How to Make a Personal AI Assistant That Knows You

You have already met the version that does not know you. It writes a reply that is technically fine and unmistakably not yours. It flags a newsletter as urgent. It books over the hour you keep free every Thursday. Nothing is broken, and every output needs enough editing that you start wondering why you bothered.

The gap is not intelligence. It is context. A new human assistant would be exactly this useless on day one and nobody would panic about it, because everyone understands that a person needs a few weeks to learn the job. The only difference is that you can shorten those weeks to days by writing things down.

On Sistava, the Personal Assistant role has her own private memory attached to your workspace. Alice remembers what you told her, what she has seen in your connected accounts, and what you corrected last Tuesday. She is not a shared chat that starts from zero every visit, which is the entire reason personalisation sticks.

What does it mean for an AI assistant to actually know you?

It means four concrete things, not a feeling. She knows who your important people are by name. She knows how you sound in writing. She knows what your week is supposed to look like. And she knows what you never want to be interrupted for. Everything people call personality is downstream of those four facts.

Notice that none of them are mysterious. They are things you could write on an index card. The reason assistants feel generic is almost never that the facts are hard to state, it is that nobody ever stated them, and a worker with no information defaults to the average of everyone.

What are the three layers of knowing you?

Told, observed, and corrected. What you told her is your brief. What she observed comes from the inbox and calendar you connected. What you corrected is the feedback you give when something comes back wrong. Each layer covers what the other two miss, and the third is the one people skip.

The layers compound. A rule you wrote on day one gets sharpened by something she saw on day three and corrected on day five. That is why an assistant at week four feels like a different worker to the one you met on Monday, even though nothing about her changed except what she knows.

How do you teach it to sound like you?

Paste in three emails you were happy with, then describe the rules you follow without thinking. Length, greeting, sign-off, whether you use exclamation marks, whether you open with pleasantries, what you never say. Examples plus rules beats either one alone, every time.

Pick the three examples carefully. One to a client or customer, one internal, one where you had to say no to something. Those three cover most of the range you actually write in, and the third one is the most useful because saying no is where tone goes wrong fastest.

Then be blunt about the small things. "Never open with I hope this finds you well." "Sign off with just my first name." "No more than five sentences unless I say so." These sound petty written down and they are the difference between a draft you send and a draft you rewrite.

A worked example: Ken and the two clients with the same first name

Ken is a commercial property surveyor working solo in Vancouver, mostly on warehouse and light industrial buildings. He has two clients called David: one is a developer who wants everything in bullet points and answers within the hour, the other runs a family logistics business and finds bullet points cold. For a week his assistant treated them identically, which was worse than useless.

The fix was two sentences added to the brief, written the morning he noticed. David at the development firm gets bullets, numbers, and no preamble. David at the logistics company gets full sentences, a line of small talk, and never a bulleted list. From the next run onward the drafts came out right, and Ken has not thought about it since.

The observed layer added something he had not thought to write. His assistant noticed that he replies to site engineers within about twenty minutes during the day and to everyone else in the evening batch. She started ordering the morning brief the same way. Ken had never articulated that habit, and would not have, because he did not know he had it.

By week three his brief had grown from twelve sentences to about thirty, and roughly two thirds of drafts went out with no edit. He is clear that this did not happen because the assistant got smarter. It happened because he spent about four minutes a week saying why something was wrong, and those four minutes were the entire investment.

How do you build the memory in the first fortnight?

Deliberately, in five steps, over two weeks rather than one sitting. The setup is thirty minutes, but knowing you is a fortnight of small corrections, and treating it as a one-off configuration job is the most common way people end up with a generic assistant.

1. Hire the assistant and connect two accounts — Email and calendar. This is the observed layer and nothing else can substitute for it. She cannot learn how you actually work from a description of how you work, any more than a new colleague could.
2. Write the brief, including the petty details — Who you are, what you are working on, your important people by name, your tone rules, and what must never interrupt you. Include the small preferences that feel too minor to mention. Those are the ones that make drafts sound like you.
3. Paste in three emails you were proud of — One external, one internal, one where you turned something down. Examples teach voice faster than adjectives do, because you cannot describe your own rhythm accurately and you do not have to.
4. Correct out loud for two weeks — Every time something is wrong, say why in one sentence rather than just fixing it silently. A silent edit teaches nothing. Ten seconds of explanation stops the same mistake permanently.
5. Reread the brief at the end of week two — It will have grown, and some of it will contradict itself. Ten minutes of tidying, deleting what turned out not to matter and sharpening what did, is the single highest value thing you do in the first month.

What will it never know about you, and should not?

It will never know what you did not write down or correct, and it will never know why. Memory is a record, not intuition. An assistant that has learned you well is still guessing at the reason behind every pattern, and that gap is exactly where confident mistakes come from.

• The unstated reason. She sees that you always reply to one person quickly. She does not know it is because they are ill, or because they nearly left last year. Say why when the why matters.
• The room. A tense negotiation, a client who is quietly unhappy, a colleague having a bad month. None of that is in the text. Hard conversations stay yours.
• What you deliberately keep out. You can exclude senders, labels, and whole accounts from her view, and you should for anything genuinely private. What she cannot see, she cannot use, by design.
• Whether a specific fact is true. Knowing you well does not make her accurate about numbers, dates, or names. Familiar tone plus a wrong figure is a worse outcome than an obvious stranger, so keep checking specifics.
• Anything with real consequence. Money moving, legal, medical, and tax decisions stay with you and a qualified human, no matter how well she knows your preferences.

There is one failure mode worth naming. If you write down how you wish you worked instead of how you actually work, she will learn the fictional version and every draft will be slightly off in a way you cannot place. Write the real one, including the parts you would not put in a bio.

Knowing you is not a feature you switch on. It is a fortnight of saying why, one sentence at a time, until the drafts stop needing you. Write the honest page, connect the two accounts, and spend four minutes a week correcting. That is the whole method.