| name | lazyreel-video-editor |
| description | Assemble and cut UGC into a finished video that wins, at any length. Takes the generated clips (any number, from a few for a short ad to dozens for a 60-90 second-plus video) from a video-gen MCP (Higgsfield, Seedance) or raw UGC footage and edits them into a fast-cut, sound-off-legible, hook-first 9:16 piece that obeys the validated breakout laws. Real FFmpeg and Remotion, not hand-waving. Use when the user has clips to stitch, wants to cut footage into a short or a long video, asks to "edit the video," "stitch these clips," "add captions," "make it 9:16," or hands off rendered clips from the director skills. Pairs with the LazyReel MCP (breakout_laws for the cut rhythm) and the director skills upstream. |
LazyReel video editor
The cut is half the video. A video-gen MCP gives you clips; the edit is what makes them a breakout. The single biggest mistake is shipping one continuous clip. Real video that travels is many clips pieced together with hard cuts, because per-frame novelty is one of the two strongest things we measured (see references/cut-rhythm.md). The clip count follows the length: a 12-15s ad is a handful, a 60-90s video is dozens. The cut rhythm (a cut every ~1.5 to 3s) is the same at any length; a longer video is simply more clips on the timeline. Your job is to assemble those clips into a piece whose first 3 seconds earn the watch and whose cut rhythm holds it the whole way.
This is editing, not generation. The value here is compression and assembly: take what was generated or shot and turn it into the version the feed rewards.
When to activate
- The user has 2 or more clips to stitch into one video
- A director skill (Higgsfield or Seedance) just rendered a multi-clip set
- The user wants to cut raw UGC footage into a short
- "edit the video", "stitch these clips", "add captions", "make it 9:16", "cut this down"
- Reframing for TikTok, Reels, or Shorts
Core rule: never ship one clip
A video is many clips, each a distinct shot or angle, cut hard, however many the length needs (a 10-15s ad is ~3 to 5, a 60s video is 12 to 18+). One static clip dies on per-frame novelty (law 3) and reads as AI. If the director only produced one clip, send it back for the full sequence, or cut the one clip into multiple framings (push-in, reframe, macro insert) so the edit still changes every few seconds.
The cut plan, before you touch FFmpeg:
- Clip 1 (0 to 3s): the hook. The strongest, most unresolved shot. Opens a loop. No title card, no format label.
- Middle clips: escalation. Each a new framing or beat, as many as the script has. Product enters here as a helper, not the star. Cut every 1.5 to 3 seconds.
- Final clip: the payoff the hook teased, then a hard end. No drawn-out CTA.
For a long video the plan is the same with more middle clips; keep the first third dense, and let the energy arc (not the length) decide where a beat can hold a little longer. The concat list below just grows, the pipeline does not change.
The pipeline
generated clips (Higgsfield / Seedance) OR raw UGC footage
-> plan the cut (the brief + breakout_laws laws)
-> FFmpeg: trim, reframe to 9:16, concat, normalize audio (deterministic)
-> captions + overlays (burned-in or Remotion) (sound-off survival)
-> audio (native, or ElevenLabs VO if scripted)
-> polish (CapCut / Descript) only if needed
Do not try to make one tool do everything. FFmpeg does the deterministic cutting and assembly. Remotion does programmable overlays and captions. A human-facing editor (CapCut, Descript) is only for final taste, and often you do not need it.
Layer 1: plan the cut
Pull the cut rhythm from the data, do not guess. If the LazyReel MCP is connected, call breakout_laws for the first-3-seconds laws and the per-frame-novelty finding, and use the brief's beat structure for clip order. Write a tiny edit decision list: which clip is the hook, the order, the cut points, where the product enters, what the on-screen caption says in the first second.
Layer 2: deterministic cuts (FFmpeg)
Trim each clip to its beat
ffmpeg -i clip_01.mp4 -ss 0 -to 2.8 -c:v libx264 -crf 18 -c:a aac trimmed_01.mp4
Reframe anything to vertical 9:16 (crop, do not letterbox)
ffmpeg -i clip.mp4 -vf "crop='min(iw,ih*9/16)':'min(ih,iw*16/9)',scale=1080:1920:flags=lanczos" -c:a copy vertical.mp4
Concatenate the clips into the cut (hard cuts)
for f in trimmed_01.mp4 trimmed_02.mp4 trimmed_03.mp4 trimmed_04.mp4; do echo "file '$f'"; done > cut.txt
ffmpeg -f concat -safe 0 -i cut.txt -c copy assembled.mp4
(Re-encode instead of -c copy if the clips differ in codec or resolution: drop -c copy and add -c:v libx264 -crf 18 -c:a aac.)
