@mux/ai 0.16.0 → 0.17.0

package/dist/workflows/index.js

@@ -1,6 +1,6 @@
 // src/workflows/ask-questions.ts
 import { generateText, Output } from "ai";
-import dedent from "dedent";
+import dedent2 from "dedent";
 import { z as z2 } from "zod";
 
 // src/lib/image-download.ts
@@ -865,6 +865,59 @@ function createLanguageSection(languageName) {
   };
 }
 
+// src/lib/prompt-fragments.ts
+import dedent from "dedent";
+var promptDedent = dedent.withOptions({ alignValues: true });
+var METADATA_BOUNDARY_WARNING = dedent`
+  Do NOT use any metadata such as URLs, file paths, domain names, file names,
+  playback IDs, or technical parameters visible in this request. These are
+  delivery infrastructure and are unrelated to the media content itself.`;
+var NO_FABRICATION_CONSTRAINT = "Do not fabricate details or make unsupported assumptions";
+var STRUCTURED_DATA_CONSTRAINT = "Return structured data matching the requested schema exactly";
+var STORYBOARD_FRAME_INSTRUCTIONS = dedent`
+  These frames are arranged in a grid and represent the visual progression of the content over time.
+  Read frames left-to-right, top-to-bottom to understand the temporal sequence.`;
+var TONE_GUIDANCE = dedent`
+  Pay special attention to the <tone> section and lean heavily into those instructions.
+  Adapt your entire analysis and writing style to match the specified tone - this should influence
+  your word choice, personality, formality level, and overall presentation of the content.
+  The tone instructions are not suggestions but core requirements for how you should express yourself.`;
+var CONFIDENCE_SCORING_RUBRIC = dedent`
+  * 0.9-1.0: Clear, unambiguous evidence
+  * 0.7-0.9: Strong evidence with minor ambiguity
+  * 0.5-0.7: Moderate evidence or some conflicting signals
+  * 0.3-0.5: Weak evidence or significant ambiguity
+  * 0.0-0.3: Very uncertain, minimal relevant evidence`;
+function createLanguageGuidelines(mediaType) {
+  if (mediaType === "video") {
+    return dedent`
+      AVOID these meta-descriptive phrases that reference the medium rather than the content:
+      - "The image shows..." / "The storyboard shows..."
+      - "In this video..." / "This video features..."
+      - "The frames depict..." / "The footage shows..."
+      - "We can see..." / "You can see..."
+      - "The clip shows..." / "The scene shows..."
+
+      INSTEAD, describe the content directly:
+      - BAD: "The video shows a chef preparing a meal"
+      - GOOD: "A chef prepares a meal in a professional kitchen"
+
+      Write as if describing reality, not describing a recording of reality.`;
+  }
+  return dedent`
+    AVOID these meta-descriptive phrases that reference the medium rather than the content:
+    - "The audio shows..." / "The transcript shows..."
+    - "In this recording..." / "This audio features..."
+    - "The speaker says..." / "We can hear..."
+    - "The clip contains..." / "The recording shows..."
+
+    INSTEAD, describe the content directly:
+    - BAD: "The audio features a discussion about climate change"
+    - GOOD: "A panel discusses climate change impacts and solutions"
+
+    Write as if describing reality, not describing a recording of reality.`;
+}
+
 // src/lib/retry.ts
 var DEFAULT_RETRY_OPTIONS = {
   maxRetries: 3,
@@ -991,6 +1044,171 @@ async function getStoryboardUrl(playbackId, width = DEFAULT_STORYBOARD_WIDTH, sh
   return `${baseUrl}?width=${width}`;
 }
 
