Is it writing as me?
It is writing what you meant, for one specific reader. Length, order, bluntness and how much background it gives are all fair game. What is off limits: the meaning, the facts, the numbers, the names, the dates, the conditions and the ask. If the only way to make something easier to hear is to make it less true, it stays true and lands hard. A message that stings and is understood beats a comfortable one that is not.
Where does my own voice go?
Nowhere. It carries your words, your idiom and your language, and follows you when you switch mid-sentence. It is not smoothing you into house style. It is deciding running order and weight for one reader it happens to know something about.
What does it write down about the people I talk to?
Three lines each, at most. What form of a message they take in. What sort of message they bother to answer. What they still remember a week later. It never records what kind of person they are, and it never keeps the content of what passed between you as a description of them.
How do I get the number back to zero?
Answer things, or say out loud that you are not going to. The second one removes it from your count and does nothing else — no note is sent, the other person is not told they were dismissed, and it is never marked as answered, because it was not. Zero is always reachable.
What if they are not on Chatter?
Then it cannot reach them and it says so, immediately, instead of failing quietly. It will not invent a person to send to and it will never tell you something went out when it did not.
There is genuinely no list of messages?
Genuinely. Nothing to scroll and nothing to open. What arrived is read to you when you ask for it, and the dot in the menu bar carries a number so you know there is something worth asking about. The history exists and you get at it by asking, not by browsing.
Can it send something it should not?
Money, passwords and keys, bad news about a person, anything long enough to be a document, and any name that could be two people — all of those come back to you first, out loud, before anything moves. That check is a plain function with tests around it, not a decision the model is trusted to make.
Is it only for messages?
No. It makes things and it drives this machine. Pictures, video with sound, a song with the words you wrote actually sung, a Word document Word opens, a spreadsheet, a deck, a PDF, code, an SVG, a 3D model a picture at a time that Blender will open, and a recording turned back into text. On the machine itself: dark mode, volume, brightness, Wi-Fi, the desktop picture, what is running, a note into Notes, and the page in your front browser tab. Every one of those sits behind a switch, and when a switch is off it says so instead of trying.
What does it run on?
macOS 13 Ventura or newer on Apple silicon or Intel, and Windows 10 or newer. The Mac build is signed and notarised, so it opens without an argument. You sign in with Google, and only because reaching another person needs an identity — everything else works signed out.
What happens when it works in the background?
You hand it a job — go through a folder, build a script, chase down a comparison — and a worker runs it while you keep talking. The worker has the same switched-on tools you gave the assistant and its own folder. Whether it stops to ask is one switch in Settings, "Let jobs act without asking". It ships on, and then a job sends, mails, messages and deletes as it goes. With it off, the job stops, the assistant reads you exactly what it wants to do, and it waits for your yes. Every job is listed in Settings with a cancel.
