Recipe: Response Guidance

This recipe detects when the bot is drifting from a character's defined personality or voice, then nudges it back via a Stage Direction — without the user seeing the nudge.

It also demonstrates dynamic response guidance: reading what the user is doing and giving the bot matching writing instructions.

The config

{
  "variables": [
    {
      "name": "userAction",
      "initialValue": "\"talking\""
    },
    {
      "name": "driftDetected",
      "initialValue": "false"
    }
  ],
  "classifiers": [
    {
      "name": "UserIntent",
      "inputTemplate": "{{content}}",
      "inputHypothesis": "The user is primarily {}.",
      "classifications": [
        {
          "label": "examining or observing something closely",
          "category": "intent",
          "threshold": 0.6,
          "updates": [{ "variable": "userAction", "setTo": "\"examining\"" }]
        },
        {
          "label": "trying to end the scene or move on",
          "category": "intent",
          "threshold": 0.6,
          "updates": [{ "variable": "userAction", "setTo": "\"transitioning\"" }]
        },
        {
          "label": "taking a direct action or attacking",
          "category": "intent",
          "threshold": 0.6,
          "updates": [{ "variable": "userAction", "setTo": "\"acting\"" }]
        },
        {
          "label": "talking or asking a question",
          "category": "intent",
          "threshold": 0.5,
          "updates": [{ "variable": "userAction", "setTo": "\"talking\"" }]
        }
      ]
    },
    {
      "name": "CharacterDrift",
      "responseTemplate": "{{content}}",
      "responseHypothesis": "This response sounds {}.",
      "classifications": [
        {
          "label": "out of character or too formal",
          "threshold": 0.7,
          "updates": [{ "variable": "driftDetected", "setTo": "true" }]
        }
      ]
    }
  ],
  "content": [
    {
      "category": "Stage Direction",
      "condition": "userAction == \"examining\"",
      "modification": "\"The user is examining something. Write a detailed, sensory description. Focus on what {{char}} notices.\""
    },
    {
      "category": "Stage Direction",
      "condition": "userAction == \"transitioning\"",
      "modification": "\"The user is trying to move on. Wrap up the current moment naturally and introduce the next scene.\""
    },
    {
      "category": "Stage Direction",
      "condition": "userAction == \"acting\"",
      "modification": "\"The user is taking action. Write a dynamic, punchy response. React to the action directly.\""
    },
    {
      "category": "Stage Direction",
      "condition": "driftDetected",
      "modification": "\"Important: stay in character as {{char}}. Speak in their established voice and manner.\""
    }
  ]
}

How it works

1. UserIntent reads the user's message and classifies what they are trying to do. It updates userAction to one of four values.

2. Four Stage Direction rules each fire for a different userAction value, giving the bot a specific writing instruction tailored to what the user is doing.

3. CharacterDrift reads the bot's response (note responseTemplate) and flags whether it sounds out of character. If so, driftDetected is set to true. (Statosphere auto-creates variables that classifiers write to, but declaring driftDetected explicitly in variables as shown above is good practice — it makes the config easier to read.)

4. A final Stage Direction fires when driftDetected is true, reminding the bot to stay in character. This fires on the next turn after drift was detected — Statosphere cannot modify the response that was already generated.

Limitations

The drift detection fires after the response is already shown. It cannot retroactively change the bad reply; it prevents the next one from drifting. For more immediate correction, consider using useLlm: true on the classifier, which may detect tone issues more reliably.

Customizing the character voice reminder

Replace the generic drift reminder with something specific to your character:

{
  "category": "Stage Direction",
  "condition": "driftDetected",
  "modification": "\"Remember: {{char}} speaks in short, clipped sentences. They never apologize and they never explain themselves. Get back on track.\""
}