Personas: Different Assistants for Different Jobs

A persona is a named system prompt with its own temperature and its own list of tools the model may use. Switching persona switches what kind of assistant you are talking to, per conversation, without re-explaining yourself every time.

Why this matters more on a small model

Large cloud models infer tone from context. A 2B model needs telling. "Answer in at most three sentences, no preamble, no apology" is the difference between a useful local assistant and one that spends forty tokens agreeing that your question is a great one — and on a phone, forty wasted tokens is a wasted second and a warmer battery.

What you can put in one

• The prompt — who it is and how it should answer.
• Temperature — low for extraction and code, higher for brainstorming.
• A tool allowlist — a journalling persona has no business calling web search, and a persona that cannot reach the network is a persona whose conversations definitionally stay on the phone.

Some that make sense

A terse coding assistant at low temperature with the calculator enabled and nothing else. A patient tutor that explains rather than answers. A journalling companion with every network tool switched off and memory on. A translator. An editor that only ever rewrites what you paste and never comments on it.

Personas ship with the app, they are all editable, and you can write your own. If you want an assistant that also plans and delegates rather than just answering, that is what custom agents are for.