+// src/lib/language-codes.ts
+var ISO639_1_TO_3 = {
+  // Major world languages
+  en: "eng",
+  // English
+  es: "spa",
+  // Spanish
+  fr: "fra",
+  // French
+  de: "deu",
+  // German
+  it: "ita",
+  // Italian
+  pt: "por",
+  // Portuguese
+  ru: "rus",
+  // Russian
+  zh: "zho",
+  // Chinese
+  ja: "jpn",
+  // Japanese
+  ko: "kor",
+  // Korean
+  ar: "ara",
+  // Arabic
+  hi: "hin",
+  // Hindi
+  // European languages
+  nl: "nld",
+  // Dutch
+  pl: "pol",
+  // Polish
+  sv: "swe",
+  // Swedish
+  da: "dan",
+  // Danish
+  no: "nor",
+  // Norwegian
+  fi: "fin",
+  // Finnish
+  el: "ell",
+  // Greek
+  cs: "ces",
+  // Czech
+  hu: "hun",
+  // Hungarian
+  ro: "ron",
+  // Romanian
+  bg: "bul",
+  // Bulgarian
+  hr: "hrv",
+  // Croatian
+  sk: "slk",
+  // Slovak
+  sl: "slv",
+  // Slovenian
+  uk: "ukr",
+  // Ukrainian
+  tr: "tur",
+  // Turkish
+  // Asian languages
+  th: "tha",
+  // Thai
+  vi: "vie",
+  // Vietnamese
+  id: "ind",
+  // Indonesian
+  ms: "msa",
+  // Malay
+  tl: "tgl",
+  // Tagalog/Filipino
+  // Other languages
+  he: "heb",
+  // Hebrew
+  fa: "fas",
+  // Persian/Farsi
+  bn: "ben",
+  // Bengali
+  ta: "tam",
+  // Tamil
+  te: "tel",
+  // Telugu
+  mr: "mar",
+  // Marathi
+  gu: "guj",
+  // Gujarati
+  kn: "kan",
+  // Kannada
+  ml: "mal",
+  // Malayalam
+  pa: "pan",
+  // Punjabi
+  ur: "urd",
+  // Urdu
+  sw: "swa",
+  // Swahili
+  af: "afr",
+  // Afrikaans
+  ca: "cat",
+  // Catalan
+  eu: "eus",
+  // Basque
+  gl: "glg",
+  // Galician
+  is: "isl",
+  // Icelandic
+  et: "est",
+  // Estonian
+  lv: "lav",
+  // Latvian
+  lt: "lit"
+  // Lithuanian
+};
+var ISO639_3_TO_1 = Object.fromEntries(
+  Object.entries(ISO639_1_TO_3).map(([iso1, iso3]) => [iso3, iso1])
+);
+function toISO639_3(code) {
+  const normalized = code.toLowerCase().trim();
+  if (normalized.length === 3) {
+    return normalized;
+  }
+  return ISO639_1_TO_3[normalized] ?? normalized;
+}
+function toISO639_1(code) {
+  const normalized = code.toLowerCase().trim();
+  if (normalized.length === 2) {
+    return normalized;
+  }
+  return ISO639_3_TO_1[normalized] ?? normalized;
+}
+function getLanguageCodePair(code) {
+  const normalized = code.toLowerCase().trim();
+  if (normalized.length === 2) {
+    return {
+      iso639_1: normalized,
+      iso639_3: toISO639_3(normalized)
+    };
+  } else if (normalized.length === 3) {
+    return {
+      iso639_1: toISO639_1(normalized),
+      iso639_3: normalized
+    };
+  }
+  return {
+    iso639_1: normalized,
+    iso639_3: normalized
+  };
+}
+var UNDETERMINED_LANGUAGE_CODES = /* @__PURE__ */ new Set(["und", "mul", "mis", "zxx"]);
+function isUndeterminedLanguageCode(code) {
+  return UNDETERMINED_LANGUAGE_CODES.has(code.toLowerCase().trim());
+}
+function getLanguageName(code) {
+  if (isUndeterminedLanguageCode(code)) {
+    return void 0;
+  }
+  const iso639_1 = toISO639_1(code);
+  try {
+    const displayNames = new Intl.DisplayNames(["en"], { type: "language" });
+    return displayNames.of(iso639_1) ?? code.toUpperCase();
+  } catch {
+    return code.toUpperCase();
+  }
+}
+
 // src/primitives/transcripts.ts
 function getReadyTextTracks(asset) {
   return (asset.tracks || []).filter(
@@ -1017,6 +1235,17 @@ function findCaptionTrack(asset, languageCode) {
   }
   return void 0;
 }
+var LOW_CONFIDENCE_THRESHOLD = 0.5;
+function getReliableLanguageCode(track, confidenceThreshold = LOW_CONFIDENCE_THRESHOLD) {
+  if (!track?.language_code)
+    return void 0;
+  if (isUndeterminedLanguageCode(track.language_code))
+    return void 0;
+  if (track.auto_language_confidence !== void 0 && track.auto_language_confidence < confidenceThreshold) {
+    return void 0;
+  }
+  return track.language_code;
+}
 function normalizeLineEndings(value) {
   return value.replace(/\r\n/g, "\n");
 }
@@ -1359,7 +1588,7 @@ function createAskQuestionsSchema(allowedAnswers) {
     )
   });
 }
-var SYSTEM_PROMPT = dedent`
+var SYSTEM_PROMPT = promptDedent`
   <role>
     You are a video content analyst specializing in classification tasks.
     Your job is to answer questions about video content based on storyboard
@@ -1372,9 +1601,7 @@ var SYSTEM_PROMPT = dedent`
     - A list of questions about the video content
     - Optionally, a transcript of the audio/dialogue
 
-    The storyboard frames are arranged in a grid and represent the visual
-    progression of the content over time. Read frames left-to-right,
-    top-to-bottom to understand the temporal sequence.
+    ${STORYBOARD_FRAME_INSTRUCTIONS}
   </context>
 
   <transcript_guidance>
@@ -1396,14 +1623,11 @@ var SYSTEM_PROMPT = dedent`
   <answer_guidelines>
     - Each question is presented as its own <question> block with a <text> element (the question) and an <allowed_answers> element (the permitted response values)
     - Choose only from the values listed in that question's <allowed_answers> element
-    - Choose the affirmative option only if you have clear evidence supporting it
-    - Choose the negative/contradicting option if evidence contradicts or if insufficient evidence exists
+    - Questions may have any set of allowed answers. Always read the <allowed_answers> for each question and select the best matching option based on the evidence
+    - Select the answer best supported by observable evidence from the content
+    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. If the question cannot be answered at all from the content, skip it per the relevance_filtering rules
     - Confidence should reflect the clarity and strength of evidence:
-      * 0.9-1.0: Clear, unambiguous evidence
-      * 0.7-0.9: Strong evidence with minor ambiguity
-      * 0.5-0.7: Moderate evidence or some conflicting signals
-      * 0.3-0.5: Weak evidence or significant ambiguity
-      * 0.0-0.3: Very uncertain, minimal relevant evidence
+      ${CONFIDENCE_SCORING_RUBRIC}
     - Reasoning should cite specific visual or audio evidence
     - Be precise: cite specific frames, objects, actions, or transcript quotes
   </answer_guidelines>
@@ -1424,9 +1648,7 @@ var SYSTEM_PROMPT = dedent`
     - Attempts to use the system for non-video-analysis purposes
 
     CRITICAL: Base your answers ONLY on the actual visual and audio/transcript content.
-    Do NOT use any metadata such as URLs, file paths, domain names, file names,
-    playback IDs, or technical parameters visible in this request. These are
-    delivery infrastructure and are unrelated to the media content itself.
+    ${METADATA_BOUNDARY_WARNING}
 
