Ugc Editing

UGC Editing

Core principle

Editing is footage-source-agnostic. The same assemble → trim → layout → caption pipeline applies whether the clips came from ugc-video-prompts or a real phone shoot. This skill picks up wherever footage already exists and turns it into a finished vertical post. The single biggest editing mistake in UGC is padding — long pauses, repeated stumbles, slow setups. Cut harder than feels comfortable. Authentically raw ≠ unedited; it means tight pacing with natural-sounding joins.

Tool model (current Palmier): timeline positions are project frames; source trims are seconds via source: [start, end]. Linked A/V is folded into the video clip as audio: { id, … } — mute/edit audio by that nested id. Mutations return a timeline delta — patch from it; don't re-get_timeline after every call. Index 0 renders on top.

Step 0: Pre-flight — always run before touching the timeline

1. get_timeline — check resolution / canGenerate. Must be 9:16 (typically 1080×1920). If not:

   set_project_settings({ aspectRatio: "9:16", quality: "1080p" })
   

   Do this before placing clips. Explicitly configured settings stick (the first clip no longer snaps the project to its source aspect — placement may only note a mismatch).

2. get_media — identify A-roll and b-roll. Note durationSeconds, width/height, hasAudio.

3. inspect_media on the A-roll:

   • First: overview: true — storyboard + segment timestamps.
   • Then: wordTimestamps: true (window long clips with startSeconds/endSeconds) — read words as prose before cutting. Look for retakes, fillers, false starts, trailing dead air, weak sections.

Step 1: Place A-roll only — no b-roll yet

add_clips({
  entries: [{
    mediaRef: AROLL_ID,
    startFrame: 0,
    // whole asset, or trim in SOURCE seconds (not project frames):
    source: [SKIP_PREROLL_S, END_S]   // omit for full file
  }]
})
// omit trackIndex on every entry → auto-creates video + linked audio tracks

• Never pass trackIndex here when you want a fresh storyline track. Omitting it auto-creates one shared video track (and linked audio). Specifying an existing index can overwrite whatever is already there.
• There is no durationFrames / trimStartFrame on add_clips — use source: [startSeconds, endSeconds] for source span, or endFrame (mutually exclusive with source) for an exact timeline occupy range.
• 16:9 A-roll in a 9:16 project may letterbox. Cover-crop the talking head to full frame:

  apply_layout({
    layout: "full",
    slots: [{ slot: "main", clipIds: [AROLL_CLIP_ID], anchor: "center" }]
  })
  

  Prefer apply_layout over set_clip_properties transforms — layout sets transform and crop together (transform-only leaves a stale crop and can stretch).

Step 2: Cut bad takes and filler — words first

Prefer remove_silence first for dead air (no transcript), then word cuts.

Read first, cut once per pass. get_transcript → mark indices → one remove_words call. Indices shift after every cut — re-read before the next pass.

get_transcript()
remove_words({
  words: [
    [FIRST_RETAKE_START, FIRST_RETAKE_END],
    FILLER_INDEX,
    [FALSE_START_START, FALSE_START_END],
  ],
  cutAggressiveness: "tight"   // "balanced" / "loose" if joins feel clipped
})
// bulk fillers: remove_words({ matches: ["um", "uh", "hmm"] })

What to cut: duplicate retakes (keep the most fluent), fillers, false starts, trailing soft endings, off-topic tangents that break hook → value → CTA.

What to keep: natural pauses between real thoughts; best take of each section; the outro/CTA.

Captions block ripple: if a caption track exists, remove it before cutting:

manage_tracks({ remove: [{ trackId: CAPTION_TRACK_ID }] })
// or remove: [index] — prefer trackId from get_timeline

Then cut, then re-add captions. Never caption before cutting is finished. There is no remove_tracks tool.

Step 3: Pick a layout format

Format 0 — Straight intercut (default)
B-roll is supplementary, not simultaneous. Hard cuts alternating full-frame A-roll and b-roll on the same video track every 3–5s. Use add_clips (overwrite landing region) or insert_clips (ripple — requires trackIndex) — no layout tool.

Format 1 — Stacked split, b-roll top / talking head bottom
B-roll is proof (product/result/location). Viewer needs both at once. layout: "top_bottom", b-roll → top, head → bottom.

