localize-content

name	localize-content
description	Produce existing content in more languages — translated copy, re-generated voiceover in the same brand voice, per-language captions and publish packs. Use when the user says "make this in Spanish", "translate my ad", "localize for Thailand", "multi-language version".
allowed-tools	Bash Read Write Edit Glob Grep

Localize Content

One piece of content, N markets. The visuals are already paid for — localization re-generates only the cheap layers: script, voiceover (~$0.01), captions (free), and copy. A $10 video becomes 3 markets for ~$0.10.

STEP 1: Scope

Ask once: which languages/markets? Then identify what the piece contains (check its manifest — see /revise-content STEP 1):

Layer	Localize how	Cost
Script / voiceover	translate → generateVoiceover per language	~$0.01 each
On-screen captions (.srt)	generateCaptions from translated script	free
Caption/post copy + hashtags	translate + adapt (see STEP 2)	~$0.001
Text baked INTO the image/video	can't swap text in-place — flag it	see below

Baked-in text warning: if the visual has burned-in text (headline on a thumbnail, text overlay in a scene), tell the user up front: that layer needs regeneration per language (/revise-content for that piece, translated prompt). Get approval before this — it's the only expensive part.

STEP 2: Translate like a marketer, not a dictionary

Use generateText (or translate yourself if fluent) with instructions to:

• Keep the brand tone from brand.md — a warm-casual brand must stay warm-casual in every language.
• Transcreate hooks, idioms, and CTAs — a literal pun translation is worse than a new hook with the same intent.
• Respect restrictions (never_say lists apply in every language; add market-specific legal terms if the user names any).
• Keep timing: a voiceover script must fit the same video duration — translations run 10-30% longer in many languages, so tighten wording, don't rush the voice.
• Localize hashtags (research/ask — don't literally translate tags) and numbers/currency formats.

Show the user each translation before generating audio. If they don't read the language, say what you adapted and why.

STEP 3: Regenerate the cheap layers per language

# Voiceover — SAME locked voice id from the asset registry for brand continuity
node workflows/cli.cjs generateVoiceover '{"script":"<translated script>","voiceName":"Kore","outputPath":".../es/voiceover.wav"}'

# Remix onto the existing video (local, free)
node workflows/cli.cjs mixVideoAudio '{"videoPath":".../final.mp4","voiceoverPath":".../es/voiceover.wav","musicPath":".../music.wav","outputPath":".../es/final-es.mp4"}'

# Captions from the translated script (local, free)
node workflows/cli.cjs generateCaptions '{"script":"<translated script>","totalDuration":30,"outputPath":".../es/captions.srt"}'

Check the voice supports the language (Gemini TTS is multilingual, but listen to the first output before batching the rest). Non-Latin scripts (Thai, Myanmar, Arabic…): confirm caption font/rendering — see the content-preflight edge-case checklist.

STEP 4: Package per language

/package-content each localized cut into output-contents/{piece}/publish/{lang}/{platform}/ with the translated caption, localized hashtags, and link. Then report:

✅ LOCALIZED — summer-ad → ES, TH, MM
  es/  final-es.mp4 + captions + IG/TikTok packs   $0.011
  th/  final-th.mp4 + captions + IG/TikTok packs   $0.012
  mm/  final-mm.mp4 + captions + FB pack           $0.012
  Flagged: thumbnail has baked-in English headline — regenerate per language? (~$0.07 each)

recordCost the spend. Save translated scripts alongside outputs so future revisions localize from the same source.

Pipeline-first + prompts.txt (mandatory)

• Author <content-id>.pipeline.json in each language's folder BEFORE generating (nodes = CLI commands — voiceover → mix → captions — {{node.data.field}} refs), then node workflows/cli.cjs runPipeline @<file>. See workflows/pipelines/README.md.
• Each language folder gets a prompts.txt with the translated script (audio tags included) and voice params; update it on every retry with a one-line RESULT note.