Agent Skills › Forward-Future/scarlet › dialogue-pass

dialogue-pass

GitHub

用于视频对话剪辑,通过修剪停顿、填充词和冗余片段优化对话流畅度,同时保留说话人语气与节奏。

skills/dialogue-pass/SKILL.md Forward-Future/scarlet

Trigger Scenarios

需要清理录音中的口误或冗余内容 调整对话节奏以增强叙事连贯性

Install

npx skills add Forward-Future/scarlet --skill dialogue-pass -g -y
More Options

Use without installing

npx skills use Forward-Future/scarlet@dialogue-pass

指定 Agent (Claude Code)

npx skills add Forward-Future/scarlet --skill dialogue-pass -a claude-code -g -y

安装 repo 全部 skill

npx skills add Forward-Future/scarlet --all -g -y

预览 repo 内 skill

npx skills add Forward-Future/scarlet --list

SKILL.md

Frontmatter
{
    "name": "dialogue-pass",
    "category": "Story and pacing",
    "description": "Make a first dialogue cut by trimming distracting pauses, fillers, redundant takes and production talk while preserving the speaker’s flow."
}

Dialogue pass

Make the conversation easier to follow with cuts alone. Keep the speaker’s meaning, the order of the exchange and the timing between synchronized tracks. The editor’s brief sets the pace and identifies any passage to protect. Use straight picture and audio cuts. Smoothing, mixing, color, framing, cutaways, music and captions each need an explicit request.

When someone talks to camera, check once per recording that their lips keep time with the sound (lip_sync) before cutting it. Put a recording that is plainly off back in time with timeline_picture_sync, and ask the editor when the measurement leaves you unsure.

An abandoned start can come out when the next take completes the thought. A retake can begin just beyond the passage under review. Compare what each delivery actually says, particularly corrections, qualifications and changes of mind. A line the speaker takes back and says again is a retake, even when something they just noticed corrected it ("oh yeah", "wait, actually"): keep the corrected line, and cut the wrong one with the moment of noticing. Keep both only when the mistake itself matters to the story. Keep repetition used for emphasis and the other speaker’s reply. A direction to the recording crew can come out when the surrounding dialogue makes its purpose clear. One the speaker says at full voice as part of the show, telling the viewer what they’re about to see (“put the other ones on the screen so we can compare them”), plays: keep it and do what it asks.

When a line is barely audible beside the dialogue on either side, it is usually an aside: cut the whole line and the wait around it. Speakers often say a note to their editor this way ("you can get a screenshot of that"), so quote each such line in the edit's summary, and flag it when it asks for something to be done. Keep a deliberate whisper or a quiet reply the audience needs to follow the exchange.

Cut tight. A pause earns its length by doing something: a beat before a reveal, a reaction, a breath between two ideas, time to take in what just appeared on screen. A speaker’s own short pauses are how they talk, so leave them as said. A pause that waits long enough to play as dead air comes down to the beat the delivery around it needs: next to nothing inside a sentence, a little more between sentences, more again where the subject changes. The quiet either side of a join plays as one pause, so judge the air left after the last word and before the next together, as it will be heard. At a cut the picture jumps, and a pause held before it plays as the speaker waiting or breathing on camera: cut where the last kept word's sound ends, before any breath, and bring the next word in right after, about a tenth of a second of quiet in all. Give a longer pause only where it earns one, such as a beat before a reveal. Where Scarlet places a join's edges, leave about 0.05 s of room tone after the last kept word, for its decay, and 0.05 s before speech resumes, so the voice comes in right after the cut. As a guide, not a rule, about a second with nothing happening in it plays as dead air. Every cut is a jump the viewer sees, and cuts that land one after another play as a stutter: let the speech between cuts run as it was said, and take a pause beside another cut out in that cut or leave it. Before finishing, account for every stretch of two seconds or more that still plays without a word: cut it, or keep it for what the viewer sees or hears in it. When nobody is speaking, keep only as long as the viewer needs to see the result: typing, loading, searching and waiting cut to the moment the result appears. A sigh, a cough, a lip smack or a throat-clear between lines goes out with the wait around it, and so does a breath or any other sound a join leaves playing beside the kept words. So does a sound that isn’t the speaker’s, such as a system chime or a remark from someone off-mic, when nothing the audience needs is said over it. Cut an “um” or “uh” when the line flows better without it. A filler can also carry hesitation or hold a speaker’s turn. Use the words, measured sound and picture to decide. A gap in the transcript can contain speech or action.

Apply confident cuts with dialogue_apply_cuts. Mark a meaningful pause or protected phrase as a keep. When removing a passage could change the meaning or lose a reaction, stage the cut for the editor. Use dialogue_review_inspect and dialogue_review_save to record a recommendation that names the words or reaction at stake. Recommend keeping a cut only when its passage belongs in the video. A cut that should go but would clip a word, or leave a piece of one, is yours to repair before anyone hears it: save it again naming words that leave every kept word whole. When a word or a line was said twice, the first delivery is usually the one to take out, unless it is clearly the better performance: cleaner edges don't decide it. In a stumble like “who who thinks of this”, taking out the first “who” tends to play best, since the join lands where the speaker broke off and the second runs on into the rest of the line as said, while cutting the second can jump from the broken word straight into the next. That is a guide, not a rule: use the words, measured sound and picture to choose the join that plays most naturally. Ask the editor only for what is theirs to decide: which delivery they meant, or whether a line belongs. A join you could fix is not a question for them. A cut you decided whose join Scarlet can't place cleanly, with no quiet at an edge or an edge it has to move well off where you put it, waits for the editor to hear it. Don't trade it for a cut with cleaner edges that leaves part of the retake playing: the editor would rather hear a join in running speech than a try left in (“…the video and tell it ‖ and give it as a reference”).

Check the joined dialogue in context. Each speaker’s words should play whole, with questions and answers still making sense together. A clipped word or an accidental restart that the cuts introduced or left behind is not finished work. Finish when the requested dialogue cuts and their review are ready.

When a picture cut returns to a speaker, the brief sets how quickly their voice should enter. timeline_join_audit reports the time from the incoming picture to the first word of that take. Use the sound around its onset to judge a trim, especially for quiet consonants or a word that begins before the picture changes.

Version History

  • 27a27c9 Current 2026-11-21 16:18

    修正重复单词的剪辑逻辑(优先切除首个错误发音),并调整对未完成的取样的剪辑策略,避免留下残留片段。

  • bb0c550 2026-11-21 03:58

    修正了对同义不同次录音的评分逻辑,并明确保留全音量向编辑提出的指令性台词。

  • bbe94a5 2026-11-20 15:35

    新增 lip_sync 和 timeline_picture_sync 工具,用于检测并校正音视频不同步问题。

  • 6e93a9e 2026-11-19 15:04

    明确保留说话人的短停顿,并规定避免连续剪辑导致画面跳跃感。

  • b6bdb58 2026-11-18 20:45

Same Skill Collection

skills/add-captions/SKILL.md
skills/b-roll/SKILL.md
skills/fit-length/SKILL.md
skills/frame-screen/SKILL.md
skills/make-shorts/SKILL.md
skills/matt-talking-head/SKILL.md
skills/motion-graphics/SKILL.md
skills/punch-in/SKILL.md
skills/speed-through-waits/SKILL.md

Metadata

Files
0
Version
27a27c9
Hash
d31a31d5
Indexed
2026-11-18 20:45

inicio - Wiki
Copyright © 2011-2026 iteam. Current version is 2.156.0. UTC+08:00, 2026-11-22 01:09
浙ICP备14020137号-1