edit-jianying-manga-narration

# 剪映漫画解说制作 Treat the user's supplied prose and image-to-paragraph mapping as exact input. Do not rewrite, summarize, “correct,” or reorder them unless explicitly asked. ## Fixed defaults - Continue the existing Jianying project; do not start a new project. - Treat the user's original image-to-paragraph mapping as the narration manifest: one image (or one explicitly mapped image group) corresponds to one complete source paragraph and one narration clip. - Generate narration from the complete source paragraph. Never use exported SRT, auto-captions, or display-caption chunks as TTS input; captions are derived only after the paragraph audio exists. Treat duration-weighted caption times as estimates unless reviewed `caption_cues` were supplied. - Captions: at most 9 Unicode characters per item, black text, solid white background plate, horizontally safe with visible margins. - For caption-only work, split the authoritative paragraph semantically before opening Jianying. Require exact concatenation, no punctuation-only item, at most 9 Unicode characters per item, relative timestamps beginning at zero, and an end time matching the existing audio segment. - Preserve the supplied image order and paragraph mapping. - Keep later material synchronized when replacing or inserting content. - Make a recoverable project backup before editing and verify the saved draft afterward. - Default to an external-first atomic segment: one image or explicit image group, one complete paragraph narration, and its burned captions in one `preview.mp4`. Jianying should normally receive one finished MP4 per independently replaceable segment. - For `解说小帅`, use Jianying only to generate one complete-paragraph audio clip, then feed that clip back into the external package builder. Do not build the pictures or display captions in Jianying merely because the voice came from Jianying. - If no external narration clip is supplied, the package builder uses the generated source voice `Tingting`; disclose this. Use separate clean video, audio, and SRT assets only when the user explicitly needs editable captions or independently editable audio. ## Execution 1. Read [references/workflow.md](references/workflow.md) before operating Jianying or editing its draft data. 2. Use Computer Use for Jianying UI actions. Re-query accessibility state after every UI change; coordinates are hints only. 3. Use `scripts/backup_project.sh` before mutation. 4. Resolve or create a paragraph manifest from the original prose and image mapping before generating audio. Stop if the complete source paragraphs cannot be recovered; do not reconstruct narration from subtitle fragments. 5. For each manifest item, generate one complete narration clip from that item's paragraph. Build captions afterward from the same paragraph text. Supply reviewed relative `caption_cues`, or disclose the duration-weighted estimate; set `require_provided_caption_timing` when estimates are forbidden. Never claim supplied audio says the source text unless it was actually checked. 6. Build and review the atomic `preview.mp4` outside Jianying. It already contains narration and burned captions; never add its source audio again, or the audio will be duplicated. 7. Append atomic MP4s in order. For a middle insertion, ripple every later video, audio, caption, and outro segment by exactly the MP4 duration before filling the gap. Compare a before/after per-track segment inventory; endpoint equality alone cannot prove that no middle segment was omitted. 8. For caption-only insertion, follow the dedicated procedure in `references/workflow.md`. The order is mandatory: validate the SRT, activate the project timeline and confirm its target timecode, then click only the subtitle tile's small add control. Never position the project after opening the subtitle asset preview; its `currentProgress` belongs to the asset, not the project. Immediately undo if project duration increases. 9. For the `解说小帅` route, create one temporary text item from each complete source paragraph, use `文本朗读`, choose `解说小帅`, and recover the resulting paragraph audio. Pass that audio as the paragraph's `audio` field to `prepare_package.js`; derive picture timing and display captions from its measured duration. Do not use `换音色` or create one voice clip per display subtitle. 10. When narration and captions stay unchanged but one existing long visual must show several images at specified caption boundaries, use the equal-duration silent replacement-video procedure in `references/workflow.md`. In Jianying's replacement dialog, turn off `复用原视频效果`; otherwise the old still can remain visible even though the new media was selected. 11. Use deterministic draft editing only for insertion into an existing long timeline when direct import cannot preserve the required offset, and only when `draft_info.json` parses as JSON, the app is closed, and a backup exists. 12. Finish only after checking caption-text exactness, declared narration-text verification, one-paragraph-per-audio structure, maximum caption length, caption timing source, image order and timing source, original start offset, new total duration, every later segment's shift, caption plate, audio presence, asset availability, and a successful save/reopen check. Verify `解说小帅` only when that route was requested. Report partial completion honestly. Do not claim the voice or white plate is applied merely because draft fields were written; verify them in Jianying.