旁白版竖屏短剧工作流

Generate a complete, editable 9:16 short-drama episode in narration mode: AI-generated
video clips + story-voice narration + background music + hand-written Motion Graphics + serif-band
subtitles, assembled into a ChatCut timeline and exported as 1080p H.264. Every step lands in a
real, re-editable ChatCut project (never a flattened local file).

When to use

• User wants an episode built the way the first three 《重生2026》 episodes were (旁白驱动).
• User says "按前三集模式生成", "生成旁白版", "做下一集", or picks a character photo and says "用这个生成下一集".
• Do not use for dialogue-driven / character-locked / no-narration requests — route those to
  reborn2026-dialogue-mode instead.

Prerequisites (verify before starting)

• ChatCut MCP tools are reachable via DeferExecuteTool under the mcp__chatcut__ prefix
  (e.g. mcp__chatcut__create_project). If a call returns "not found in the deferred tools index",
  run ToolSearch with the exact tool name, then retry — the index intermittently drops.
• VideoGen / ImageGen are reachable via DeferExecuteTool.
• Local media upload helper exists at:
  ~/.workbuddy/plugins/marketplaces/experts/plugins/chatcut-video-editor/skills/asset-import/scripts/upload-media.mjs
  (or follow the asset-import skill). It needs ffmpeg/ffprobe binaries and a ChatCut import token.
• ffmpeg/ffprobe binaries (win32-x64) live under ~/.workbuddy/binaries/node/workspace/node_modules/@ffmpeg-installer and @ffprobe-installer. Pass their absolute Windows paths (C:/...) to the uploader.

Workflow (high level)

1. Confirm the story. If no script exists, draft a narration script (one section per clip,
   ~80–110s total, ~430 Chinese chars of pure narration) and confirm with the user. Reuse prior
   episode scripts for tone/format. Save it as a .md file.
2. Generate 6 video clips with VideoGen (text-to-video for establishing/insert shots;
   image-to-video with a character reference PNG to lock the face for that character's shots).
   - 720P, 5s, enable_audio:false, watermark:false.
   - Portrait trap: text-to-video often outputs 1280×720 landscape even when the prompt says
   "9:16". Always pass aspect_ratio:"9:16" for text-to-video. Image-to-video must NOT pass
   aspect_ratio (it inherits the reference image's 2:3 ratio).
   - Rename outputs to epN-segK-*.mp4 and verify with ffprobe (width,height,duration).
3. Generate narration via mcp__chatcut__submit_voice (provider doubao, voiceId
   qingcang, emotion:"storytelling", emotionScale:3, a performancePrompt for story-telling
   tone). Generate music via mcp__chatcut__submit_music (Mureka-style prompt: piano + strings,
   emotional, loopable). Bind both to the project with projectId. Wait with
   mcp__chatcut__track_progress {action:"wait", target:"generation", jobIds:"..."}.
4. Create project + upload videos. mcp__chatcut__create_project
   (compositionWidth:1080, compositionHeight:1920, fps:30). Get an import token via
   mcp__chatcut__import_media {action:"create_session", projectId} (token valid ~30 min), then
   upload the 6 clips in batches of ≤4 with the upload helper. Capture each returned assetId.
5. Author 4 Motion Graphics via mcp__chatcut__create_motion_graphic_from_code. Free (no
   credits). Use inline JSX with Remotion globals; root element MUST be a div with style={rootStyle}
   (NOT AbsoluteFill — validation fails). Noto Serif SC font. Typical set: opening title /
   character-name reveal / key-quote emphasis / ending freeze.
6. Assemble the timeline with a single mcp__chatcut__edit_item call (see critical gotcha
   below). Read tracks first with mcp__chatcut__read_project {view:"timeline"}.
7. Transcribe + subtitles. Generated voice is NOT auto-transcribed. Kick ASR with
   mcp__chatcut__manage_transcript {action:"retry_transcription", asset:"(narration asset id)"},
   wait ~35s, then mcp__chatcut__edit_captions {action:"enable", preset:"serif-band"} and
   {action:"source_set", json:{"sources":[{"trackId":"A1"}]}} to lock subtitles to the narration
   track. Fix ASR name-mishearing with edit_captions {action:"display_text", json:{overrides:[...]}}.
8. Export via mcp__chatcut__submit_export {format:"video", codec:"h264", resolution:"1080p", fps:30} and wait with mcp__chatcut__track_export. Download the S3 URL and verify with ffprobe.

CRITICAL gotchas (learned the hard way)

• edit_item auto-track-shuffle (severe): when updating/adding an item without trackId, the
  tool may auto-move it to a new track to avoid same-track overlap, scattering the timeline. Always
  pass trackId on every add. Prefer one big adds:[...] call with all 12 items (6 videos + 4 MG +
  2 audio), each carrying its trackId.
• No cross-dissolve on slow-stretched clips: stretching a 5s clip to cover ~7s leaves no extra
  handle at the ends, so cross-dissolve reports "seam has 0 feasible visual transition frames". Use
  hard cuts (segments meet end-to-end) for narration-mode episodes.
• Track roles: A1 = role:"anchor" (narration), A2 = role:"follower" (music) → automatic
  ducking. Create them explicitly with mcp__chatcut__edit_track.
• ASR mishears names: narration TTS "紫日"→"子日", "白露"→"白鹿", "怔住"→"正住". Fix with
  edit_captions display_text overrides (read word keys via read_captions {words:true}). This
  edits only the displayed text, not the audio.
• submit_video needs Pro: native ChatCut video generation requires a Pro plan. Use VideoGen
  (DeferExecuteTool) + the upload helper instead.
• Upload path pitfall: the upload helper runs under node — pass Windows-style absolute paths
  (C:/Users/...), not Git-Bash /c/... (node mangles /c/ into C:\c\).
• Index flakiness: edit_track / edit_item / track_progress occasionally drop from the
  deferred index mid-session. On "not found", ToolSearch the name and retry.

Verification

• read_project {view:"timeline"}: confirm 4 tracks, 12 items, durations add up, no stray tracks.
• read_captions {fullText:true}: confirm subtitle text is complete and names are correct.
• After export: ffprobe the downloaded file — expect H.264, 1080×1920, 30fps, duration ≈ narration
  length, size ~38–45 MB.

References

See references/workflow.md for the full per-step parameter recipes (exact JSON shapes, frame math
for slow-stretch, and the serif-band caption fix).