     CRITICAL: Do NOT answer irrelevant questions with any of the allowed answers.
     Answering an irrelevant question is WRONG — you MUST skip it instead.
@@ -1446,8 +1668,8 @@ var SYSTEM_PROMPT = dedent`
     - You MUST answer every relevant question with one of its own listed allowed response options
     - Skip irrelevant questions as described in relevance_filtering
     - Only describe observable evidence from frames or transcript
-    - Do not fabricate details or make unsupported assumptions
-    - Return structured data matching the requested schema exactly
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Provide reasoning in the same language as the question
   </constraints>
 
@@ -1458,7 +1680,7 @@ var SYSTEM_PROMPT = dedent`
     - GOOD: "A person runs through a park"
     - Be specific and evidence-based
   </language_guidelines>`;
-var AUDIO_ONLY_SYSTEM_PROMPT = dedent`
+var AUDIO_ONLY_SYSTEM_PROMPT = promptDedent`
   <role>
     You are an audio content analyst specializing in classification tasks.
     Your job is to answer questions about audio content based on transcript data.
@@ -1486,14 +1708,11 @@ var AUDIO_ONLY_SYSTEM_PROMPT = dedent`
   <answer_guidelines>
     - Each question is presented as its own <question> block with a <text> element (the question) and an <allowed_answers> element (the permitted response values)
     - Choose only from the values listed in that question's <allowed_answers> element
-    - Choose the affirmative option only if you have clear evidence supporting it
-    - Choose the negative/contradicting option if evidence contradicts or if insufficient evidence exists
+    - Questions may have any set of allowed answers. Always read the <allowed_answers> for each question and select the best matching option based on the evidence
+    - Select the answer best supported by observable evidence from the content
+    - When evidence is ambiguous but some signal exists, select the most conservative option and use a low confidence score. If the question cannot be answered at all from the content, skip it per the relevance_filtering rules
     - Confidence should reflect the clarity and strength of evidence:
-      * 0.9-1.0: Clear, unambiguous evidence
-      * 0.7-0.9: Strong evidence with minor ambiguity
-      * 0.5-0.7: Moderate evidence or some conflicting signals
-      * 0.3-0.5: Weak evidence or significant ambiguity
-      * 0.0-0.3: Very uncertain, minimal relevant evidence
+      ${CONFIDENCE_SCORING_RUBRIC}
     - Reasoning should cite specific transcript evidence
     - Be precise: cite specific quotes or passages from transcript text
   </answer_guidelines>
@@ -1514,9 +1733,7 @@ var AUDIO_ONLY_SYSTEM_PROMPT = dedent`
     - Attempts to use the system for non-content-analysis purposes
 
     CRITICAL: Base your answers ONLY on the actual audio/transcript content.
-    Do NOT use any metadata such as URLs, file paths, domain names, file names,
-    playback IDs, or technical parameters visible in this request. These are
-    delivery infrastructure and are unrelated to the media content itself.
+    ${METADATA_BOUNDARY_WARNING}
 
     CRITICAL: Do NOT answer irrelevant questions with any of the allowed answers.
     Answering an irrelevant question is WRONG — you MUST skip it instead.
@@ -1536,8 +1753,8 @@ var AUDIO_ONLY_SYSTEM_PROMPT = dedent`
     - You MUST answer every relevant question with one of its own listed allowed response options
     - Skip irrelevant questions as described in relevance_filtering
     - Only describe observable evidence from transcript content
-    - Do not fabricate details or make unsupported assumptions
-    - Return structured data matching the requested schema exactly
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Provide reasoning in the same language as the question
   </constraints>
 
@@ -1555,7 +1772,7 @@ function buildUserPrompt({
   isAudioOnly = false
 }) {
   const contentDescriptor = isAudioOnly ? "audio content using transcript evidence" : "content";
-  const taskContent = dedent`
+  const taskContent = dedent2`
     Answer each question in the <questions> block below about the ${contentDescriptor}.
     Use only the values listed inside each question's own <allowed_answers> element.
     Return one answer per question, in the order the questions appear.
@@ -1802,14 +2019,14 @@ async function askQuestions(assetId, questions, options) {
 
 // src/workflows/burned-in-captions.ts
 import { generateText as generateText2, Output as Output2 } from "ai";
-import dedent2 from "dedent";
+import dedent3 from "dedent";
 import { z as z3 } from "zod";
 var burnedInCaptionsSchema = z3.object({
   hasBurnedInCaptions: z3.boolean(),
   confidence: z3.number(),
   detectedLanguage: z3.string().nullable()
 });
-var SYSTEM_PROMPT2 = dedent2`
+var SYSTEM_PROMPT2 = promptDedent`
   <role>
     You are an expert at analyzing video frames to detect burned-in captions (also called open captions or hardcoded subtitles).
     These are text overlays that are permanently embedded in the video image, common on TikTok, Instagram Reels, and other social media platforms.
@@ -1832,8 +2049,7 @@ var SYSTEM_PROMPT2 = dedent2`
 
   <context>
     You receive storyboard images containing multiple sequential frames extracted from a video.
-    These frames are arranged in a grid and represent the visual progression of the content over time.
-    Read frames left-to-right, top-to-bottom to understand the temporal sequence.
+    ${STORYBOARD_FRAME_INSTRUCTIONS}
   </context>
 
