sidhantpande/AI-Video-Editor

Tighten interviews, podcasts, tutorials, and talking-head edits in Adobe Premiere Pro by detecting silence, preserving natural speech margins, keeping an untouched recovery sequence, and verifying sync. Use for pause removal, dialogue cleanup, jump-cut planning, or rough-cut acceleration through the Premiere Pro MCP server.

AI-Video-Editor 是什么？

AI-Video-Editor is a Claude Code agent skill that tighten interviews, podcasts, tutorials, and talking-head edits in Adobe Premiere Pro by detecting silence, preserving natural speech margins, keeping an untouched recovery sequence, and verifying sync. Use for pause removal, dialogue cleanup, jump-cut planning, or rough-cut acceleration through the Premiere Pro MCP server.

兼容平台~Claude Code~Codex CLI~Cursor

npx skills add https://github.com/sidhantpande/AI-Video-Editor/tree/HEAD/skills/premiere-dialogue-cut

文档

Premiere Dialogue Cut

Build an evidence-backed cut plan before touching the timeline. Favor natural pacing over maximum compression.

Required server profile

Start or restart the MCP server with MCP_TOOL_PROFILE=dialogue. The default standard profile also contains this workflow, but dialogue is the smaller dedicated surface.

Workflow

Call premiere_ping, premiere_get_project, and premiere_get_timeline. Stop if the host, project, or target sequence is unavailable.
Resolve the active sequence index with premiere_get_sequence_list. Before duplication, capture the source settings and a canonical timeline inventory: track counts, clip order, source-media paths, source in/out ranges, timeline start/end ranges, speed, and duration. Call premiere_duplicate_sequence, re-read the list to identify exactly one new copy, and activate it with premiere_set_active_sequence. Re-read the same canonical fields and require them to match while ignoring the clone name and sequence/clip identities. Do not compare sequence hashes across distinct sequences: the hash intentionally includes sequence and clip identity, so a valid clone has a different hash. Leave the source sequence untouched as the recovery boundary. premiere_snapshot_timeline and premiere_save_sequence_version may record before/after audit data, but they cannot restore a whole sequence.
List the dialogue clips with premiere_get_clips_on_track, then call premiere_get_clip_info to resolve each source-media path, source in/out range, speed, and timeline range.
Call premiere_analyze_waveform once per unique source path. Use its decoded-audio silence_regions; the legacy premiere_detect_silence host command only returns clip metadata and is not silence evidence. Stop if waveform analysis is unavailable or empty for a clip that should contain audio.
Intersect source silence with each used source range and map it to timeline time. Stop on reversed, retimed, or variable-frame-rate media unless that mapping can be verified. Keep 80–150 ms before speech, 120–250 ms after speech, and preserve longer rhetorical pauses unless the user asks for an aggressive cut.
Show a proposed plan with edit count, removed duration, shortest retained pause, affected tracks, and protected markers before applying changes.
Apply approved edits in reverse timeline order with premiere_razor_clip, premiere_remove_clip_from_track, and premiere_ripple_delete_gap, or use premiere_ripple_trim when it expresses the same edit safely.
Re-read the timeline and sequence hash. Verify duration change, A/V sync, clip ordering, and that no edit overlaps protected markers.

Guardrails

Preserve linked audio/video and room tone unless explicitly told otherwise.
Do not treat premiere_detect_silence as analysis, and do not call premiere_smart_cut, premiere_smart_trim, or premiere_generate_rough_cut when the server reports them as unsupported.
Stop on the first failed edit; report the last verified time range and the untouched recovery-sequence name.
Add short audio crossfades only after cuts are verified, using the readback-verified premiere_add_audio_transition route.