The frame
Why default Claude feels mid
Out of the box, Claude has three specific handicaps. It does not know who you are, so every answer is written for a stranger. It cannot touch any of your actual tools, so it can only talk about work instead of doing it. And it dies the second you close your laptop, so nothing happens while you are away.
Saving more prompts to your notes app does not fix any of those. Neither does building the perfect skill. Those are things you do after the foundation is in.
Each of the three settings below kills exactly one handicap:
Setting 1
Context
It knows who you are in every chat, forever.
Setting 2
Tools
It can reach your real apps instead of just describing them.
Setting 3
Continuity
It keeps working after you shut the laptop.
Setting 1 — about 5 minutes
Global Instructions
Global Instructions is a text box that loads at the top of every single conversation you have. It is the single highest-leverage thing in your entire account, and most people leave it empty.
Here is the part people get wrong: they try to write it themselves, stare at the box, type "I'm a marketer, be concise," and call it done. That is worth almost nothing.
Do not write it. Have Claude interview you and write it for you. It knows what it needs to know about you better than you do. Paste this prompt into a new chat and answer its questions honestly for a few minutes.
Copy-paste prompt
You are going to write my Global Instructions: the standing block of context that Claude loads at the start of every conversation I ever have with you. You do not know anything about me yet, so do not write a single word of it until you have interviewed me. HOW TO RUN THIS INTERVIEW: - Ask me ONE question at a time. Wait for my answer. Then ask the next one. - Never dump a list of questions on me. This is a conversation, not a form. - If my answer is vague, ask a follow-up until it is concrete. Push me for real numbers, real names, real examples. "I run a business" is not an answer. "I run a 4-person landscaping company in Tampa doing about $600k a year" is an answer. - Roughly 10 questions. Go deeper on whatever I seem most invested in. - Talk like a sharp friend who is genuinely curious, not a customer service bot. COVER THESE, in whatever order the conversation naturally goes: 1. Who I am. Name, where I am, what I actually do all day. 2. My work. What I produce, who I serve, how I get paid or judged. 3. What I am working toward in the next 90 days. What "winning" looks like specifically. 4. Where my hours actually go, and which of those hours I resent. 5. What I am good at, and what I would hand off tomorrow if I could. 6. How I write and talk. Words I use. Words I hate. 7. How I want you to respond by default: long or short, direct or gentle, push back or agree, ask before acting or just go. 8. The recurring stuff I will ask you about over and over. 9. Anything a great chief of staff would need to know about me that I have not mentioned. WHEN THE INTERVIEW IS DONE: Say "I have what I need. Here are your Global Instructions." Then output ONE clean block I can copy in a single click, written as direct instructions to you, in this shape: WHO I AM [2 to 3 factual sentences. No adjectives, no fluff.] WHAT I AM WORKING TOWARD [My real goals, specific and dated where possible.] HOW TO TALK TO ME [My tone, my format defaults, my length preference.] ALWAYS [The standing rules. Things to do in every reply without being asked.] NEVER [The hard rules. Words, formats, and behaviors I do not want, ever.] Include this line word for word inside ALWAYS: "Address me by my first name in every single reply." That line is a free warning light. The moment you stop using my name, the conversation has run too long and lost the thread, and I know to start a fresh chat. RULES FOR THE FINAL BLOCK: - Under 400 words. This loads on every message, so every word costs me. - Plain language. Write it the way I talk, not the way a consultant writes. - No hedging, no "try to," no "when appropriate." Give it spine. - Specifics only. "Writes marketing emails for B2B SaaS clients" beats "works in marketing." Start now. First question only.
When it hands you the finished block, copy it. Then put it where it actually loads:
Open Settings → Cowork
In the Claude desktop app, click your name, then Settings, then Cowork in the left menu.
Click "Edit" next to Global instructions
Paste your block into the box and hit Save. That is it. It now loads on every conversation.
If you cannot find it
Depending on which build you are on, the same box also shows up under your profile settings as "instructions for Claude" or "personal preferences." Same box, same effect, applies to both Chat and Cowork. Paste it wherever your app puts it.
Why the prompt makes it say your name
That line is not vanity. It is a free warning light. Claude will keep using your first name while it still has a good grip on the conversation. The moment it stops, the chat has run too long and started dropping context. That is your cue to open a fresh one before it starts making things up.
Setting 2 — about 2 minutes
One connector for every tool
Claude ships with a handful of built-in connectors, and you can add more one at a time. That path gets old fast. Every new tool is another connector to find, add, authorize, and babysit.
