Shape your summaries with templates

A template decides which sections a meeting's summary has and what goes in each of them. Five ship with the app and you can write your own. This page is also what an assistant is pointed at before it writes one for you.

The two halves of a template

A template has two pieces of text, and they do different jobs.

When this type applies is read by the model that decides which template a meeting gets. The editor says it plainly: Read by the model when it detects a meeting's type. Describe the meeting, not the document. Write what the conversation sounds like, not what you want the write-up to look like.

Sections is the shape of the document itself: The sections this summary should have, and what to look for in each. Markdown headings.

Write sections a small model can follow

The model that writes your summary may be the one that ships inside the app, and it runs on your Mac rather than in a data centre. These rules are what the five built-in templates already do, and they matter most for that model.

One ## heading per section, and a short line under it saying what belongs there. The heading is the instruction; the line under it is the scope.
Put a placeholder in brackets rather than an example: [What was discussed], not a sample sentence.
Never write a person's name, a company or a date as an example. A small model copies it straight into your summary as a participant who was never in the meeting. None of the built-in templates contains one, and that is why.
Say "or none" where a section can legitimately be empty, as in [Feedback given, or "none"]. Without it a model invents something to fill the heading.
Ask for a table only where the content is genuinely tabular, such as tasks with an owner and a date.

Keep it short

Every section you add is text sent to the model on top of the transcript, and the editor shows the total while you type. Its warning is the whole reason to keep a template tight: A longer prompt leaves a local model less room for the transcript.

For scale, the five built-in templates run between 444 and 628 characters. An assistant writing a template through a connector is refused above 4,000 characters; you are not capped in the editor, but the same arithmetic applies to what you write by hand.

The two switches

Each template has an enabled state and Auto-detect, and they are independent. Enabled decides whether you can pick the template for a meeting. Auto-detect decides whether the model may choose it for you. Turning Auto-detect off for a type you rarely record keeps it available by hand without letting the detector reach for it.

Built-in templates are copied, not edited

A template marked Built-in has its wording refreshed by the app whenever it improves, so an edit to one would be reverted. Use Duplicate to Edit and change the copy. Its two switches are yours either way, and they apply to every workspace, because one built-in row is shared across all of them. A template you write belongs to the workspace you wrote it in.

The General template cannot be switched off. Every meeting the detector cannot place falls back to it, so something has to answer.

What an assistant may change

An assistant you have connected can read your templates and, if you have separately allowed changes, add and edit them. Three rules bound it, and they are the same ones above:

It cannot delete a template. That stays with you.
It cannot edit a built-in's wording, only its two switches, and it must copy the built-in to change its text.
A change it makes is refused if you edited the same template since it read it, rather than overwriting you.

Every change an assistant makes raises a notice with an Undo.

Result

Settings for a template live under Templates in the main window, and the template a meeting used is shown on the meeting itself. A new or edited template changes how future summaries are written and leaves existing ones alone.