القائمة
Use when rough-cutting Chinese talking-head video with local FunASR transcription, silence trimming, repetition/filler detection, subtitles, and an audit page.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Use when Dasheng Publish needs to route Xiaohongshu publishing or content access through API-first CLI, MCP, or browser fallback bridges.
Use when Dasheng Publish needs to prepare, upload, or package Bilibili video submissions from completed transwrite video channel packs.
Use when running Stage 3 Draft from approved Brief topics and producing Reasoning Sheets, standard drafts, and draft quality gates.
Use when the workflow enters Postmortem after Publish and published content needs to be evaluated with performance data, Publish Guard status, accuracy checks, and knowledge-base writeback suggestions.
Use when Dasheng draft or transwrite needs HTML Anything templates, editable article HTML, social cards, posters, or visual HTML references.
Use when Dasheng transwrite needs no-human or human-material talking-head videos rendered through the local html-video project.
| name | dasheng-video-roughcut |
| description | Use when rough-cutting Chinese talking-head video with local FunASR transcription, silence trimming, repetition/filler detection, subtitles, and an audit page. |
这是 Transwrite 的口播视频辅助环节，处理“用户已经有真人口播素材”的场景。
目标是先给出可审核的粗剪版本，不直接替代人工精剪。
它不是独立主阶段，而是 transwrite -> talking_head_video 的“初检/粗剪”子环节。后续视频视觉层、HTML 动画、字幕烧录和平台导出继续留在 video lane 内组合。
.mp4 / .movparaformer-zh + fsmn-vad + ct-puncafftdn、dynaudnorm、acompressor、loudnorm、alimiter当前口播粗剪保留两条并行路径：
FunASR -> Agent 字幕/语义整理 -> FFmpeg 精剪 -> HTML 审核页，作为可复现主链路。剪映专业版 -> 时间线片段右键 -> 智能剪口播，作为实验通路，用于测试剪映内置去水词、去重复和停顿识别能力。剪映路径只测试剪映自己的智能剪口播结果，不再叠加 Agent 根据原提纲做二次语义检查。
.venv_media/bin/python scripts/video_roughcut_funasr.py \
--input-video "/path/to/source.mp4" \
--output-dir "产物/06_转写生产/<run_id>/talking_head_video/roughcut" \
--mode balanced
先生成 Agent 整理输入：
python3 scripts/video_roughcut_agent_align.py \
--source-video "/path/to/source.mp4" \
--segments-json "roughcut/work/segments.json" \
--output-dir "roughcut_agent_refine"
Agent 输出 agent_plan.json 后渲染：
python3 scripts/video_roughcut_agent_align.py \
--source-video "/path/to/source.mp4" \
--segments-json "roughcut/work/segments.json" \
--agent-plan "roughcut_agent_refine/agent_plan.json" \
--output-dir "roughcut_agent_refine"
首次使用媒体环境：
.venv_media/bin/python -m pip install -r requirements-media.txt
roughcut_manifest.jsonwork/funasr_raw.jsonwork/segments.jsonwork/delete_segments.jsonwork/keep_segments.jsonfinal/*_roughcut_funasr.mp4final/*_roughcut_funasr.srtfinal/*_roughcut_funasr_softsub.mp4review/3_review_live.htmlreview/review_server.jsreview/start_review.commandreview/2_candidates.jsonreview/3_delete_segments.jsonreview/3_delete_segments.reviewed.json（人工保存后生成）review/reviewed_output_loud.mp4（保存并重剪后生成）review/reviewed_output_loud.srt（保存并重剪后生成）review/reviewed_output_loud_softsub.mp4（保存并重剪后生成）review/proofread_agent_input.md（保存并重剪后生成）cd "<output-dir>/review"
PORT=8899 node review_server.js
浏览器打开 http://localhost:8899/。
3_delete_segments.reviewed.json。适用条件：
操作顺序：
注意：
delete_segments.json，供审核页追溯。highpass -> lowpass -> afftdn -> dynaudnorm -> acompressor -> loudnorm(I=-14) -> alimiter。dasheng-video-open-filter。review/proofread_agent_input.md / review/proofread_agent_input.json。requirements-media.txt，不要静默退回低质量转写。