When someone asks a twin a question, the twin finds the parts of its creator's material most relevant to the question, places them in front of the AI model together with the creator's voice and tone guidance, and asks the model to answer from that material as the creator would. The model's answer is shaped by three things: what material exists on the topic, how well the voice guidance captures the person, and the question itself.
The five steps
- The question arrives. Someone types a question on the twin's page, in any of the languages the twin supports. The whole conversation so far comes with it.
- The relevant material is found. The twin searches the creator's uploaded material for the passages that bear on the question: a few paragraphs of a book, a section of a talk, the creator's written answer to a similar question.
- The model is briefed. Those passages are placed in the AI model's context, alongside the creator's answers to the guided questions about their voice, how formal they are, how they open and close, what they would never say.
- The answer is written in the creator's voice. The model composes a reply from the material, in the creator's manner, and in the language the person asked in.
- The conversation is stored according to its mode. In normal mode the creator can read it and follow up; in private mode they cannot. Either way, nothing is trained into the model.
What makes an answer good or weak
Because the twin answers from material, the quality of the answer is mostly the quality of the material on that topic:
- Well covered topics produce grounded, specific answers. If you have written three chapters on forgiveness, the twin will have a lot to work with.
- Thin topics produce general answers. If you have only mentioned something in passing, the twin has little to go on. This is where a twin is most likely to drift toward what any AI would say.
- Your most-asked questions are the best investment. Writing out, in your own words, how you answer the thirty or forty questions you get most is often worth more than uploading another book.
Why it sounds like you
Two things carry your voice: your material, which is full of your phrasing and your way of thinking, and your answers to the guided questions during setup. The model is good at matching a voice when it has examples of it in context. Why a twin sounds like you goes deeper.
What the twin does not do
It does not look things up on the open web, so it will not bring in facts you never gave it. It does not learn from conversations. It does not cite its sources in the answer. It does not hand the conversation off to you or notify you; you read the shared conversations in your dashboard on your own schedule. These limits are deliberate: a twin is meant to represent your point of view, not to be a general search engine with your name on it.
Improving it after launch
Read normal-mode conversations in your dashboard. Where an answer was vague, upload something you have written or said on that topic. Revisit the voice questions if the tone is off. The twin improves because the material behind it gets more complete, not because it has "learned" on its own.
Common questions
Does a twin search the internet to answer?
No. A twin answers from the material its creator has given it, plus the conversation so far. It will not bring in outside facts or news the creator never addressed.
Can a twin answer in other languages?
Yes. People can talk with a twin in their own language, including English, Spanish, French, German, Portuguese (Brazilian and European), Vietnamese and Korean, even if the creator's material is in English.
What happens when a twin is asked something the creator never covered?
It has little material to draw on, so the answer will be general or the twin will say it does not have the creator's view on that. Adding material on the topic is the fix.