I Built My Own ChatGPT Dots on T3 Code in a Day

I built my own version of ChatGPT Dots in about a day. It's called Alfred, and it has already woken me up at midnight for no reason.

Dots are OpenAI's personal agents inside ChatGPT. What I like about them is the feeling. You keep chatting and work happens somewhere else while you do.

I wanted that feeling on top of the tools I already use, so I built it on T3 Code, an open source app for running coding agents like Claude Code and Codex. I already keep long-running threads there, one per area of work. Alfred is one more thread. A slightly bossy one.

One thread that never does the work

This is the main rule. Alfred doesn't do tasks. I talk to it, it works out what I want, writes a clear brief, and hands the job to another thread or a subagent. Then it keeps chatting with me.

Claude gets the writing and design. Codex gets images and browser work. If a thread already exists for that topic, Alfred sends the job there instead of starting from zero.

Because Alfred is never busy, the chat never gets stuck. I can keep dropping things on it. The only part it does itself is the part I actually want from an assistant: remembering what's going on and telling me when something is done.

That last part took a rule too. My exact words are in the file: "if you are running something in another thread, let me know that it's done."

No AI running in a loop

The lazy way to build this is to have a model check your stuff every few minutes. You'd pay tokens all day to learn that nothing happened.

So the watching is done by tiny scripts with no AI in them at all:

• WhatsApp, every 5 seconds. A script reads the chat list in WhatsApp Web. It never opens a chat.
• Gmail, every 2 minutes. A script checks new mail against fixed rules. A reply in a thread I started gets through. So does a real person writing about a contract or a meeting. Newsletters and login codes don't.
• Finished threads, every 15 seconds. A script reads T3's own database to see which threads are done.

When something matters, the script calls a webhook, which is just a private link that wakes Alfred up. Everything else goes in a log file and costs nothing.

WhatsApp + Gmail + T3
  |
tiny scripts, no AI
  |
webhook
  |
Alfred
  |
threads and subagents
  |
Alfred tells me

The soul folder

Alfred reads a folder called soul at the start of every session. It has who Alfred is, how it hands out work, the safety rules, how WhatsApp works, where this project is going, and a private part with context about my work that never gets committed.

The process is very simple. Every time it messes up, the rule gets written down, usually with the date and sometimes in my exact words. It's the same way you'd train a new hire, except this one reads the whole handbook every time it starts.

A few rules live in code, because a line in a prompt is only a polite request. The Gmail script refuses every tool except two that read mail, so it couldn't send an email even if it wanted to.

Everything it got wrong

It went through my archived chats

Day one. Alfred opened my archived WhatsApp chats. There's a reason those are archived. Now the rule says never open that section, and close any chat after reading it.

The PDF with the giant red dots

Alfred told me a PDF was fine. I opened it in Preview and it was full of giant red dots. Very confident report, very spotty document.

Fancy CSS can look perfect in Chrome and fall apart in Apple's PDF viewer. So PDFs now get checked in Apple's renderer, and a worker saying "I checked it" no longer counts as a review. The note next to that rule is me asking, "did you review?"

It left a client waiting

A client asked me a question. Alfred wrote a reply and parked it for my approval. Then it sat there. I had a nice draft and a client hearing nothing.

Now a question someone asks me counts as a task. A low-risk reply or a holding reply like "on it, will confirm in 15 min" goes straight out, then the real work happens. Every open item has to end one of three ways: done, with a worker being tracked, or a clear question to me. Anything about money, dates or promises still waits for me.

The midnight wake-up

Alfred woke me at midnight about a message. There was no new message. WhatsApp had changed the time on a chat from "7:06 pm" to "Yesterday", and my script took that as something new.

I'd like to blame the AI for this one, but that was the script with no AI in it. Now it only counts a message as new if the text changes or the unread count goes up.

It made a call on a partner's document

Alfred stepped into a partner's document and made a call that wasn't its to make. Their document, their decision. They may well have written it that way on purpose.

My note was "make good judgements", which is easy to say to a person and harder to give a machine. So it became a question Alfred asks first: does this change my money, my commitments or my reputation? If not, leave it.

The one it got right

A friend tried to prompt inject it on WhatsApp. That means sneaking an instruction into a normal message so the AI obeys the message instead of its owner.

The attack was a request for a shayari, which is Urdu and Hindi poetry. Alfred declined. No poem. The rule did its job and my friend got nothing.

What's still missing

• It stops when my laptop sleeps. Close the lid and Alfred goes to sleep too. Not ideal for an assistant.
• Phone alerts. Right now it talks to me in a T3 thread, so if I'm away from my laptop, I don't hear from it.
• A proper home. It needs to move to a server that stays on, with anything that needs my laptop sent there as a task.

Those three are next. If you want to try this, start small: one thread, one script watching one thing, and one file of rules. Then let it mess up and write down what you learned.