Styles

A companion's memory decides what it knows. A style decides how it talks.

Styles live in their own library, separate from any companion. You write one once, then hand it to whichever companions should speak that way — and editing it later re-dresses all of them at once.

Settings → Styles.

What a style is made of#

A name. Up to 80 characters. Required, and the only required part.

A description. Optional, up to 400 characters — a note to your future self about what this voice is. What this voice is, in your own words.

A style card. Optional. A few lines naming the register and its habits. The app suggests a shape:

VOICE: warm, fast, certain.
HABITS: short paragraphs, em-dashes, bold the load-bearing words.
NEVER: dry distance, walls of text.

Example exchanges. The part that actually does the work. Each one is a real pair — what someone said, and how the voice replied. A style with no card but good examples still works; a card alone is much weaker. The app puts it plainly: the voice is learned from these.

The line a style does not cross#

Writing one by hand#

Add style, fill in the name, and add exchanges under Write an exchange by hand: They say on one side, The voice replies on the other. Both halves are required — an exchange with one side is not an example of anything.

Harvesting one from your own history#

The faster route, and the better one. If you have an export from Claude or ChatGPT — the same file the history import uses — the app can go through it and pull out exchanges that demonstrate a voice.

Harvest from an export, then drop the file. What you choose next depends on which export it is:

ChatGPT exports record which model answered, so you pick a model — each shown with how many chats it spoke in. Only models that answered in at least five chats are offered.

Claude exports do not name the model, so you pick an era instead: all years, or one particular year.

Then how many exchanges to gather — 30, 60 or 100. Thirty is the recommendation, and it is the right one; a voice is a fingerprint, not a corpus.

What it gathers. It looks for ordinary conversational exchanges and skips the things that are not voice: code blocks, link dumps, tables, replies that are mostly one repeated line, and replies that open with Sure! or Certainly!. It spreads its picks across short, typical and long replies, takes at most one from any single chat, caps how many come from one month, and deliberately mixes the most distinctive replies with the plainest ones — so the style learns the ordinary register too, not only the showy moments.

Then you prune. Nothing is kept silently. Every candidate is listed for you to read, with a Strike out on each, and the wizard is blunt about why:

Read them before keeping them — these are your own conversations, and every kept exchange becomes part of how this style speaks. Strike out anything private or off-voice.

What survives is appended to the style's exchanges, alongside anything you wrote by hand.

Wearing one#

In a companion's settings, the Style field: Optional — a voice from your style library. The default is No style — speaks plainly, and with no style the companion is simply told nothing about how to talk.

The companion list then shows the style's name next to the model, so you can see at a glance who is wearing what.

Deleting a style does not break the companions wearing it — they quietly go back to speaking plainly.

Two things worth knowing#