How do you add a new scene to an AI video project using a single instruction?

Write the instruction as a delta command to the invideo agent — something like "insert a 4-second establishing shot between scene 2 and scene 3, matching the existing character, lighting, and tone." Because the project's context, character sheet, and shot timeline are already in memory, that one line triggers a plan update, the new scene generation, and a voiceover-timeline regeneration around it.

Phrase the instruction as a delta — what changes, where it goes, and what stays — not a fresh prompt. The invideo agent is an agentic video tool that holds the project's live brain: brand context, character sheet, locked scenes, voiceover timeline, music bed. A single natural-language line flows through that shared context to all the sub-agents at once, so you don't re-attach references or re-describe the look.

Write the delta clearly. Name the insertion point ("between scenes 2 and 3", "at the top before the current opener", "after the CTA"), the scene's content in one beat, the duration, and the carry-overs ("same character v3, same location, same lighting, same VO tone"). The sharper the brief, the more proactive and specific the next steps from the invideo agent — vague deltas produce vague inserts.

Let the invideo agent return a plan before it generates. Ask for a sitrep — what's locked, what's open, and an estimated generation breakdown (how many images, how many video clips, the credit estimate) for the new scene only. This is the same pre-flight pattern used across documented productions: "it gives me an estimated generation breakdown… so I know what I'm roughly going to spend now before I actually spend it." Confirm, then trigger generation.

Insert at the image stage first, then animate. Have the invideo agent generate the new scene as a still keyframe inside your existing storyboard order, with the locked character sheet and location sheet as references. Lock the frame, then animate that single shot — usually a 5-second clip out of Seedance 2.0 — rather than regenerating neighboring scenes. Spending video credits only on the new locked frame is what keeps a mid-project insert cheap; in one documented production this discipline held per-ad cost around $125.

Let it regenerate the voiceover timeline around the insert. If the project has VO, the same instruction tells the invideo agent to re-slice the VO so existing lines re-align to their original scenes and any new line for the inserted scene lands inside it. Upload the full VO file once and tag which line belongs to the new shot — the invideo agent trims, feeds it into Seedance 2.0 with the character reference, and returns the lip-synced clip without manual audio editing.

Pick the model the insert needs. invideo holds the current generation models — Seedance 2.0, Kling, Veo, Runway — and the invideo agent routes the new shot to the right one without you switching tools. Seedance 2.0 reference-to-video carries character context from your locked sheet across the new clip; Kling 3.0 handles a multi-cut insert (e.g. a 3-beat montage scene) natively in one pass; for an image-first insert, GPT-Image-2 or Nano Banana locks the keyframe before any video credits are spent.

Stitch into the existing edit. Pull the new clip into Slate (or Premiere Pro) in the position the delta specified — same order the invideo agent planned. Concurrent edit assembly works here too: by the time the new scene renders, the surrounding cut is already laid up, and you only drop the insert into the gap.

These are the moving parts of a one-instruction insert — what's in your delta and what's already locked in the project's brain decides how much regenerates.

Watch some of these to see what works for you:

This one change many updates behavior is exactly what it is doing in practice because Agent One keeps a live project brain.

— invideo's creative team, on how a single instruction cascades updates through a project