Build Your Own SIM Feed

In the first two parts I argued that personal AI apps are about to look like a social feed: a stream of what your AI did for you, tagged IRL, and what it would do for you, tagged SIM, each with one button. Part two argued that the button is the whole business, because every tap hands off another job. This part is practical. Here is how to build a crude version for yourself this week.

This morning I did exactly this with my own email, calendar and bookings. The cards that came back were not chores. They were jobs I had been carrying for weeks without noticing how much they weighed. That is the test of whether you built it right.

What you need

• A personal AI that can read your accounts. Claude, ChatGPT or Gemini, with your email and calendar connected through its connectors or apps settings. The new always-on agents will do this by default; today you turn it on yourself.
• Two sources, at minimum. Email and calendar carry most of a life. Add notes, a voice recorder, texts or a bookings system later if your AI can reach them.
• Twenty minutes, once. After that it is a minute a week.

Leave out anything you would not hand a new assistant on their first day: banking passwords, ID numbers, medical portals. A SIM feed works fine without them, and nothing in it ever needs a password.

Step 1. Ask it to notice

Paste this into your AI. It is the whole engine. Read it once before you do, because every line is there for a reason: it asks for jobs instead of chores, it demands evidence so the AI cannot make things up, and it forbids starting work, so nothing happens until you agree.

Read my last 30 days of email and calendar, and the next 14 days of my calendar.

Find jobs in my life that I keep doing, keep putting off, or keep paying for in time and stress. Not small chores. Whole jobs I would gladly hand to a capable assistant: preparing for something important, running a recurring responsibility, replacing a tool I keep fighting, keeping people from slipping through the cracks.

For each one, write a SIM card:
- Title: the job, as a role you would play for me ("Be my legal advisor", "Run my move").
- One line: what comes off my plate.
- What you saw: 2 to 4 specific pieces of evidence from my own email or calendar, with dates.
- What I would get: 3 to 5 concrete deliverables.
- Time back: an honest estimate.
- Still checks with me: what you will never do without my yes.

Give me the best five, most valuable first. Do not invent anything you did not see. Do not start any work yet.

Step 2. Read the cards like a hiring manager

You are not reviewing a to-do list. You are interviewing for a job, and the AI is the candidate. Each card should hold up to four questions.

If the cards come back small, it is not a failure. Tell it so: these are chores, not jobs; find bigger ones. If a card cites something that is not in your inbox, delete it and tell the AI it guessed. Two rounds of that and it learns what you mean.

Step 3. Say the two words

Pick one card. Not five. One. Then hand it over with a prompt that does what the button will do in the apps that are coming: hire the AI for the whole job, take away its excuse to keep asking you questions, and keep anything irreversible behind your approval.

Try it: [card title].

Do the whole job now. Do not ask me questions; make every decision you reasonably can, and tell me afterward what you decided and why.

Anything that sends a message, spends money, signs, files or cancels: prepare it completely as a draft and list it for my approval.

When you finish, give me one IRL line: what is now off my plate.

What comes back should feel strange the first time. Not a plan for you to carry out. The thing itself, done, with a short list of what needs your yes. That feeling, a weight you did not have to lift, is the whole product. Everything else is plumbing.

Step 4. Keep the feed alive

A feed is something that keeps arriving. Once a week, or every morning if your AI can run on a schedule, give it the standing prompt below. Ask it to remember your Hired list and the cards you passed on, so it never pitches the same job twice.

Run my SIM feed.

1. IRL: tell me, one line each, what my hired jobs did since last time.
2. SIM: notice up to three new jobs since last time, using the same card format. Skip anything already on my Hired list and anything I passed on.
3. Anything waiting for my approval, at the top.

Remember my Hired list and my passes between runs.

Unplug [job]. Stop doing it, tell me what you were in the middle of, and go back to asking me.

That fourth prompt matters more than it looks. People hand over more when they know they can take it back. Keep the off switch close and you will find you rarely use it.

The mistakes everyone makes first

1. Writing the cards yourself.

   The moment you start listing what your AI could do, you are back to staring at a blank box. Let it notice. Your job is to agree or pass.

2. Accepting chores.

   A card that saves two minutes trains you to ignore the feed. Ask for jobs worth an afternoon. They are there; most of us just stopped seeing them.

3. Hiring five at once.

   Start with one, see it done, then the next. Trust grows in steps, and the steps are where you learn what to hand over.

4. Letting it ask permission for everything.

   If every result comes back as questions, you have hired a very polite intern. Tell it to decide, and keep only the irreversible things behind your yes.

Where this goes

What you have built is a hand-cranked version of what the always-on personal agents will do by themselves: watch, notice, audition, and wait for a tap. When those apps reach your phone, you will already know how to read a SIM card, which ones to hire, and how much of your life you are willing to put on the other side of one word. Most people will learn that by accident. You will have learned it on purpose.

Notice, don’t imagine. Then say yes.