Personalizing AI: Make It Work Like You Do
Why AI answers feel generic, and the four levers — instructions, context, memory, and examples — that make an assistant work the way you do. With a method for checking that it actually helped.
The assistant does not know you
It is not being lazy. It genuinely has nothing to go on. Ask an assistant “how should I open this email?” and you get advice that would suit anyone on earth. It feels like the model is holding back. It is not. When you opened that chat, the model received your sentence and almost nothing […]
The four levers
Four levers, four different jobs People often reach for one lever and conclude that personalization “does not really work”. Usually they pulled the wrong one. Each lever answers a different question, and they are not interchangeable. 1. Instructions — who am I and how should you answer me? A short standing description that is attached […]
Writing a profile that actually changes the output
Most custom instructions are wasted words The usual first attempt looks like this: “I am a professional. Please be helpful, accurate, and concise.” That changes nothing. It contains no fact the assistant can act on, and “be helpful” was already the plan. A profile earns its place when every line would visibly change an answer. […]
Teaching it your voice by showing, not telling
Describing your style barely works. Showing it works immediately. Ask for “a warm but professional tone” and you get someone’s average of those words. Paste two emails you actually wrote and ask for a third in the same voice, and the difference is obvious in one attempt. This is the examples lever, and it works […]
Testing whether personalization actually helped
“It feels better” is not evidence You spent an hour on a profile, a project, and a voice sample set. Did the answers improve? Almost nobody checks, and the reason is that checking sounds like it needs a methodology. It needs about fifteen minutes. Build a tiny personal eval set Write down five tasks you […]
The limits: privacy, drift, and being agreed with
Personalization has two failure modes and one hard boundary The boundary: what you hand over Every lever works by putting information somewhere a system will read it later. That makes “what should I share?” a real question rather than a paranoid one. The principle to borrow is data minimisation, set out in Article 5 of […]