I'm Neeraj. I spent 11 years as an engineer at Goldman Sachs and JPMorgan and now I build AI agents that run my own business. A chatbot answers when you type. An agent has a job, the tools to do it, and permission to run on its own. Below is how you set one up in the browser, four levels, with no terminal and nothing to install. Each level ends with how you know it worked, so you never have to guess.
→Level 1: do the job once by hand in Claude or ChatGPT. Ten minutes, and you keep whatever it makes.
→Level 2: save the job as a Project so you never explain it twice. Free.
→Level 3: give it hands. Connect your Gmail, calendar and Drive. Also free, and it asks before it sends anything.
→Level 4: put it on a schedule so it runs while you sleep. This one takes the $20 plan.
🔒Get the four-level setup
The prompt shape, the Project instructions to paste, the connector steps, and the schedule setup. Yours immediately.
Join 3,000+ builders already getting this. No spam.
THE SETUP — 6 STEPS
01
Level 1: do the job once, by hand
Open Claude or ChatGPT, whichever you already have. Pick a real task off today's list, something you'd otherwise do yourself: messy meeting notes that need to become action items, or a long proposal you have to understand before a call. Give it the situation, then what you have, then what you need, in that order. Paste the text or upload the file with the paper clip icon. The first answer is a draft, so tell it what to fix. You'll know it worked when you're holding a one-page thing you'd send to someone. Keep it. Every level below is this same job, done with less of you.
Here is the situation: I just got off a 40-minute call with a client and my notes are a mess.
Here is what I have: (paste the notes)
Here is what I need: a clean list of action items with who owns each one, the decisions we made, and the three things I should follow up on this week.
For today, use normal business documents and keep client financials out of it.
02
Level 2: save the job as a Project
A Project is a folder in Claude with standing instructions. Every chat you start inside it already knows the job, so you paste the notes and get the output without explaining anything. It's free, you get five. In Claude, click Projects in the left sidebar, then New Project, name it after the job, like Meeting notes, then Set instructions and paste the block below. Write the instructions like a job description for a new hire: what comes in, what goes out, what to flag. You'll know it worked when you open a fresh chat inside the Project, paste raw notes with no other words, and get the same clean list as level 1.
You turn my raw meeting notes into a follow-up plan.
Every time I paste notes, return:
1. Action items, one line each, with an owner and a due date if one was mentioned
2. Decisions made
3. Open questions nobody answered
4. Anything that sounds like a risk or a deadline I could miss
Keep it under one page. Plain language. Ask me nothing unless the notes are unreadable.
03
Level 3: give it hands
Right now the agent only sees what you paste. Connectors let it see your actual email, calendar and files. Gmail, Google Calendar and Google Drive connectors work on every Claude plan, including free. In any chat, click the plus sign, hover Connectors, and switch on the ones you want. Google will ask you to sign in and approve. Then ask for something you'd normally do by hand, like the prompt below. Before it sends, replies to or forwards any email, Claude asks your permission. It drafts, you approve. You'll know it worked when it lists real emails from today and hands you drafts that are sitting there waiting for your click.
Read my unread emails from today. Tell me which three actually need a reply from me, and draft each reply in my voice. Don't send anything.
04
Level 4: put it on a schedule
This is the level where it becomes an employee. In the Claude desktop app there's a mode called Cowork with a Scheduled section in the left sidebar. Click New task, pick Create with Claude, describe the job and the cadence in plain English. Tasks run on Anthropic's servers, so they fire even when your laptop is closed. This needs the Pro plan, $20 a month. If you're on ChatGPT instead, the same idea is called Tasks and lives in its sidebar on the Plus plan. You'll know it worked when the first run lands in your Scheduled list at the time you set, with the output attached, and you didn't touch anything.
Every weekday at 7am, read my calendar for today and my unread email from the last 24 hours. Give me a short brief: meetings with who and why, the emails that need a reply, and anything due this week. Keep it under 200 words.
05
The two rules that keep it from going wrong
Rule one: it proposes, you approve. Nothing that sends, pays or deletes runs without your click, even after it's earned your trust. Every one of my own agents drafts into a queue and I clear the queue. Rule two: one agent, one job. The Project above does meeting notes. A second Project does proposals. A third does your morning brief. A single do-everything agent gets vague instructions and gives vague work back. Three small ones with clear job descriptions stay sharp, and when one breaks you know exactly which.
06
Level 5, when you want it: always on
There's a level past scheduling where the agent watches for events (a transcript lands, a form gets submitted) and acts the moment they happen, and a level past that where one agent hands work to a team of them. That's how my business runs: a Mac mini on my desk, agents that read my inbox overnight, cut my reels and file my meeting notes. It takes a terminal and a tool called Claude Code, and I wrote the technical version of this page for it. Link in the sources. Do levels 1 through 4 first. Most businesses I look at never need level 5, and the ones that do are much better at it for having done the four.
SOURCES — ALL OFFICIAL, PLUS THE TECHNICAL VERSION
