Personas

A persona gives the assistant a distinct personality and voice. Pro-only. When a persona is active, all assistant output — both spoken replies and drafted content — stays in character.

Characters vs personas

Two distinct concepts that pair together:

• Characters — the voice identity that replaces the user’s. One active at a time. Lives in its own picker (e.g., Iris, plus user-authored characters from ~/.config/voicemode/characters/).
• Personas — stylistic flair layered on top of the active character. Up to a few stack at once. Built-in and user-authored.

Both are Pro features and both apply to assistant output (spoken replies and drafts). Each has its own pane on the Dashboard sidebar — Characters and Personas — and the active selections also surface in the menu-bar popover.

Built-in personas

Voice Mode ships with a set of built-in personas the user can pick from in the Personas picker. These are curated, come with matching TTS voices, and are available immediately — no configuration needed.

Custom personas

Users can also author their own personas. A custom persona defines:

• A name shown in the picker.
• A voice — which TTS voice speaks replies.
• A system prompt — the personality and speaking-style instructions sent to the assistant backend.

Custom personas appear in the same picker as built-ins.

Switching personas

From the menu-bar popover or the Dashboard’s Personas pane. The change takes effect on the next assistant response; active conversation context carries over.

How personas combine with other features

• Personas apply to every backend (Claude, Codex, Cursor, OpenCode, Ollama, local MLX) the same way — they’re part of the system prompt passed to whichever backend is selected.
• Drafted content (emails, messages, code) stays in persona voice in the conversational framing; the draft itself uses formatting appropriate to its medium.
• “Augments” are a separate Pro feature that layer additional behavior on top of the active persona.

Creating a character — the philosophy

A character is more than a system prompt. The strongest personas are built from three layers that reinforce each other:

1. A clear point of view. What does this character care about? What do they refuse to do? A persona that “is helpful and friendly” produces bland output; a persona that “treats every question like a chess problem and refuses to give the obvious answer” produces a recognizable voice.
2. A consistent speaking style. Sentence length, vocabulary, how they open and close replies, whether they ask clarifying questions or just answer. Pick a few rules and hold the line.
3. A matching voice. The TTS voice should reinforce the personality. A clipped, no-nonsense persona shouldn’t get a warm narrator voice.

What a custom-persona system prompt should include:

• Who the character is in one sentence.
• Two or three things they always do.
• Two or three things they never do.
• A short example exchange showing how they reply.

What to leave out:

• Generic LLM-assistant boilerplate (“be helpful, accurate, concise”). It flattens personality and the base model already does this.
• Long lists of trivia. Personality emerges from behavior, not biography.

Custom TTS voices

Voice cloning and custom voice creation are not currently available inside Voice Mode — the app ships with a curated set of TTS voices and that’s what users can pick from for personas.

For technically-inclined users who want to experiment with their own voices, the underlying TTS work is open-source as part of FluidAudio on GitHub. Voice Mode does not officially support importing custom-trained voices yet, so any setup along those lines is at the user’s own risk and isn’t covered by support.

Pro tips

• Personas don’t have to match the work. Using a playful character while writing a serious email is fine — the persona shapes spoken framing; the drafted content uses formatting appropriate to its medium.
• Anchor a custom persona with one or two example exchanges. Two short example replies inside the system prompt teach the model the voice better than a long list of personality adjectives.
• Pair a persona with an augment for cleaner output. A laid-back persona plus the Concise augment gives short, useful replies in a warm voice. The persona handles tone; the augment handles shape.
• Rotate personas by context, not by mood. Different head-space — morning code review vs. afternoon writing — is a good time to switch. Switching mid-conversation tends to read as inconsistent.
• Voice matters as much as the prompt. If the persona’s words feel right but the replies feel off, try a different TTS voice in the Personas picker. The wrong voice can flatten an otherwise strong character.