questions first" to the end of everything you send.

Week 2. Build one artifact. The calculator, the checklist, the one-pager. Something you'll actually open again.

Week 3. Download the desktop app. Set the shortcut. Let it clean one folder. Connect one app you already use and set its permissions properly.

Week 4. Take the job you've now done five times and turn it into a skill. Then run it.

When you get stuck

Ask Claude. Out loud, in plain words, including "I don't understand what you just said, explain it like I've never done this before." It's better at explaining itself than any guide is, including this one.

Settings and menus were accurate as of October 2026. Anthropic changes both, often. If a screen doesn't look like this guide says it should, trust the screen.