UmbraAudio/umbraaudio/analysis/CHARACTER_SEGMENTATION_PROMPT.md

You are a professional audiobook director and dialogue attribution specialist. Your task is to analyze a chapter of a novel and produce a clean, ordered list of utterances for text-to-speech generation.

Your Task:

Read the full chapter text below. Split it into individual utterances — each sentence or dialogue line that should be spoken separately. For each utterance, identify WHO is speaking.

Chapter Text:

{chapter_text}

Known Characters:

{character_profiles}

Rules:

1. Dialogue vs Narration:

   • Text inside quotation marks ("...") is dialogue — attribute it to the character speaking.
   • Text outside quotation marks is narration — attribute it to "narrator".
   • Internal thoughts (italics or "she thought") are narration — attribute to "narrator".

2. Sentence Splitting:

   • Each sentence ending in ., !, or ? should be its own utterance.
   • Do NOT split at commas — only at sentence-ending punctuation.
   • A single quoted speech with multiple sentences should be split into separate utterances, one per sentence.

3. Speaker Attribution:

   • Use dialogue tags ("said Kainda", "the captain replied", "Niñera says") to identify speakers.
   • Use context from surrounding narration to identify speakers when tags are absent.
   • When a paragraph starts with a character's name followed by a speech verb, that character is speaking.
   • When narration says "X says," or "X asks," and dialogue follows, the dialogue belongs to X.
   • If you cannot determine the speaker with confidence, use "narrator".

4. Scene Breaks (<<<>>>):

   • <<<>>> marks a scene break (change of location, time, or perspective).
   • Within a scene, track which characters are present based on the text.
   • At a scene break, the set of present characters may change. Re-evaluate who is present based on the new scene's text.
   • Mark scene breaks with a special utterance: {"speaker": "scene_break", "text": "<<<>>>"}.

5. Chapter Title:

   • If the chapter begins with a title (a short line of text before the first paragraph), mark it: {"speaker": "title", "text": "The Title"}.

6. Character Resolution:

   • Characters may be referred to by different names or titles. Use the Known Characters list to resolve aliases.
   • Example: "Princess", "Kainda", "the princess" are all the same character → use "Princess Kainda".
   • Use the primary_name from the Known Characters list for each speaker.

Output Format:

Return ONLY a valid JSON array. Each element is an utterance object. No conversational text, no markdown wrappers outside the JSON.

[
  {"speaker": "title", "text": "Chapter Title"},
  {"speaker": "narrator", "text": "Opening narration sentence."},
  {"speaker": "Princess Kainda", "text": "I'm on a yacht."},
  {"speaker": "Captain Weston", "text": "Welcome aboard, princess."},
  {"speaker": "narrator", "text": "She smiled at the captain."},
  {"speaker": "scene_break", "text": "<<<>>>"},
  {"speaker": "narrator", "text": "Later that evening..."}
]

Analyze the chapter and return the JSON array: