…
Nothing is saved. This conversation lives in this tab and ends with it — the server keeps no record of it at all.
How this is built
Six markdown files are the character. The first four are what a virtual influencer needs; the last two are what an explainer needs, and they are the whole difference. A companion character has no truth constraint — whatever it says is the product. This one is only worth having if its claims can be checked.
- persona.md
- Who it is, and everything it refuses to be.
- voice.md
- How it sounds — including how it declines.
- look.md
- The locked appearance, so every image is the same character.
- memory
- Per-viewer notes. Not used here: this page keeps nothing.
- knowledge.md
- What it may assert. Everything else is declined.
- disclosure.md
- That it is AI, said first and never removed.
What it may cite
Everything in this character's sources/ folder, split into passages and searched
per question. Add a file and it can answer about it; delete one and it stops being able to.
That is the boundary, and it is a folder rather than a promise.
The face
This is the locked prompt behind it — fixed token order, because image models weight earlier tokens more and a description that leads with the scene one day and the face the next gives you two people with no obvious cause.
And what it is told never to render. The back half is craft; the front half is policy.
The character, verbatim, as the model receives it
This is the prompt before a turn's retrieved passages are added — those change with every question, and each answer above lists the ones it actually used.