Live Slice

直播切片

Use this skill to convert a local livestream video into transcript-backed short-video clip plans and exports. Local media work uses direct ffmpeg/ffprobe commands only; MCP tools handle OSS upload, TingWu analysis, and LLM JSON planning.

Default Artifacts

Use one output directory per source video:

Workflow

1. Check media tools: command -v ffmpeg && command -v ffprobe

   If either command is missing, tell the operator to install FFmpeg. Python is not required.

2. Prepare artifacts:

   mkdir -p "$DIR" "$DIR/exports"
   ffprobe -v error -show_format -show_streams -of json "$VIDEO" > "$DIR/metadata.json"
   ffmpeg -y -i "$VIDEO" -vn -ac 1 -ar 16000 -codec:a libmp3lame -q:a 4 "$DIR/audio.mp3"
   ffmpeg -y -ss 0 -i "$VIDEO" -frames:v 1 -q:v 2 "$DIR/cover.jpg"
   

3. Upload audio: Call upload_live_audio(file_path="$DIR/audio.mp3"). If storage is local, ask the operator to configure OSS or provide a direct audio_url.

4. Create TingWu task: Call create_live_analysis_task(audio_url=..., auto_chapters_enabled=true, summarization_enabled=true, meeting_assistance_enabled=true, diarization_enabled=false, script_template_enable=true).

5. Poll analysis: Call query_live_analysis_task(task_id=...) until status is COMPLETED, then save the JSON to analysis.json.

6. Plan cleanup and cuts:

   • Call recognize_live_invalid_sentences(sentences=analysis.sentences) and save invalid-sentences.json.
   • Remove invalid indexes from analysis.sentences.
   • Call recognize_live_segments(sentences=valid_sentences, ask=optional_user_goal) and save segments.json.
   • Detect source orientation with ffprobe -v error -select_streams v:0 -show_entries stream=width,height -of csv=s=x:p=0 "$VIDEO" (store $SRC_W/$SRC_H; $SRC_W > $SRC_H is landscape).
   • Call build_live_clip_plan(sentences=analysis.sentences, segments=segments.segments, invalid=invalid.invalid, video_path="$VIDEO", output_dir="$DIR", target_mode="vertical", vertical_fill="blur", source_width=$SRC_W, source_height=$SRC_H, target_width=1080, target_height=1920, normalize_audio_loudness=true, head_padding_seconds=0.15, tail_padding_seconds=0.30, min_duration_seconds=5, max_duration_seconds=120) and save clip-plan.json.
   • Use recognize_live_subjects and complete_live_subject when the user wants topic-driven clips instead of broad segments, then call build_live_subject_clip_plan(sentences=analysis.sentences, completions=subject_completions, invalid=invalid.invalid, video_path="$VIDEO", output_dir="$DIR", target_mode="vertical", vertical_fill="blur", source_width=$SRC_W, source_height=$SRC_H, target_width=1080, target_height=1920, normalize_audio_loudness=true, head_padding_seconds=0.15, tail_padding_seconds=0.30, min_duration_seconds=5, max_duration_seconds=120) and save subject-clip-plan.json.
   • Never hand-convert non-contiguous subject scripts into segments.json; use build_live_subject_clip_plan.

7. Cut clips: Read each item in the selected plan (clip-plan.json or subject-clip-plan.json). Before each command, create parent directories for planned output, parts[].output, and concat_list_path. For single-part clips, run fast_cut_shell first, then use ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "$OUT" to read output duration. If the command fails, the output is missing or empty, ffprobe fails, or actual duration differs from planned duration by more than max(1.0s, duration*0.05), run accurate_cut_shell and read duration again. For multi-part clips, run every parts[].accurate_cut_shell, write concat_list_content to concat_list_path, run concat_shell, then read the final duration. Save one execution result per clip to clip_results.json with actual_duration_seconds; include part_results only for multi-part clips. When fast_cut_shell is empty (the tool set method="encode" because the source is landscape and must be verticalized with a filter), skip straight to accurate_cut_shell — the vertical blur/scale filter requires re-encoding and -c copy is invalid. The tool-generated accurate_cut_args already include -pix_fmt yuv420p, -crf 20, -movflags +faststart, the vertical -vf filter, and -af loudnorm.

