I spent two months outside Silicon Valley. Nobody around me talked about AI, and when it did come up, it was to say that AI means nothing.
People there were all about which beach was the hottest. (Not which model, haha.) Nobody sat in cafes with laptops. Another world.
The other world where "AI" is nobody's daily word.
Then I came back home to Silicon Valley, where Claude and Codex are practically part of my family. To be honest, I like it here. But two months away showed me how small the AI bubble is.
The gap shows up even in the US, the moment I leave the valley. The country that builds most of these tools ranks 24th in the world for actually using them. In the US, 28% of people use AI. In Singapore it's 61%. In the UAE, 64%.
And even the people who use it are mostly doing small things with it. The biggest real-world use of AI this year is therapy and emotional support. After that, troubleshooting. Then a category researchers actually named "fun and nonsense." Autonomous agents, the thing that's supposed to mean you've arrived, didn't even make the top ten.
The most common ways people actually use AI. "Running autonomous agents" isn't in the top ten.
The gap is smaller than it feels. Agents aren't at the bottom of that list because they're hard. They're there because almost nobody has sat down and built one.
I'll show you with the thing I do every week: writing this newsletter. Yours might be a weekly report, a research summary, a running budget.
By the end you'll have an agent: a workflow that knows your work, follows your rules without being told, takes a one-line prompt instead of a page, and keeps running on a schedule while you sleep.
One thing first: I'll do this in Claude. If ChatGPT is your tool, stay there. Both start you in the exact same place, a Project. Steps 2 through 4 go by different names there, but they do the same job. Open whichever one you already pay for and build it with me.
1. Start with a Project
If you've been pasting your writing samples into a fresh chat every time and re-explaining yourself, a Project ends that.
This is where Projects live: a space you build once and work inside, separate from the everyday chat.
Create a new Project and name it after the real thing you do: Newsletter, Client emails, Weekly report.
A Project has two parts, and both matter.
The materials: your reference shelf.
This is what Claude pulls from when it needs to know how you work. Feed it your actual work, not a description of it.
For a newsletter: a few pieces you're proud of, or one issue you wish yours looked like.
For client emails: real emails you've sent, in the tone you want to keep.
For a weekly report: past reports or the raw files behind them.
Here's what actually lives in my Newsletter Project:
Past issues: so it knows the format and rhythm before it writes a line.
My top ten subject lines by open rate: so it learns what my audience opens, not what I assume they like.
My copywriting rules: the persuasion principles I hold every draft to, so I'm not re-teaching taste every week.
My voice rules: so Claude writes like me.
That's the whole folder. I don't drop generic files in there.
The instructions: how you want it to work.
In the Project's instructions box, write a short standing brief for this project: what you're making, the rules it should always follow, and how to use the materials you uploaded. Claude reads these instructions every time you work in this project.
If a blank box is intimidating, talk to Claude in the chat first ("here's my work and the task, help me write the instructions for this project") and paste back what it gives you. One or two paragraphs is plenty to start.
A Project has two halves: the files you upload on one side, the standing instructions on the other. Claude reads both every time you work here.
Once both parts are in, you have a working setup. A one-line prompt is enough, and it already knows what you mean: "here's an idea, draft the next issue." Or "here are last month's numbers, draft the report."
Next level: I keep an about-me file that I upload to every project, so Claude knows exactly who I am, what my business is, and what I'm aiming for. Do a short interview with Claude to build it once, and then add it to each Project.
2. Create Skills
A skill is a text file where you write down, once, how you want a task done. Claude follows it every time.
A skill works in any chat, even outside this project. It starts itself when the task matches, so you don't write the prompt again.
You can download a generic skill, and that's enough to start. Here's a starter newsletter skill you can download and adapt: here.
I'd customize it to your own standard for how the task should be done. My skills also point to my Project files. For example:
when it suggests titles, it studies my top ten by open rate.
before drafting, it follows my voice rules.
Go to Settings → Capabilities → Skills, and add your file.
Skills live in Settings, not inside any one project. Add one for each task you run regularly.
You don't need a skill on day one. The project instructions you wrote in step 1 already gave you a working setup. Aim to build a skill for each task you run regularly.
Next level: a skill is never really finished. When you catch yourself giving the same note twice, ask Claude to fold that feedback into the skill, then upload the updated file. It gets sharper each week, so you stop repeating yourself.
3. Schedule what’s on repeat
Some tasks are the same small job every time, no creativity needed. Those go to Cowork, on a schedule.
Switching to Cowork is one click. Setting the schedule is one sentence: "do this every Monday."
Set a task once and Cowork runs it on schedule, even when you're not at your desk.
A couple of mine:
-a weekly report on open and click rates, how the A/B subject lines did, and where new subscribers came from.
-every week, Claude scans our past issues and logs which use cases and examples we've already used, so when I sit down to write I can see what ran and when, and I don't tell you the same story twice.
My two Monday tasks in the Newsletter Project: the performance report and the use-case log.
Next level: Want more scheduled-task ideas? I built the whole newsletter into a content agent and wrote up the full step-by-step, with every task it runs. Read it here: Claude agent. One evening.
4. Connect tools
Everything so far lives inside Claude. This step is what turns a setup into an agent: it can pull what it needs and act, without you in the room.
Connect your tools: email, calendar, Notion. Then hand it a real job.
I can ask it to pull the numbers on our last 10 newsletters, or check my email for a brand-deal confirmation, without opening Gmail or beehiiv myself.
Connectors live in Settings. Flip them on and Claude can reach your email, calendar, and notes.
Bonus: Make it visual
Some people are happy with a wall of numbers. I like the numbers, I just want to see them.
Remember presentations and charts in PowerPoint? Claude does all of it, a few levels up. Type "show me in HTML" and watch your idea, report, or even a long text turn into a clean web page.
Take our quarterly goals. I could leave them as rows in a spreadsheet, or have Claude build them as a page in our brand colors.
Where our subscribers come from today, and where we're aiming next. The same numbers, built as a page instead of a spreadsheet.
That's the whole thing. You opened Claude, did one real task, and walked away with an agent that knows your work and runs the repetitive parts on its own. The thing that didn't make the top ten, you just built. You're not behind. You never were.
Build one this week and tell me what you made.
💜 Marina
P.S. Don't stop here! Your agent only gets more useful the more of your own tools you connect, so go see which ones are worth adding: The 13 tools my team can't work without.
