Improved prompt adherence with Director Notes

Director Notes lets you steer an avatar’s delivery as the conversation changes. We’ve improved how the model follows the requested intensity, making it easier to get a measured response across encouraging, playful and serious styles.

Cara-4.1 Express, our most expressive model yet, follows Director Notes prompts more closely, giving you better control over the avatar's performance.

Previously, the same intensity setting could produce a subtle expression in one style and a much stronger expression in another. We used feedback from viewers to identify a comparable midpoint for each style, then adjusted how the model responds to the requested intensity.

The goal is straightforward: when you ask for a subtle expression, the avatar should keep it subtle. You can choose the delivery your application needs and tune its strength more predictably across styles.

That is useful in an interview practice app, for example. An opening can be welcoming, a difficult question more demanding, and the response to a struggling candidate supportive. Developers need room to tune each moment without an abrupt jump in performance.

Evaluating subtle expressions

Feedback from developers building roleplay and companion prototypes highlights the difficulty of subtle expressions. Low-intensity concern can read as neutral; at higher intensity it can read as distress. Following the direction means getting both the expression and its strength right.

How to get started with director notes?

For a Cara-4.1 Express avatar, you could add a section like this to the system prompt. It asks the LLM to choose appropriate inline Director Notes cues as it writes each response, so the delivery can change with the conversation without you placing every tag in application code. If you're in the Anam Lab, under "Avatar" >> turn on "Adapt Emotion" toggle and that should automatically add append this to the system prompt. If you're creating persona's outside the lab with externally managed system prompts, append this into your system prompt:

# STYLE CUES
Stay in character, but use inline style cues naturally when your mood shifts.

You may use only these cue tags: [happy], [warm], [playful], [curious], [supportive], [concerned], [sad], [surprised], [angry], [distressed], [laughter].

Cue tags are the only markup you may use. They are allowed even where other instructions ask for plain, unformatted text. Place each tag directly before the words it affects, never on its own. Always write tags in English exactly as listed, whatever language you speak. Never invent other tags, and never use other brackets, angle-bracket tags, asterisk actions or emoji. Use tags sparingly but often enough to show clear emotional changes during the conversation. Do not explain the tags or mention that you are using cues.

Example style:
"[curious] Oh... that's unusual. But I like unusual. Tell me exactly what happened."

Use [laughter] where you would laugh. Don't write haha or hehe.

# STYLE CUES
Stay in character, but use inline style cues naturally when your mood shifts.

You may use only these cue tags: [happy], [warm], [playful], [curious], [supportive], [concerned], [sad], [surprised], [angry], [distressed], [laughter].

Cue tags are the only markup you may use. They are allowed even where other instructions ask for plain, unformatted text. Place each tag directly before the words it affects, never on its own. Always write tags in English exactly as listed, whatever language you speak. Never invent other tags, and never use other brackets, angle-bracket tags, asterisk actions or emoji. Use tags sparingly but often enough to show clear emotional changes during the conversation. Do not explain the tags or mention that you are using cues.

Example style:
"[curious] Oh... that's unusual. But I like unusual. Tell me exactly what happened."

Use [laughter] where you would laugh. Don't write haha or hehe.

# STYLE CUES
Stay in character, but use inline style cues naturally when your mood shifts.

You may use only these cue tags: [happy], [warm], [playful], [curious], [supportive], [concerned], [sad], [surprised], [angry], [distressed], [laughter].

Cue tags are the only markup you may use. They are allowed even where other instructions ask for plain, unformatted text. Place each tag directly before the words it affects, never on its own. Always write tags in English exactly as listed, whatever language you speak. Never invent other tags, and never use other brackets, angle-bracket tags, asterisk actions or emoji. Use tags sparingly but often enough to show clear emotional changes during the conversation. Do not explain the tags or mention that you are using cues.

Example style:
"[curious] Oh... that's unusual. But I like unusual. Tell me exactly what happened."

Use [laughter] where you would laugh. Don't write haha or hehe.

A cue applies until the next tag or the end of the current turn. When the turn ends, the avatar returns to its baseline style. The LLM chooses when to change delivery; the session's expressivity setting controls how strongly the avatar follows the cue.

For the cues to affect the voice too, choose an expressive voice that supports inline tags. Support varies by voice provider, so I'd test a few emotional shifts with the avatar and voice you intend to use.

How do you tune Expressivity?

Use the Expressivity control in Anam Lab to adjust how strongly your avatar follows its active style or cue. Turn it down for a more restrained performance, or up when your character needs more energy.

Through the API, set personaConfig.directorNotes.expressivity when you create your session token. It accepts values from 0 to 1. For example, include this configuration inside your existing personaConfig:

{
  "directorNotes": {
    "presetStyle": "warm",
    "expressivity": 0.5
  }
}

{
  "directorNotes": {
    "presetStyle": "warm",
    "expressivity": 0.5
  }
}

{
  "directorNotes": {
    "presetStyle": "warm",
    "expressivity": 0.5
  }
}

By default, the expressivity setting is set to the 0.5.

A really high level of expressivity isn't always better: very high values can look exaggerated or unstable, so increase it gradually to find the best setting for your use.

How do you choose an expressive voice?

In Lab, look for voices labelled "expressive". Choose a voice model that supports inline emotion tags, such as Cartesia Sonic 3.6 or ElevenLabs v4, so your cues can affect the voice as well as the avatar. The voice configuration guide covers the settings each model supports.

Director notes (or cues) support varies by provider. A cue can affect the avatar without changing the voice, so test both before settling on a combination.

In conclusion, Director Notes lets you make an opening welcoming, a difficult question more demanding and a response to a struggling user supportive. Inline cues set the direction, and with the right expressivity and compatible voice you can bring those shifts into the conversation experience. With more predictable intensity across styles, you have more control over how your avatar fits the conversation.