8. Build delivery files: Call build_live_clip_manifest(source_video="$VIDEO", tingwu_task_id="$TINGWU_TASK_ID", analysis_title=analysis.title, sentences=analysis.sentences, invalid=invalid.invalid, warnings=plan.warnings, rejected=plan.rejected, clips=plan.clips, clip_results=clip_results) and write:

   • clip_manifest to clip-manifest.json as a JSON array with each clip transcript.
   • transcript_markdown to transcript.md.
   • summary_markdown to summary.md.
   • clip_notes_markdown[].markdown to clip_notes_markdown[].markdown_path.

9. Export to CapCut (optional): For each successful clip in clip_results.json, create a CapCut/JianYing draft using the capcut-draft skill. Each clip becomes its own draft with:

   • Video segment: the clip's MP4 file (type: "video"). The exported MP4 is already 1080×1920, so place it at scale=1.0 (no hard upscale).
   • Subtitles: from the clip's transcript array, applying the Short-Video Captions style from the capcut-draft skill (font_size≈10, bold_width=0.1 for weight, white with #FFD60A for selling points/numbers, 0.12 black stroke, 0.8 shadow, transform.y=-0.78 lower-third, <size=10.0>...</size> content — do NOT use <b> tags, they render literally). Split sentences >8 chars into ≤2 chunks at natural pauses, each its own text segment; time offsets are relative to the clip's start (subtitle_start = (chunk.start - clip.start) × 1,000,000 microseconds).
   • Cover: the clip's smart keyframe cover-NN.jpg (see Smart Cover below), falling back to $DIR/cover.jpg.
   • Canvas: vertical 9:16 (1080×1920) for short videos
   • No effects, transitions, or background music added automatically (audio is already loudness-normalized) Save results to clip-draft-results.json. Skip this step entirely if CapCut draft root directory is not found.

Cutting Commands

Fast cut:

ffmpeg -y -ss "$START" -i "$VIDEO" -t "$DURATION" -c copy "$OUT"

Accurate fallback when copy-mode cuts poorly or fails:

ffmpeg -y -ss "$START" -i "$VIDEO" -t "$DURATION" -c:v libx264 -c:a aac "$OUT"

The templates above are the conceptual minimums. In practice, prefer the accurate_cut_args returned by build_live_clip_plan/build_live_subject_clip_plan — when the source is landscape and target_mode="vertical", the tool emits a full re-encode: the chosen vertical -vf filter (blur-fill by default), -c:v libx264 -preset veryfast -crf 20 -pix_fmt yuv420p, -c:a aac -b:a 128k, -af loudnorm=I=-16:TP=-1.5:LRA=11, and -movflags +faststart. In that case fast_cut_shell is empty and method="encode"; run accurate_cut_shell directly (filtering is incompatible with -c copy). The multi-part concat final pass stays -c copy because every part is already yuv420p.

Single clip example:

START="12.300"
DURATION="38.700"
OUT="$DIR/exports/01-product-proof.mp4"
ffmpeg -y -ss "$START" -i "$VIDEO" -t "$DURATION" -c copy "$OUT"

For automated batches, prefer the fast_cut_shell and accurate_cut_shell fields returned by build_live_clip_plan; do not hand-calculate durations.

