LLM-assisted editing

Type your intent in chat. The LLM proposes a timeline JSON patch and applies it after you confirm.

How it works

Your chat message goes to the Bom.Ag Worker. The Worker calls the bom-av MCP tool, which sends the current timeline JSON and your intent to the LLM. The LLM returns a JSON patch you preview in a diff panel before clicking Apply.

  • CLIs used: Claude Code (preferred) · Codex, with automatic failover
  • Change the provider order in Settings → CLI priority

Common command patterns

  • "Trim the first 3 seconds" — cut 3 s off the first clip on track 1
  • "Add a 0.5 s fade between every scene" — fadeblack 0.5 s on every clip boundary
  • "Enable BGM ducking with reduction -10 dB" — auto-duck the music track
  • "Add a subtitle 'intro end' at 1:30" — new subtitle clip
  • "Reframe to vertical 9:16 for Shorts" — change project aspect ratio
Tip

Be specific with time: "first 3 seconds" / "from 01:23" / "last 5 seconds".

Preview the diff

After the LLM responds, a diff panel appears below the message.

  • Green (+) — new clips or keyframes
  • Red (−) — items to remove
  • Yellow (~) — properties to change
  • [Apply] applies it; [Discard] dismisses it

Safety guards

  • Destructive operations (clear all clips, delete a whole track) require extra confirmation
  • 50-step Undo for everything
  • Malformed JSON patches are rejected; the original is auto-backed-up
  • Auto-backup every 5 minutes to ~/.bom/timelines/.backup/

Inspect the active model

A model badge appears next to the connection dot in the top toolbar. Click it for token usage and per-call latency.

Prompt-writing tips

  • Be specific with time: "first 3 seconds", "from 01:23", "at the 1:30 mark"
  • Name tracks explicitly: "the BGM track", "the narration track"
  • Specify output format: "vertical 9:16", "1080p", "60 fps"
  • Split big edits into smaller asks — 1 or 2 changes per turn