Format 2 — Stacked split, head top / b-roll bottom
B-roll is filler/retention. Same layout, slots flipped.

Format 3 — Full-bleed b-roll, floating head
Not clean via apply_layout alone — needs chroma/apply_effect + manual framing. Skip unless asked.

Step 4: Place b-roll

Format 0 — intercut on the storyline track

add_clips({
  entries: [
    { mediaRef: BROLL_1, startFrame: F0, source: [0, 4] },
    { mediaRef: BROLL_2, startFrame: F1, source: [0, 3] },
  ]
})
// same track as A-roll → overwrites those ranges (split/trim under the hood)

Mute any linked b-roll audio (nested audio.id on each placed video, or mute the b-roll audio track via manage_tracks).

Format 1–2 — stacked split (critical: span-scope the head)

apply_layout on an existing clip is whole-clip (time-invariant). If you layout the entire A-roll into bottom, the head stays half-frame for the whole video — including stretches with no b-roll (black / empty top).

Do this instead:

1. split_clips the A-roll at each b-roll span boundary so overlay segments are their own clips.
2. add_clips each b-roll tile on a new track (omit trackIndex on the whole batch) covering exactly those frames — use startFrame + source or endFrame.
3. Mute b-roll audio immediately:

   set_clip_properties({
     clipIds: [BROLL_AUDIO_NESTED_IDS],  // video.audio.id from timeline/delta
     volume: 0
   })
   

4. For each overlapping pair, apply_layout:

   apply_layout({
     layout: "top_bottom",
     slots: [
       { slot: "top",    clipIds: [BROLL_CLIP_ID], anchorY: 0.4 },
       { slot: "bottom", clipIds: [AROLL_SEGMENT_ID], anchor: "center" }
     ]
   })
   

5. Leave non-overlay A-roll segments full-frame (layout: "full" only if something dirtied them).

Do not use insert_clips for overlay b-roll — it ripples the storyline and breaks sync. Do not build splits with set_clip_properties / set_keyframes transforms.

Optional shortcut for a single continuous stacked section: apply_layout placement mode with mediaRef in every slot + startFrame/endFrame creates stacked tracks for that range only — useful when A-roll isn't already on the timeline for that span. Don't mix mediaRef and clipIds across slots in one call.

Step 5: Captions

add_captions finds spoken audio itself — there is no clipIds filter. Mute all b-roll / non-dialogue audio before captioning or captions will pick up the wrong speech.

add_captions({
  animation: "highlightPop",          // or "wordPop"
  highlightColor: "#FFD700",
  maxWords: 3,
  transform: { centerY: 0.5 },        // 0.5 seam (split); 0.82 lower-third (full-frame)
  style: {
    color: "#FFFFFF",
    bold: true,
    fontSize: 48,                     // ~48 split; 60+ full-frame
    fontCase: "mixed"                 // NOT "uppercase" — natural case
  }
})

• Restyle later: update_text({ captionGroupId, style: {…}, transform: {…}, animation }).
• fontCase is display casing (mixed / uppercase / lowercase). Prefer mixed.
• Caption clips show as captionGroups summaries on get_timeline — use captionDetail: true only when editing individual caption clips.

Step 6: Verify

inspect_timeline({ startFrame: 0, endFrame: totalFrames, maxFrames: 6 })

Check samples for: b-roll filling its half only during overlay spans; talking head full-frame elsewhere; captions at the seam/lower-third; no stretched clips (stale crop — reset with apply_layout full).

Common failure modes

Track order — typical end state

Index 0 = top (what you see on top in the preview).

Pacing benchmarks

• Hook in the first 2–3 seconds.
• A cut, b-roll switch, or caption change roughly every 3–5 seconds.
• Most UGC ads: 15–60s (30–45s sweet spot for Reels with CTA).
• Captions are not optional — most social video is watched muted.
• AI talking heads: keep faces to 2–3s before cutting to b-roll.

Guardrails

• Build fictional, generic personas — don't edit in a specific real, identifiable person without consent, and don't recreate a named public figure's likeness.
• Keep wardrobe/pose/action choices brand-safe and non-sexualizing by default; only go further if the user explicitly asks.