Multi-part subject clips use planned part commands and concat fields:

# for each part
ffmpeg -y -ss "$START" -i "$VIDEO" -t "$DURATION" -c:v libx264 -c:a aac "$PART_OUT"

# after writing concat_list_content to concat_list_path
ffmpeg -y -f concat -safe 0 -i "$CONCAT_LIST" -c copy "$OUT"

Duration probe:

ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "$OUT"

Smart Cover

cover.jpg is the global first-frame fallback. For each clip, grab a representative keyframe instead of frame 0: sample near the clip's ~40% point with scene-change detection, falling back to the midpoint:

PIVOT=$(awk 'BEGIN{printf "%.2f", '"$CLIP_START"' + '"$CLIP_DURATION"' * 0.4}')
ffmpeg -y -ss "$PIVOT" -i "$CLIP_PATH" -vf "select=gt(scene\,0.3)" -frames:v 1 -q:v 2 "$DIR/cover-NN.jpg" || \
  ffmpeg -y -ss "$(awk 'BEGIN{printf "%.2f", '"$CLIP_DURATION"' / 2}')" -i "$CLIP_PATH" -frames:v 1 -q:v 2 "$DIR/cover-NN.jpg"

NN is the two-digit clip index. Use cover-NN.jpg as that clip's CapCut draft cover; fall back to cover.jpg if generation fails.

Text Export

For Markdown transcript exports, write the content directly from analysis.sentences:

# {analysis.title}

- [1] 0.00-3.20s 字幕
- [2] 3.20-7.80s 字幕

For clip notes, write the clip_notes_markdown returned by build_live_clip_manifest; do not invent note content by hand.

JSON Shapes

analysis.json uses: chapters, sentences, subjects, segments, invalid, qas, topics, words, silents, keywords, key_sentences, templates, audio_info.

Sentence objects: {"index":1,"start":0.0,"end":3.2,"text":"字幕"}

Segment objects: {"title":"片段标题","description":"说明","thoughts":"剪辑思路","start":1,"end":8}

Invalid objects: {"index":1,"reason":"直播间欢迎语"}

Clip plan objects: {"index":1,"title":"片段标题","sentence_start":1,"sentence_end":8,"start":12.3,"end":51.0,"duration":38.7,"output":"...mp4","fast_cut_shell":"ffmpeg ...","accurate_cut_shell":"ffmpeg ...","parts":[...],"transcript":[...]}

Multi-part subject clip fields: parts, concat_list_path, concat_list_content, concat_shell, concat_args, script_notes.

Clip manifest is a JSON array of delivered clip objects with status, method, output, duration, actual_duration_seconds, transcript, and optional error fields. clip_notes_markdown contains ready-to-write Markdown notes for each clip: {"index":1,"title":"片段","markdown_path":"...md","markdown":"..."}.

Cutting Rules

• Keep transcript index boundaries intact; never invent timestamps.
• Prefer valid sentences after removing invalid indexes.
• Pass head_padding_seconds=0.15, tail_padding_seconds=0.30 to the plan tools so cuts don't clip the first or last word.
• Target short-video sweet spots: pass max_duration_seconds=120 (15-60s is ideal for livestream hooks). Treat clips >120s as too long and record a warning; drop clips shorter than 5s — the MIN is 5s everywhere (draft-skip and red-flag thresholds agree).
• Aim for semantic diversity across clips (distinct sentence ranges or topics) so the batch is not a set of near-duplicates.
• Favor segments whose first sentence is a conclusion, opinion, pain-point, or contrast (a hook); complete_live_subject already forces a 3-5s opening hook for subject clips.
• Verticalize for the target platform: pass target_mode="vertical" and vertical_fill="blur". Landscape sources get a blur-fill re-encode; already-vertical sources get only scale. Audio is loudness-normalized (EBU R128) on re-encode paths, never on -c copy.
• Use -c copy first for speed only when fast_cut_shell is non-empty; retry with re-encoding (or go straight to it) if a filter is needed or timestamp accuracy is poor.
• For subject clips, preserve selected sentence order unless the LLM explicitly provides a better narrative order.
• Do not cut live-only greetings, thanks, countdowns, real-time stock claims, or room-specific promos into short videos unless the user explicitly asks.