Normalize loudness so it is not jarring on a muted-then-unmuted scroll
ffmpeg -i assembled.mp4 -af loudnorm=I=-14:TP=-1.5:LRA=11 -c:v copy normalized.mp4
Burn sound-off captions (most viewers start muted)
ffmpeg -i normalized.mp4 -vf "subtitles=captions.srt:force_style='Fontname=Inter,Fontsize=16,Bold=1,PrimaryColour=&H00FFFFFF,BorderStyle=1,Outline=2,Alignment=2,MarginV=120'" -c:a copy final.mp4
The first caption must carry the hook on its own. Keep it high-contrast and inside the safe area (above the platform UI).
Layer 3: programmable overlays (Remotion, optional)
Use Remotion when burned-in subtitles are not enough: animated text reveals, a breakout-counter style number tick, lower thirds, or a product end-card. Compose clips as Sequences on a timeline, keep each Sequence short to preserve the cut rhythm. Skip this layer entirely if FFmpeg captions are enough; do not add motion graphics that slow the cut.
Layer 4: audio
- Keep native clip audio when it sells realism (the silence and room tone are trust signals).
- If the brief is scripted, generate the VO with ElevenLabs and lay it under the cut, then re-run loudnorm.
- No music bed unless the format calls for it. A trending sound can be added in CapCut at the very end.
Worked example
Input: 4 Higgsfield clips for a matcha kit, brief says PAS, testimonial angle.
Cut plan:
- Clip 1 (0 to 2.8s): macro of clumpy matcha mid-whisk, no caption yet. Hook, unresolved. Caption at 0.5s: "still clumpy at 2pm?"
- Clip 2 (2.8 to 6s): hand reaches for the product, new angle. Product enters as helper.
- Clip 3 (6 to 10s): smooth pour, macro. The switch.
- Clip 4 (10 to 14s): the finished cup, one sip, knowing look. Payoff. Hard end.
Commands: trim each to its beat, crop all to 1080x1920, concat in order, loudnorm, burn the caption track. Result: a 14s, 4-cut, sound-off-legible 9:16 ad. Four clips pieced together, never one static shot.
ffmpeg -i c1.mp4 -ss 0 -to 2.8 -c:v libx264 -crf 18 -c:a aac t1.mp4
ffmpeg -i c2.mp4 -ss 0 -to 3.2 -c:v libx264 -crf 18 -c:a aac t2.mp4
ffmpeg -i c3.mp4 -ss 0 -to 4.0 -c:v libx264 -crf 18 -c:a aac t3.mp4
ffmpeg -i c4.mp4 -ss 0 -to 4.0 -c:v libx264 -crf 18 -c:a aac t4.mp4
for f in t1 t2 t3 t4; do echo "file '$f.mp4'"; done > cut.txt
ffmpeg -f concat -safe 0 -i cut.txt -c:v libx264 -crf 18 -c:a aac assembled.mp4
ffmpeg -i assembled.mp4 -af loudnorm=I=-14:TP=-1.5:LRA=11 -c:v copy normalized.mp4
ffmpeg -i normalized.mp4 -vf "subtitles=captions.srt:force_style='Fontsize=16,Bold=1,Outline=2,Alignment=2,MarginV=120'" -c:a copy final.mp4
Quality checklist (the cut-level breakout gate)
References
references/cut-rhythm.md: the measured cut rhythm, the sound-off rule, the 9:16 safe area, and how each breakout law maps to a concrete edit decision. Read it before you plan a cut.- Upstream: the
lazyreel-higgsfield-director and lazyreel-ugc-ad-director skills produce the multi-clip prompts this skill assembles.
- Live insight: the LazyReel MCP
breakout_laws tool serves the first-3-seconds laws.
