The assistant I wanted doesn't have a chat box
For the last three months I have had an assistant that works while I'm asleep. It isn't a person and it isn't an app I open. It is a handful of scheduled jobs that read my email, my calendar, a few job boards and a Notion workspace, and every morning it leaves me a short list: here are the three things only you can decide, here is what I already did about everything else. It is a personal assistant that gets a little more access each month, as trust builds.
I built it because I was between jobs and had the time, but the longer I run it the less it feels like a side project and the more it feels like the actual product. So this is the first of a few posts about what it is, how it works, and what I think it says about where AI assistants are going.
The chat box is the wrong shape
Every AI assistant I have used asks me to start. I open a window, I remember what I needed, I type it, I read the answer, I close the window. It is like having a very good reference librarian. But a reference librarian is not what I was missing. What I was missing was the colleague who notices that the presale for a band I like opens Thursday at 10, who already read the forty-message thread with the architect, who says "you have two decisions today and I prepped both."
The difference is who holds the initiative. In a chat box the human holds it, which means the assistant can only ever be as useful as my memory of what to ask. As anyone who is married can tell you, there is a world of difference between doing what your spouse asks and noticing that something needs doing and handling it before they have to ask. Just ask my wife which one she'd like more of. The assistant I wanted holds the initiative itself and hands it back to me only at the points where my judgment is required.
I don't think the chat box is wrong, exactly. It is a great starting interface, the way the command line was. But the parts of my life the assistant helps with don't change much from year to year: health, money, the house, work, the people I care about. Fixed categories want a fixed place on the screen, not a scrolling conversation. So the design question changed from "what should it answer?" to "what should it be allowed to do on its own, where exactly does it stop and wait for me, and what does the thing I look at every morning actually look like?"
What it actually does
Here is the current set, in the order I built it.
The Command Center. A single page that shows every open project in my life as a tile in a fixed place, like a site plan, with one marker per item that needs something. The rule for the visual hierarchy is strict: a hard deadline, like a payment due, inside three days is the loudest thing on the page; things waiting on me glow softly; items the agents are working on recede; soft deadlines never turn red; fun projects, like a purchase advisor I'm building, live in their own quiet corner. If I only look at it for five seconds I should know whether anything is on fire and whether anything is mine.
The attention queue behind it. Every item on the page answers two questions: is this on me or on the agent, and if it's on me, what kind of answer does it need. The agent can move an item into my column with a recommendation at the top and whatever it already prepared underneath. I can reply to any item right from the page, so the loop is: it prepares, I decide in one line, it carries on.
Project Books. Every project, from a kitchen renovation to the job search to a dinner planner, gets one Notion page with the same four parts: a brief, a decision log, open questions, and a session log. An agent reads the book before it does anything and writes a line in the log when it finishes. This is also how I talk to it. Instead of returning to one ever-growing chat thread that eventually forgets its own beginning, I open a fresh conversation, point it at the book, and it has the latest context in a few seconds. Memory as a page you can read beats memory as a transcript you have to scroll.
Job Radar. A script pulls every product management posting from about twenty companies I think are doing interesting work, every morning. A second job reads each posting against a rubric I wrote about what I actually want from the next role (room to originate bets weighs most, then level, then problem space), scores it 1 to 5, and writes anything worth a look into a table with a one-line "why" a recruiter could read. Anything that scores a 4 or 5 lands in my attention queue. I mark each one Interested or Not for me, sometimes with a note, and the rubric gets revised from my calls. It has been running since September and reads about a hundred postings a week so I don't have to. Every so often I review the roles it passed on, looking for false negatives. So far, none.
Smaller ones. A morning brief. A weekly sweep of what fell through the cracks. The renovation paperwork: proposals, permit timelines, the vendor email that needs a reply. Concert presales for artists I'd buy tickets for, with the presale code sitting right on the tile. Not everything has to be useful.
The rules it runs under
The part I'm proudest of is not any one agent. It is the small set of rules they all obey, which is what makes me increasingly willing to let them run unattended:
- It never sends anything to another person. It drafts the email, in my voice, and leaves it for me. I have had it draft a decline to a recruiter, a note to our architect, a message to our daycare. I read and send every one myself. This is the rule that makes everything else safe, because the blast radius is limited to a bad draft.
- It never closes my decisions for me. It makes recommendations. It cannot mark a job "Not for me" or tick something off my to-do list.
- It never spends money. It can pressure-test a purchase or comparison shop for the best price. The card stays in my wallet.
- It writes down what it learned. Each project has its book. And it watches the delta between the emails it drafts and the ones I actually send, so it sounds more like me over time. This might be the most useful feature of all.
- It brings me decisions, not updates. A status report is something I have to read. A decision with a recommendation at the top and the prep already done is something I can act on in thirty seconds.
Those rules are the actual product. The scripts are replaceable.
What surprised me
The hard part is not the model. It is the rubric. The scoring job got good the week I stopped describing jobs I'd like and started writing down what I was actually optimizing for, including the uncomfortable parts: pay floor, hours, how AI-native a team is. The model was fine from day one. My own clarity was the bottleneck. I suspect this is true of every "AI agent" anyone is building right now.
It is agile the way the manifesto meant it. The Command Center has had something like a hundred changes since I started, and there are a hundred more I want. The pattern is always the same: I use it for a few days, something feels off, I describe the change, it ships in minutes, I use it for a few more days. Filters. Sorting. A checkbox to mark things done. Replying inline instead of switching apps. A rule that every recommendation leads with the answer and puts the evidence underneath, which is Barbara Minto's pyramid principle applied to a to-do list. No single change was the breakthrough. A thousand small tweaks were. For most of my career, getting a product in front of a user and iterating a hundred times in a few weeks was the thing we said we did. This is the first time it has been literally true.
Product taste is the scarce input. Because building is cheap now, the time goes to deciding what should exist, where it should live on the screen, and what the agent should be allowed to touch. I have spent far more hours on what the morning page looks like than on any script behind it, and I think that ratio is right. The agent is only as good as the judgment about what to show me and when.
Where this goes
The obvious next version is one that runs the same loop for health, money, the house and the people I'm bad at keeping up with. I'm building those next and will write them up as they land. The Job Radar scaffold goes up on GitHub in a couple of weeks with the rubric and the evals I use to check the scoring, because the interesting part is not my code but the shape: fetch, score against a written point of view, bring only the decisions to a human.
There's a second thing I noticed along the way, about what happens to discovery when the assistant is the one choosing the gym, and it deserves its own post. Soon.
If you're building something in this shape, or run a team that should be, I'd like to compare notes. Email is at the top of the page.