Composio collapses that into one. You add a single custom connector, and from then on any app you want to connect goes through it. Gmail, Slack, Notion, Linear, HubSpot, Calendar, hundreds more. You are managing one connection instead of fifty.
Open Customize → Connectors
In the Claude desktop app, click Customize in the left sidebar, select Connectors, then click the + icon at the top.
Click "Add custom connector"
Name it Composio and paste the server URL below.
Click Connect, then authorize in your browser
A browser window opens. Sign in to Composio, approve the permissions, and you land back in Claude.
Just ask for what you want
Do not go hunting for an integrations menu. Say "read my last 5 emails" or "create a page in my Notion." Claude prompts you to authorize that specific app right there, once, and it is connected from then on.
Composio MCP server URL
https://connect.composio.dev/mcp
The honest part
You need a free Composio account to authorize the connection, and it does sit between Claude and your apps, which is worth knowing before you hand it your inbox. If you only ever plan to use Gmail and Calendar, Claude's own built-in connectors are simpler and you can skip this one. Composio earns its place the moment you have four or more tools, or you keep trying new ones.
On Claude Code instead of the desktop app?
claude mcp add --transport http --scope user composio https://connect.composio.dev/mcp
Then run /mcp, pick Composio, and authenticate. Both flags matter. Without --transport http it errors out, and without --scope user it only works in one folder.
Setting 3 — about 30 seconds
Run new tasks in the cloud
This is the one nobody knows about, and it is a 30-second toggle.
By default, a lot of Cowork work runs on your computer. Which means it runs while your laptop is open, awake, and online, and stops when it is not. Flip this on and new tasks run on Anthropic's servers instead. Close the laptop, walk away, and the work keeps going. Scheduled tasks fire with no device online at all.
Go to Settings → Cowork and turn on "Run new tasks in the cloud."
Read this before you assume it is working
Three things trip everyone up, and none of them are in the toggle description:
Cloud tasks cannot see your local files. That is the whole tradeoff. If a task needs to read or write a folder on your Mac, it needs your computer. Ask a cloud task to touch a local file and it will quietly flip itself back to running on your computer.
It only applies to new tasks. The word "new" in that toggle is load-bearing. Scheduled tasks you already built do not migrate on their own.
An existing task with a folder attached will refuse to move to the cloud. If you try to switch an old scheduled task over and the option is greyed out, that is why. Deselect the local folder it is pointed at, save, and then you can move it.
The workaround that gets you both
If you want tasks that run overnight and touch your files, stop keeping your working folder only on your hard drive. Put it in Dropbox or Google Drive, connect that as a connector, and point the task at the cloud copy. Now Claude reaches your actual files through the connector instead of through your laptop, and the laptop can be shut. That is the whole trick, and it is the difference between a toggle you flipped and a system that runs while you sleep.
Confirm it took
The 60-second check
Do not assume all three landed. Settings fail silently, and the failure mode is you thinking you have a system when you have a chatbot. Open a fresh chat and paste this.
Copy-paste prompt
I just changed three settings and I want to confirm all three actually took effect. Run these checks and give me a straight answer on each one, no guessing. 1. CONTEXT CHECK Without me telling you anything: what is my name, what do I do for work, and what am I currently working toward? If you cannot answer all three from your instructions, say "Global Instructions are not loading" and tell me what you can see. 2. TOOLS CHECK List the connectors and tools you currently have access to. Then tell me specifically whether Composio is connected, and which of my apps are authorized through it so far. 3. CONTINUITY CHECK Tell me whether this task is currently set to run on my computer or in the cloud, and explain in one sentence what that means for whether you keep working after I close my laptop. Then give me a short punch list: anything that is not working, and the exact click path to fix it.
The checklist
All of it, in 8 minutes
Run the interview prompt, paste the block into Global Instructions
Settings → Cowork → Global instructions → Edit → Save. About 5 minutes, almost all of it answering questions.
Add Composio as a custom connector
Customize → Connectors → + → Add custom connector → paste the URL → Connect. About 2 minutes.
Turn on "Run new tasks in the cloud"
Settings → Cowork → toggle it on. 30 seconds.
Run the 60-second check
Fresh chat, paste the verify prompt, fix anything it flags.
What's next
Once these three are in
You now have the foundation: Claude knows you, it can reach your tools, and it keeps working when you are gone. That is the part that does not change no matter which model or tool comes out next month.
The next two moves are building a second brain it reads from every session, and turning the thing you do twice a week into a saved skill that runs itself. Both are in the full setup guide, along with the prompts.