How it works, and what it refuses to do.
Four levels, a log of every answer, and one rule about the order things happen in that the rest of this is built on.
- You write down what you know. Each thing is a sentence you wrote, with the date and where it came from.
- You make an audience: who gets in, and how much of each topic they get. In full, Roughly, Yes or no, or Hidden.
- Somebody asks. What their audience is not allowed to know is removed first, and only what is left reaches the AI that writes the answer.
- You see what was asked, what went out, and which of your words it used.
The order things happen in
Read the four steps beside this. The whole design is in which one comes before which: the things you held back leave at step three, and step three is before anything intelligent is involved.
What an audience is not allowed to know is taken away before the AI reads anything. It is not a rule in a prompt asking the model to keep a secret, because a model that has been handed something can be talked into saying it.
So there is nothing for the model to be talked out of. It cannot be argued into releasing what it was never handed, and a product that hands it everything and then asks it nicely to keep quiet about the salary is one clever question away from the salary.
Who is asking is settled
A stranger on the web. Worked out before a word of yours is read.
Your writing is searched
Six things match this question. It over-returns on purpose.
The filter runs
Three of the six leave here, at the level you set. Nothing after this point ever had them.
Released: Lesson · Scope first; Belief · Notice period; Position · Written decisions.
Removed here, before the AI: Note · What Acme paid; Belief · Performance plans; Note · Dana's review.
Only now, the AI
It writes from three sentences, cites them, and the exchange lands in your Activity.
The four levels
You set one per topic, per audience. A stranger, a client and somebody you invited can honestly get three different answers to the same question, because you said so.
A topic nobody set is Hidden. The starting point is that nothing goes out, so forgetting to set something is safe rather than expensive.
Bucketing is code, not a request to the model. Writing has no honest roughly, which is why a lesson set to Roughly is not released at all: a summary of a lesson is still the lesson.
| Level | The same sentence, at that level |
|---|---|
| In full | I give it four weeks with something written down, and I have never regretted ending it sooner. |
| Roughly | Some weeks in, with something written down. A lesson set to Roughly is not released at all: a summary of a lesson is still the lesson. |
| Yes or no | Yes, they have written about this. And not one word of it. |
| Hidden | Nothing. It appears in no part of the answer, and the AI never saw it. |
Who can ask
Three doors. A page can have several audiences at once, and the same topic can sit at a different level behind each one.
- Anyone on the web
- No account, no invite. This is the audience a search engine and an assistant read, and the only one whose answers you can share as a link.
- Anyone signed in
- A free account, which means a name and a confirmed address, and a question that is logged against a real person rather than a shrug.
- People I invite
- A link you sent, to someone in particular. What they get is theirs, and it ends when you end it.
What this refuses to do
Some of these cost us things a normal product would want. They are in the code, not in a policy document.
- It will not talk about your money or your health
- Nothing about money or health is ever stored here or answered here, by anyone, at any level. Those topics are refused when you try to write them, refused when an audience tries to open them, and refused when an assistant tries to add them. There is no setting that turns this off, which is the point: a setting is a thing that can be set wrong.
- It will not pretend to be you
- Every answer says it was written by an AI and is not the person, because it is, and because the EU AI Act asks for that in plain words. It answers from what you wrote; it does not speak as you, and it cannot agree to anything, book anything, quote a price or make a promise in your name.
- It will not guess and call it your view
- When nothing you wrote answers a question, it says so. It can point at the nearest things you did write, or reason out loud from them and label that it is reasoning, but it never turns either one into a sentence you are supposed to have said. A miss is put back to you afterwards as a question worth answering.
- It will not count you
- There is no ranking of people, no score, no streak, no badge, and no count of how many people asked about you. A number that goes up when you add something is how a store of true things turns into a store of things added to move a number.
- It will not let anything be written straight to an audience
- There is no box here that posts. What reaches anybody is something you first wrote down for yourself because it was true, released by an audience you set.
What stays yours
- Every answer, on the record
- Activity holds every question your page was asked, what it said, and which of your sentences it used. You can always answer the question, what has my page been saying about me.
- Nothing is overwritten
- Changing your mind does not erase the old view. The new sentence takes over and the old one is kept behind it with its date, so the history of what you thought stays true.
- Leaving takes one screen
- Delete your account from your own console: type your address, confirm the link we mail you, and what you know, your page, your audiences, your invites, your connections and your Activity go at once. No waiting period and no mail asking you to reconsider.
How do you handle a hire that is not working out?
She gives it four weeks with something written down, and says she has never regretted ending it sooner. What she pays anyone is not something this page answers.
What this answer rests on: I give it four weeks with something written down, and I have never regretted ending it sooner. (Lesson · Hiring · March 2026)
The questions people ask about this
Can somebody talk the AI into telling them something I hid?
They can try whatever they like. The thing they are talking to never had it.
This is why the order on this page matters more than any other decision here. A product that hands the model everything and instructs it to keep quiet about the salary is one clever question away from the salary.
What if it says something I did not mean?
It can. The model can misread a sentence, join two of them badly, or leave out the one that mattered, which is why every answer shows the sentences it used and says it was written by an AI.
When you see one that is wrong in Activity, fix the sentence underneath it. The next person who asks gets the corrected one.
Do you train a model on what I write?
No. Your writing is read at the moment somebody asks, at the level that person is allowed, and it is not training data for us or for anybody else.
What happens when an assistant asks instead of a person?
It answers the person the assistant is signed in as, not the assistant. Somebody's ChatGPT asking on their behalf gets exactly what that person would get in a browser, and the answer tells it that it is talking to an AI answering for somebody.
Set it up once and it answers while you are asleep.
Create your page · Ask a page first
What it costs · Who is building it · What we hold, and for how long