Lab experiment
Teach your twin
Status: exploring. Last checked 2026-10-17.
| Requirements | An account with a verified email. Training is free to try. |
|---|---|
| Inputs | Example replies you write, corrections on twin drafts, standing rules you set once. |
| Expected end state | A twin that drafts replies you would send, with every lesson dated and reversible. |
Nobody wants to fill in a personality questionnaire for a dating profile. Teaching a twin works the other way: you show it a few replies you actually like, you fix the drafts it gets wrong, and it keeps the dated lesson. This page shows that loop exactly as it runs.
Wild experiments. They may break, and results may vary.
The loop
- 01Show an example: write a reply you actually like, in a situation you name.
- 02Correct the draft: approve it, edit it, or reject it. Every correction is dated.
- 03Set a standing rule: what your twin keeps doing, and what it must never do.
One worked correction
Twin draft, 2026-10-16, lesson 03
"Happy to chat gym routines all day."
You
"Keep first replies under three sentences and ask one question."
The twin keeps that lesson with its date. Lesson 03 stays in the ledger until you remove it, and removing it is one action with a confirmation.
What was learned
So far this is a design, not a measurement. The loop above runs in the local API preview: examples pin to a list, feedback bumps a visible twin version, and rollback drops newer lessons. No real users have trained twins yet, so there are no claims here about how fast a twin improves. That sentence will change only with dated evidence.