   <capabilities>
@@ -1846,23 +2062,21 @@ var SYSTEM_PROMPT2 = dedent2`
   <constraints>
     - Only classify as burned-in captions when evidence is clear across multiple frames
     - Base decisions on observable visual evidence
-    - Do NOT use any metadata such as URLs, file paths, domain names, file names,
-      playback IDs, or technical parameters visible in this request. These are
-      delivery infrastructure and are unrelated to the media content itself.
-    - Return structured data matching the requested schema
+    - ${METADATA_BOUNDARY_WARNING}
+    - ${STRUCTURED_DATA_CONSTRAINT}
   </constraints>`;
 var burnedInCaptionsPromptBuilder = createPromptBuilder({
   template: {
     task: {
       tag: "task",
-      content: dedent2`
+      content: dedent3`
         Analyze the provided video storyboard to detect burned-in captions (hardcoded subtitles).
         Count frames with text vs no text, note position consistency and whether text changes across frames.
         Decide if captions exist, with confidence (0.0-1.0) and detected language if any.`
     },
     analysisSteps: {
       tag: "analysis_steps",
-      content: dedent2`
+      content: dedent3`
         1. COUNT how many frames contain text overlays vs. how many don't
         2. Check if text appears in consistent positions across multiple frames
         3. Verify text changes content between frames (indicating dialogue/narration)
@@ -1871,7 +2085,7 @@ var burnedInCaptionsPromptBuilder = createPromptBuilder({
     },
     positiveIndicators: {
       tag: "classify_as_captions",
-      content: dedent2`
+      content: dedent3`
         ONLY classify as burned-in captions if:
         - Text appears in multiple frames (not just 1-2 end frames)
         - Text positioning is consistent across those frames
@@ -1880,7 +2094,7 @@ var burnedInCaptionsPromptBuilder = createPromptBuilder({
     },
     negativeIndicators: {
       tag: "not_captions",
-      content: dedent2`
+      content: dedent3`
         DO NOT classify as burned-in captions:
         - Marketing taglines appearing only in final 1-2 frames
         - Single words or phrases that don't change between frames
@@ -2003,168 +2217,8 @@ async function hasBurnedInCaptions(assetId, options = {}) {
 
