Day 50: your first gate into building proactive AI

I wrote this so you can have a working proactive AI by tonight, one that catches what you commit to out loud and hands it back before you even ask for it.

All you need is Claude, one prompt and about 30 minutes.

I did $1M in dropshipping revenue before Moonshot, and I think the money any business loses is rarely in the product. It's in the follow-up somebody said out loud and nobody wrote down.

And I care about it because it covers what no app on your phone tracks: the deadline you agreed to on a call, the follow-up you promised a client, the thing you told your mom you'd handle.

I've been reading every AI launch this month by what it asks of you, and they all ask the same thing.

On September 8, Meta launched Muse, an agent that TechRepublic reports asks for your email, your calendar and your payments. On September 23 it showed Muse Charm, a keychain for the same agent that Decrypt says has cameras and a mic that listens when you tap it.

On September 14, the iOS 27 beta gave Siri what Apple calls personal context, across your messages, emails and photos.

And ChatGPT got four things in eight days, per OpenAI's release notes and BGR's DevDay coverage:

Every one of them needs something typed, timestamped, booked or tapped.

But the things I think matter most get said out loud. The deadline on a call. What your dad told you about his knee on the walk. The interview your roommate is dreading on Thursday.

People who care a lot still drop those, because they lose to logistics. And nothing launched this month hears them unless you tap, type or book first.

You need Claude (the free plan works), Apple Reminders and Apple Notes. Set the project up once on claude.ai, then use it from the Claude app on your phone.

Step 1 is a habit. Steps 2 to 5 are the setup you do tonight, and the whole thing is one prompt you paste once.

1. Say it out loud, right after. In the two minutes after a conversation that mattered, open Claude inside the project and dictate a 60-second recap: who was there, what got said, what you said you'd do, what they said they'd do.

You recap it yourself, in your own words.

If you'd rather talk into Voice Memos, record your recap there and paste its transcript into the project chat.

2. Make a Claude Project called "What I owe people." Then paste this into the project's instructions:

Rule 2 is the one I'd never cut, because a promise Claude can't quote didn't happen.

Rule 3 keeps "we should grab coffee" from turning into a Tuesday reminder you'll resent.

If someone forwarded you this, subscribe here. I send one setup like this every week, with the prompt written out so you can paste it the same night.

3. Keep one note per person. Every recap ends with People notes. Copy each one into an Apple Note named after that person, so the knee, the birthday and the new manager's name all live in one place.

Once a week, add those notes to the project's knowledge so Claude reads them next time.

4. Turn every promise into a dated reminder. Each recap ends with lines like "Remind me to call Mom about Thanksgiving on Sunday at 6pm."

Say that line to Siri, or type it into Reminders. Now the promise has a time on it, and your phone does the nagging.

5. Ask every morning. Open the project and ask "what do I owe people this week?"

If your Claude plan has memory or chat search turned on, it can look across the project's past chats. If it doesn't, keep one note called Ledger, paste each recap's commitments into it, and add it to the project's knowledge.

Either way you get a dated list, oldest first, of what you owe people and what they owe you.

Steps 2 through 5 are the easy part, because once Claude has a recap, the prompt does the sorting.

But the whole setup rests on step 1, and step 1 is you remembering to talk into your phone in the two minutes after a call, while everyone is still saying bye.

So the promise you forget to capture is exactly the one you were going to forget anyway, and the setup can't catch what it never heard.

Step 5 fails the same way from the other side. Claude won't open the conversation. It sits on your ledger until you think to ask.

What you built remembers everything you give it and brings up nothing on its own.

I think a proactive AI has to close both gaps. It hears the promise when you make it and brings it back before you ask.

Albert Borgmann called this the device paradigm in his 1984 book Technology and the Character of Contemporary Life: technology that takes over the machinery of a thing and hands you the result. He worried we'd lose the part that mattered along with the chore.

For a promise, I think the machinery is the remembering, and the part that matters is the call to your dad.

I co-founded Moonshot, so weigh this accordingly.

We're building for step 1. The Moonshot iPhone app listens through your day while listening is on, and keeps a ledger of the real promises and action items it heard, without inventing any.

So say you tell a client on a Friday call that you'll send the revised quote, and you don't recap anything. On Monday morning the reminder fires on its own.

Anything that would send a message for you waits for your yes.

Jalapeño, the earpiece we're building, is up for preorder for early December. It listens only when listening is on, and you can pause it any time and delete what it keeps.

Run the setup for seven days. Every time you remember a promise too late to have recapped it, put a tally on a sticky note. Count them on day seven.

An AI that waits for you to remember is still asking you to remember.

Next week's setup lands right here in your inbox. If you're reading this on the web or someone forwarded it to you, subscribe so it comes to you and you don't have to remember to go find it.