ajiang-koubo-cut

Ajian Talking-Head Tight Cut

Cut the raw recording into a natural, coherent talking-head video with complete information. By default, directly complete one main version and deliver same-named .mp4 and .srt; keep the original footage. When the input is clear, handle it independently, do not repeatedly confirm, and do not require the user to mark re-records in advance.

Judge by only these rules

• The recording determines the final cut; the script helps understanding. For the same meaning, choose the take that is correct, complete and smooth in information, not mechanically the last take. Keep complete and relevant improvised content; do not pad duration to match the script, and do not add sentences or numbers that are not in the recording.
• First choose the complete take, then fix breath pauses. Delete re-records, unfinished sentence starts, short repetitions and waits with no information, while keeping the original content order. Natural emphasis, restatements with different meanings and short tonal pauses do not count as re-records. Cross-take splicing is only for deduplication or completing expressions and must not change the original meaning.
• SRT and ASR are clues. ASR may automatically swallow stutters such as "你看三阶你看三阶". Repeated sentence starts, number corrections, subtitle conflicts and intra-segment seams need to be aligned against the original audio, waveform or short clips to preserve word onsets and endings.
• Subtitles follow the retained recording. Correct proper nouns and recognition errors; subtitles spanning deleted regions must first be split or rewritten into the text actually retained, then mapped in time. Do not bring an entire discarded subtitle back into the final cut just because only tens of milliseconds of overlap remain.

Execution

Read execution interface, use this skill's two scripts, and avoid rewriting the editing program each time.

1. Identify the video, Jianying SRT and verbatim transcript; save pasted scripts to the project working directory. Check the audio track, frame rate and subtitle timeline, and spot-check the correspondence at the beginning, middle and end. Use the original video audio track; externally exported audio replaces it only when alignment has been verified.
2. Read the full SRT and script; when subtitles are inaccurate or there are many re-records, do one independent transcription of the original audio and cache it. Build retained segments for plan.json and corrected source-timeline subtitles for source.json . Only add local re-checks when boundaries are doubtful; do not repeatedly transcribe the whole video.
3. prepare Create a copy and compile one shared timeline; decide whether to shorten long breath pauses based on speech rate and noise. Thresholds are only a starting point and cannot prove that no speech is present. First check the cut audio/subtitles/seams, and export a lightweight preview when the rhythm needs to be seen; export formally once the plan is stable, and re-render after content changes.
4. render Keep the original video's resolution and frame rate.verify --full-decode Check complete decoding, frame count, audio and video duration and subtitle validity; spot-check actual content seams and audio-video correspondence. Fix the corresponding segment when a problem is found, rather than rebuilding the whole process.

By default, keep the original audio and picture, and do not add music, packaging or burned-in subtitles on your own. Use loudness processing only when volume affects viewing. After completion, provide MP4/SRT links, before and after durations and necessary limitations. If only transcription/waveform re-check was done, describe it truthfully and do not claim a complete listening review.