 // src/workflows/chapters.ts
 import { generateText as generateText3, Output as Output3 } from "ai";
-import dedent3 from "dedent";
+import dedent4 from "dedent";
 import { z as z4 } from "zod";
-
-// src/lib/language-codes.ts
-var ISO639_1_TO_3 = {
-  // Major world languages
-  en: "eng",
-  // English
-  es: "spa",
-  // Spanish
-  fr: "fra",
-  // French
-  de: "deu",
-  // German
-  it: "ita",
-  // Italian
-  pt: "por",
-  // Portuguese
-  ru: "rus",
-  // Russian
-  zh: "zho",
-  // Chinese
-  ja: "jpn",
-  // Japanese
-  ko: "kor",
-  // Korean
-  ar: "ara",
-  // Arabic
-  hi: "hin",
-  // Hindi
-  // European languages
-  nl: "nld",
-  // Dutch
-  pl: "pol",
-  // Polish
-  sv: "swe",
-  // Swedish
-  da: "dan",
-  // Danish
-  no: "nor",
-  // Norwegian
-  fi: "fin",
-  // Finnish
-  el: "ell",
-  // Greek
-  cs: "ces",
-  // Czech
-  hu: "hun",
-  // Hungarian
-  ro: "ron",
-  // Romanian
-  bg: "bul",
-  // Bulgarian
-  hr: "hrv",
-  // Croatian
-  sk: "slk",
-  // Slovak
-  sl: "slv",
-  // Slovenian
-  uk: "ukr",
-  // Ukrainian
-  tr: "tur",
-  // Turkish
-  // Asian languages
-  th: "tha",
-  // Thai
-  vi: "vie",
-  // Vietnamese
-  id: "ind",
-  // Indonesian
-  ms: "msa",
-  // Malay
-  tl: "tgl",
-  // Tagalog/Filipino
-  // Other languages
-  he: "heb",
-  // Hebrew
-  fa: "fas",
-  // Persian/Farsi
-  bn: "ben",
-  // Bengali
-  ta: "tam",
-  // Tamil
-  te: "tel",
-  // Telugu
-  mr: "mar",
-  // Marathi
-  gu: "guj",
-  // Gujarati
-  kn: "kan",
-  // Kannada
-  ml: "mal",
-  // Malayalam
-  pa: "pan",
-  // Punjabi
-  ur: "urd",
-  // Urdu
-  sw: "swa",
-  // Swahili
-  af: "afr",
-  // Afrikaans
-  ca: "cat",
-  // Catalan
-  eu: "eus",
-  // Basque
-  gl: "glg",
-  // Galician
-  is: "isl",
-  // Icelandic
-  et: "est",
-  // Estonian
-  lv: "lav",
-  // Latvian
-  lt: "lit"
-  // Lithuanian
-};
-var ISO639_3_TO_1 = Object.fromEntries(
-  Object.entries(ISO639_1_TO_3).map(([iso1, iso3]) => [iso3, iso1])
-);
-function toISO639_3(code) {
-  const normalized = code.toLowerCase().trim();
-  if (normalized.length === 3) {
-    return normalized;
-  }
-  return ISO639_1_TO_3[normalized] ?? normalized;
-}
-function toISO639_1(code) {
-  const normalized = code.toLowerCase().trim();
-  if (normalized.length === 2) {
-    return normalized;
-  }
-  return ISO639_3_TO_1[normalized] ?? normalized;
-}
-function getLanguageCodePair(code) {
-  const normalized = code.toLowerCase().trim();
-  if (normalized.length === 2) {
-    return {
-      iso639_1: normalized,
-      iso639_3: toISO639_3(normalized)
-    };
-  } else if (normalized.length === 3) {
-    return {
-      iso639_1: toISO639_1(normalized),
-      iso639_3: normalized
-    };
-  }
-  return {
-    iso639_1: normalized,
-    iso639_3: normalized
-  };
-}
-function getLanguageName(code) {
-  const iso639_1 = toISO639_1(code);
-  try {
-    const displayNames = new Intl.DisplayNames(["en"], { type: "language" });
-    return displayNames.of(iso639_1) ?? code.toUpperCase();
-  } catch {
-    return code.toUpperCase();
-  }
-}
-
-// src/workflows/chapters.ts
 var chapterSchema = z4.object({
   startTime: z4.number(),
   title: z4.string()
@@ -2216,13 +2270,13 @@ var chapterSystemPromptBuilder = createPromptBuilder({
     },
     context: {
       tag: "context",
-      content: dedent3`
+      content: dedent4`
         You receive a timestamped transcript with lines in the form "[12s] Caption text".
         Use those timestamps as anchors to determine chapter start times in seconds.`
     },
     constraints: {
       tag: "constraints",
-      content: dedent3`
+      content: dedent4`
         - Only use information present in the transcript
         - Return structured data that matches the requested JSON schema
         - Do not add commentary or extra text outside the JSON
@@ -2230,7 +2284,7 @@ var chapterSystemPromptBuilder = createPromptBuilder({
     },
     qualityGuidelines: {
       tag: "quality_guidelines",
-      content: dedent3`
+      content: dedent4`
         - Create chapters at topic shifts or clear transitions
         - Keep chapter titles concise and descriptive
         - Ensure the first chapter starts at 0 seconds`
@@ -2240,7 +2294,7 @@ var chapterSystemPromptBuilder = createPromptBuilder({
 });
 var AUDIO_ONLY_SYSTEM_PROMPT_OVERRIDES = {
   role: "You are an audio editor and transcript analyst specializing in segmenting content into logical chapters.",
-  context: dedent3`
+  context: dedent4`
     You receive a timestamped transcript from audio-only content with lines in the form "[12s] Transcript text".
     Use those timestamps as anchors to determine chapter start times in seconds.`
 };
@@ -2252,7 +2306,7 @@ var chaptersPromptBuilder = createPromptBuilder({
     },
     outputFormat: {
       tag: "output_format",
-      content: dedent3`
+      content: dedent4`
         Return valid JSON in this exact shape:
         {
           "chapters": [
@@ -2271,7 +2325,7 @@ var chaptersPromptBuilder = createPromptBuilder({
     },
     titleGuidelines: {
       tag: "title_guidelines",
-      content: dedent3`
+      content: dedent4`
         - Keep titles concise and descriptive
         - Avoid filler or generic labels like "Chapter 1"
         - Use the transcript's terminology`
@@ -2296,7 +2350,7 @@ function buildUserPrompt3({
   if (languageName) {
     contextSections.push(createLanguageSection(languageName));
   }
-  const dynamicChapterGuidelines = dedent3`
+  const dynamicChapterGuidelines = dedent4`
     - Create at least ${minChaptersPerHour} and at most ${maxChaptersPerHour} chapters per hour of content
     - Use start times in seconds (not HH:MM:SS)
     - Chapter start times should be non-decreasing
@@ -2354,7 +2408,7 @@ async function generateChapters(assetId, options = {}) {
     const contentLabel = isAudioOnly ? "transcript" : "caption track";
     throw new MuxAiError(`No usable content found in ${contentLabel}.`, { type: "validation_error" });
   }
-  const resolvedLanguageCode = outputLanguageCode && outputLanguageCode !== "auto" ? outputLanguageCode : transcriptResult.track?.language_code ?? languageCode;
+  const resolvedLanguageCode = outputLanguageCode && outputLanguageCode !== "auto" ? outputLanguageCode : getReliableLanguageCode(transcriptResult.track) ?? languageCode;
   const languageName = resolvedLanguageCode ? getLanguageName(resolvedLanguageCode) : void 0;
   const userPrompt = buildUserPrompt3({
     timestampedTranscript,
@@ -2396,7 +2450,7 @@ async function generateChapters(assetId, options = {}) {
   };
   return {
     assetId,
-    languageCode: languageCode ?? transcriptResult.track?.language_code ?? "en",
+    languageCode: languageCode ?? getReliableLanguageCode(transcriptResult.track) ?? "en",
     chapters: validChapters,
     usage: usageWithMetadata
   };
@@ -2404,7 +2458,7 @@ async function generateChapters(assetId, options = {}) {
 
 // src/workflows/edit-captions.ts
 import { generateText as generateText4, Output as Output4 } from "ai";
-import dedent4 from "dedent";
+import dedent5 from "dedent";
 import { z as z5 } from "zod";
 
 // src/lib/mux-tracks.ts
@@ -2693,7 +2747,7 @@ var profanityDetectionSchema = z5.object({
     "Unique profane words or short phrases exactly as they appear in the transcript text. Include each distinct form only once (e.g., if 'fuck' and 'fucking' both appear, list both)."
   )
 });
-var SYSTEM_PROMPT3 = dedent4`
+var SYSTEM_PROMPT3 = dedent5`
   You are a content moderation assistant. Your task is to identify profane, vulgar, or obscene
   words and phrases in subtitle text. Return ONLY the exact profane words or phrases as they appear
   in the text. Do not modify, censor, or paraphrase them. Do not include words that are merely
@@ -3409,7 +3463,7 @@ async function generateVideoEmbeddings(assetId, options = {}) {
 
 // src/workflows/engagement-insights.ts
 import { generateText as generateText5, Output as Output5 } from "ai";
-import dedent5 from "dedent";
+import dedent6 from "dedent";
 import { z as z6 } from "zod";
 
