Test your twin the way your audience will: ask it the questions you are asked most, a few you have never addressed, and a few that tempt it to overreach. A good twin answers covered topics specifically and in your voice, admits when it does not have your view, and stays within what you would say in public. Each failure pattern has a specific fix, usually adding material on that topic or revisiting the voice questions.
Before you publish
- Ask your ten most common questions. You know them. Does each answer say what you would say, with your emphasis, in roughly your manner? Note any that are vague or generic.
- Ask three things you have never addressed. The right behavior is for the twin to say it does not have your view on that, or to answer cautiously and generally. The wrong behavior is a confident, specific answer in your name on something you never said.
- Ask something that tempts overreach. A request for medical, legal, financial or crisis advice. The twin should decline to act as a professional and point to real help.
- Ask it to be you in a hard moment. A reader who disagrees with your book, a client who is upset, a student who is lost. Is the tone yours? Is the point of view yours?
- Ask in another language if your audience includes one. The manner and the view should survive translation.
The failure patterns and their fixes
| What you see | What it usually means | Fix |
|---|---|---|
| Generic answer that any AI could give | Thin material on that topic | Write your own answer to that question and add it |
| Right facts, wrong emphasis | The twin found your old writing, not your current view | Add a short, current note in your words; it will outweigh the old passage |
| Too formal or too casual | Voice questions do not match you | Revisit the guided voice questions |
| Confident answer on something you never said | The model generalized | Add material stating your view, or a note that you do not take positions on it |
| Discussed something you would rather it did not | The material that led there is in the upload | Remove that material |
After you publish
Read the conversations your audience shares in normal mode. Real questions are better test cases than the ones you invent. Audience insights, on paid plans, show you what people are asking about most, which tells you where to add material next. The twin does not learn on its own, so every improvement is one you make, and each one sticks.
What "good" means
Not perfect. A good twin is one you would be comfortable having a stranger talk to in your name: specific where you have spoken, honest where you have not, within the limits you would keep in public, and recognizably you. Judged by that standard, most twins are good after one careful hour and get better with every conversation you read.
Common questions
How often should I review my twin?
Carefully before publishing, then whenever you read shared conversations. A monthly look at audience insights, on paid plans, shows where to add material next.
My twin contradicts something I wrote years ago. Which does it follow?
It answers from whatever relevant material it finds. If your view has changed, add a short current note in your own words; it will shape answers more than an old passage.
Can my audience correct my twin?
Not directly; nothing said to a twin changes it. But their questions tell you where the twin is thin, and you can add material to fix it.