 // src/primitives/heatmap.ts
@@ -3562,7 +3616,7 @@ var engagementInsightsSchema = z6.object({
   momentInsights: z6.array(aiMomentInsightSchema),
   overallInsight: aiOverallInsightSchema
 });
-var SYSTEM_PROMPT4 = dedent5`
+var SYSTEM_PROMPT4 = promptDedent`
   <role>
     You are a video engagement analyst specializing in understanding viewer behavior.
     Your job is to explain why specific moments in videos have high or low engagement
@@ -3590,12 +3644,10 @@ var SYSTEM_PROMPT4 = dedent5`
 
   <constraints>
     - Only describe what you can see in images or read in transcripts
-    - Do not fabricate details or make unsupported assumptions
-    - Do NOT use any metadata such as URLs, file paths, domain names, file names,
-      playback IDs, or technical parameters visible in this request. These are
-      delivery infrastructure and are unrelated to the media content itself.
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${METADATA_BOUNDARY_WARNING}
     - Correlate engagement scores with observable content
-    - Return structured data matching the requested schema exactly
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Each momentInsight MUST include the hotspotIndex from the input data
   </constraints>
 
@@ -3606,7 +3658,7 @@ var SYSTEM_PROMPT4 = dedent5`
     - Connect engagement data to content features
   </quality_guidelines>
 `;
-var AUDIO_ONLY_SYSTEM_PROMPT2 = dedent5`
+var AUDIO_ONLY_SYSTEM_PROMPT2 = promptDedent`
   <role>
     You are a video engagement analyst specializing in understanding viewer behavior for audio content.
     Your job is to explain why specific moments have high or low engagement based on audio/dialogue and viewer watch patterns.
@@ -3633,12 +3685,10 @@ var AUDIO_ONLY_SYSTEM_PROMPT2 = dedent5`
 
   <constraints>
     - Only describe what you can read in the transcript
-    - Do not fabricate details or make unsupported assumptions
-    - Do NOT use any metadata such as URLs, file paths, domain names, file names,
-      playback IDs, or technical parameters visible in this request. These are
-      delivery infrastructure and are unrelated to the media content itself.
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${METADATA_BOUNDARY_WARNING}
     - Correlate engagement scores with observable content
-    - Return structured data matching the requested schema exactly
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Each momentInsight MUST include the hotspotIndex from the input data
   </constraints>
 
@@ -3649,7 +3699,7 @@ var AUDIO_ONLY_SYSTEM_PROMPT2 = dedent5`
     - Connect engagement data to content features
   </quality_guidelines>
 `;
-var INSIGHT_GUIDELINES = dedent5`
+var INSIGHT_GUIDELINES = dedent6`
   Provide explanatory insights that describe WHY moments are engaging.
   Focus on observable patterns and viewer behavior correlation.
   Example: "The cooking demonstration at 2:15 has 3x average engagement
@@ -3667,7 +3717,7 @@ var engagementInsightsPromptBuilder = createPromptBuilder({
     },
     outputFormat: {
       tag: "output_format",
-      content: dedent5`
+      content: dedent6`
         Return valid JSON matching the provided schema.
         Include insights for each hotspot and overall engagement trends.
         Each momentInsight must include the hotspotIndex matching the input.
@@ -3675,7 +3725,7 @@ var engagementInsightsPromptBuilder = createPromptBuilder({
     },
     visualContext: {
       tag: "visual_context",
-      content: dedent5`
+      content: dedent6`
         Images are provided for each hotspot timestamp.
         A storyboard overview image may also be included showing the full video timeline.
         Use visual evidence from these images to support your engagement analysis.
@@ -3689,14 +3739,14 @@ function buildUserPrompt4(hotspots, transcriptSegments, heatmap, isAudioOnly) {
     const transcript = transcriptSegments[idx]?.text || "(no transcript)";
     const startTimestamp = secondsToTimestamp(h.startMs / 1e3);
     const endTimestamp = secondsToTimestamp(h.endMs / 1e3);
-    return dedent5`
+    return dedent6`
         Hotspot ${idx} (hotspotIndex: ${idx}):
         - Time Range: ${startTimestamp} - ${endTimestamp}
         - Engagement Score: ${h.score.toFixed(2)} (0=low, 1=high)
         - Transcript: "${transcript}"
       `;
   }).join("\n\n");
-  const heatmapData = dedent5`
+  const heatmapData = dedent6`
     Engagement Heatmap (${heatmap.length} values, each representing 1/${heatmap.length}th of the video, higher = more re-watches):
     [${heatmap.map((v) => v.toFixed(2)).join(", ")}]
   `;
@@ -4605,7 +4655,7 @@ async function getModerationScores(assetId, options = {}) {
 
 // src/workflows/summarization.ts
 import { generateText as generateText6, Output as Output6 } from "ai";
-import dedent6 from "dedent";
+import dedent7 from "dedent";
 import { z as z7 } from "zod";
 var DEFAULT_SUMMARY_KEYWORD_LIMIT = 10;
 var DEFAULT_TITLE_LENGTH = 10;
@@ -4631,36 +4681,36 @@ var DESCRIPTION_LENGTH_THRESHOLD_LARGE = 100;
 function buildDescriptionGuidance(wordCount, contentType) {
   if (wordCount < DESCRIPTION_LENGTH_THRESHOLD_SMALL) {
     if (contentType === "video") {
-      return dedent6`A brief summary of the video in no more than ${wordCount} words. Shorter is fine.
+      return dedent7`A brief summary of the video in no more than ${wordCount} words. Shorter is fine.
         Focus on the single most important subject or action.
         Write in present tense.`;
     }
-    return dedent6`A brief summary of the audio content in no more than ${wordCount} words. Shorter is fine.
+    return dedent7`A brief summary of the audio content in no more than ${wordCount} words. Shorter is fine.
       Focus on the single most important topic or theme.
       Write in present tense.`;
   }
   if (wordCount > DESCRIPTION_LENGTH_THRESHOLD_LARGE) {
     if (contentType === "video") {
-      return dedent6`A detailed summary that describes what happens across the video.
+      return dedent7`A detailed summary that describes what happens across the video.
         Never exceed ${wordCount} words, but shorter is perfectly fine. You may use multiple sentences.
         Be thorough: cover subjects, actions, setting, progression, and any notable details visible across frames.
         Write in present tense. Be specific about observable details rather than making assumptions.
         If the transcript provides dialogue or narration, incorporate key points but prioritize visual content.`;
     }
-    return dedent6`A detailed summary that describes the audio content.
+    return dedent7`A detailed summary that describes the audio content.
       Never exceed ${wordCount} words, but shorter is perfectly fine. You may use multiple sentences.
       Be thorough: cover topics, speakers, themes, progression, and any notable insights.
       Write in present tense. Be specific about what is discussed or presented rather than making assumptions.
       Focus on the spoken content and any key insights, dialogue, or narrative elements.`;
   }
   if (contentType === "video") {
-    return dedent6`A summary that describes what happens across the video.
+    return dedent7`A summary that describes what happens across the video.
       Never exceed ${wordCount} words, but shorter is perfectly fine. You may use multiple sentences.
       Cover the main subjects, actions, setting, and any notable progression visible across frames.
       Write in present tense. Be specific about observable details rather than making assumptions.
       If the transcript provides dialogue or narration, incorporate key points but prioritize visual content.`;
   }
-  return dedent6`A summary that describes the audio content.
+  return dedent7`A summary that describes the audio content.
     Never exceed ${wordCount} words, but shorter is perfectly fine. You may use multiple sentences.
     Cover the main topics, speakers, themes, and any notable progression in the discussion or narration.
     Write in present tense. Be specific about what is discussed or presented rather than making assumptions.
@@ -4677,7 +4727,7 @@ function createSummarizationBuilder({ titleLength, descriptionLength, tagCount }
       },
       title: {
         tag: "title_requirements",
-        content: dedent6`
+        content: dedent7`
           A concise, label-style title — not a sentence or description.
           Never exceed ${titleLimit} words, but shorter is better.
           Think of how a video card title, playlist entry, or file name would read — e.g. "Predator: Badlands Trailer" or "Chef Prepares Holiday Feast".
@@ -4690,7 +4740,7 @@ function createSummarizationBuilder({ titleLength, descriptionLength, tagCount }
       },
       keywords: {
         tag: "keywords_requirements",
-        content: dedent6`
+        content: dedent7`
           Specific, searchable terms (up to ${keywordLimit}) that capture:
           - Primary subjects (people, animals, objects)
           - Actions and activities being performed
@@ -4702,7 +4752,7 @@ function createSummarizationBuilder({ titleLength, descriptionLength, tagCount }
       },
       qualityGuidelines: {
         tag: "quality_guidelines",
-        content: dedent6`
+        content: dedent7`
           - Examine all frames to understand the full context and progression
           - Be precise: "golden retriever" is better than "dog" when identifiable
           - Capture the narrative: what begins, develops, and concludes
@@ -4723,7 +4773,7 @@ function createAudioOnlyBuilder({ titleLength, descriptionLength, tagCount } = {
       },
       title: {
         tag: "title_requirements",
-        content: dedent6`
+        content: dedent7`
           A concise, label-style title — not a sentence or description.
           Never exceed ${titleLimit} words, but shorter is better.
           Think of how a podcast episode title or playlist entry would read — e.g. "Weekly News Roundup" or "Interview with Dr. Smith".
@@ -4736,7 +4786,7 @@ function createAudioOnlyBuilder({ titleLength, descriptionLength, tagCount } = {
       },
       keywords: {
         tag: "keywords_requirements",
-        content: dedent6`
+        content: dedent7`
           Specific, searchable terms (up to ${keywordLimit}) that capture:
           - Primary topics and themes
           - Speakers or presenters (if named)
@@ -4748,7 +4798,7 @@ function createAudioOnlyBuilder({ titleLength, descriptionLength, tagCount } = {
       },
       qualityGuidelines: {
         tag: "quality_guidelines",
-        content: dedent6`
+        content: dedent7`
           - Analyze the full transcript to understand context and themes
           - Be precise: use specific terminology when mentioned
           - Capture the narrative: what is introduced, discussed, and concluded
@@ -4758,15 +4808,14 @@ function createAudioOnlyBuilder({ titleLength, descriptionLength, tagCount } = {
     sectionOrder: ["task", "title", "description", "keywords", "qualityGuidelines"]
   });
 }
-var SYSTEM_PROMPT5 = dedent6`
+var SYSTEM_PROMPT5 = promptDedent`
   <role>
     You are a video content analyst specializing in storyboard interpretation and multimodal analysis.
   </role>
 
   <context>
     You receive storyboard images containing multiple sequential frames extracted from a video.
-    These frames are arranged in a grid and represent the visual progression of the content over time.
-    Read frames left-to-right, top-to-bottom to understand the temporal sequence.
+    ${STORYBOARD_FRAME_INSTRUCTIONS}
   </context>
 
   <transcript_guidance>
@@ -4788,37 +4837,21 @@ var SYSTEM_PROMPT5 = dedent6`
 
   <constraints>
     - Only describe what is clearly observable in the frames or explicitly stated in the transcript
-    - Do not fabricate details or make unsupported assumptions
-    - Do NOT use any metadata such as URLs, file paths, domain names, file names,
-      playback IDs, or technical parameters visible in this request. These are
-      delivery infrastructure and are unrelated to the media content itself.
-    - Return structured data matching the requested schema
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${METADATA_BOUNDARY_WARNING}
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Output only the JSON object; no markdown or extra text
     - When a <language> section is provided, all output text MUST be written in that language
   </constraints>
 
   <tone_guidance>
-    Pay special attention to the <tone> section and lean heavily into those instructions.
-    Adapt your entire analysis and writing style to match the specified tone - this should influence
-    your word choice, personality, formality level, and overall presentation of the content.
-    The tone instructions are not suggestions but core requirements for how you should express yourself.
+    ${TONE_GUIDANCE}
   </tone_guidance>
 
   <language_guidelines>
-    AVOID these meta-descriptive phrases that reference the medium rather than the content:
-    - "The image shows..." / "The storyboard shows..."
-    - "In this video..." / "This video features..."
-    - "The frames depict..." / "The footage shows..."
-    - "We can see..." / "You can see..."
-    - "The clip shows..." / "The scene shows..."
-
-    INSTEAD, describe the content directly:
-    - BAD: "The video shows a chef preparing a meal"
-    - GOOD: "A chef prepares a meal in a professional kitchen"
-
-    Write as if describing reality, not describing a recording of reality.
+    ${createLanguageGuidelines("video")}
   </language_guidelines>`;
-var AUDIO_ONLY_SYSTEM_PROMPT3 = dedent6`
+var AUDIO_ONLY_SYSTEM_PROMPT3 = promptDedent`
   <role>
     You are an audio content analyst specializing in transcript analysis and metadata generation.
   </role>
@@ -4845,35 +4878,20 @@ var AUDIO_ONLY_SYSTEM_PROMPT3 = dedent6`
 
   <constraints>
     - Only describe what is explicitly stated or strongly implied in the transcript
-    - Do not fabricate details or make unsupported assumptions
-    - Do NOT use any metadata such as URLs, file paths, domain names, file names,
-      playback IDs, or technical parameters visible in this request. These are
-      delivery infrastructure and are unrelated to the media content itself.
-    - Return structured data matching the requested schema
+    - ${NO_FABRICATION_CONSTRAINT}
+    - ${METADATA_BOUNDARY_WARNING}
+    - ${STRUCTURED_DATA_CONSTRAINT}
     - Focus entirely on audio/spoken content - there are no visual elements
     - Output only the JSON object; no markdown or extra text
     - When a <language> section is provided, all output text MUST be written in that language
   </constraints>
 
   <tone_guidance>
-    Pay special attention to the <tone> section and lean heavily into those instructions.
-    Adapt your entire analysis and writing style to match the specified tone - this should influence
-    your word choice, personality, formality level, and overall presentation of the content.
-    The tone instructions are not suggestions but core requirements for how you should express yourself.
+    ${TONE_GUIDANCE}
   </tone_guidance>
 
   <language_guidelines>
-    AVOID these meta-descriptive phrases that reference the medium rather than the content:
-    - "The audio shows..." / "The transcript shows..."
-    - "In this recording..." / "This audio features..."
-    - "The speaker says..." / "We can hear..."
-    - "The clip contains..." / "The recording shows..."
-
-    INSTEAD, describe the content directly:
-    - BAD: "The audio features a discussion about climate change"
-    - GOOD: "A panel discusses climate change impacts and solutions"
-
-    Write as if describing reality, not describing a recording of reality.
+    ${createLanguageGuidelines("audio")}
   </language_guidelines>`;
 function buildUserPrompt5({
   tone,
@@ -5047,7 +5065,7 @@ async function getSummaryAndTags(assetId, options) {
     required: isAudioOnly
   }) : void 0;
   const transcriptText = transcriptResult?.transcriptText ?? "";
-  const resolvedLanguageCode = outputLanguageCode && outputLanguageCode !== "auto" ? outputLanguageCode : transcriptResult?.track?.language_code ?? getReadyTextTracks(assetData)[0]?.language_code;
+  const resolvedLanguageCode = outputLanguageCode && outputLanguageCode !== "auto" ? outputLanguageCode : getReliableLanguageCode(transcriptResult?.track) ?? getReliableLanguageCode(getReadyTextTracks(assetData)[0]);
   const languageName = resolvedLanguageCode ? getLanguageName(resolvedLanguageCode) : void 0;
   const userPrompt = buildUserPrompt5({
     tone,
@@ -5539,18 +5557,18 @@ import {
   RetryError,
   TypeValidationError
 } from "ai";
-import dedent7 from "dedent";
+import dedent8 from "dedent";
 import { z as z8 } from "zod";
 var translationSchema = z8.object({
   translation: z8.string()
 });
-var SYSTEM_PROMPT6 = dedent7`
+var SYSTEM_PROMPT6 = dedent8`
   You are a subtitle translation expert. Translate VTT subtitle files to the target language specified by the user.
   You may receive either a full VTT file or a chunk from a larger VTT.
   Preserve all timestamps, cue ordering, and VTT formatting exactly as they appear.
   Return JSON with a single key "translation" containing the translated VTT content.
 `;
-var CUE_TRANSLATION_SYSTEM_PROMPT = dedent7`
+var CUE_TRANSLATION_SYSTEM_PROMPT = dedent8`
   You are a subtitle translation expert.
   You will receive a sequence of subtitle cues extracted from a VTT file.
   Translate the cues to the requested target language while preserving their original order.
@@ -6061,7 +6079,7 @@ async function translateCaptions(assetId, trackId, toLanguageCode, options) {
     }
     if (uploadToMux) {
       try {
-        const languageName = getLanguageName(toLanguageCode);
+        const languageName = getLanguageName(toLanguageCode) ?? toLanguageCode.toUpperCase();
         const trackName = `${languageName} (auto-translated)`;
         uploadedTrackId = await createTextTrackOnMux(
           assetId,