Build the Cut First: A Continuity Workflow for AI Film

A hand closes around a brass key. Cut to a doorway. The same hand opens it.

For this scene, judge the two shots at their join. Does the key stay in the same hand? Is the character approaching the correct door? Does the second shot continue the action, or restart it?

Here is a practical AI filmmaking workflow built around those questions: write the cut before generating the clips. Give every shot an entry state, an exit state, and a reason to exist. Then review the connections as carefully as the images.

Let's take a simple example: a caretaker finds a key, approaches a locked greenhouse, and decides to enter.
Use it to plan your own sequence without confusing a convincing demonstration with guaranteed model performance.

1. Write what changes, not just what appears

Start with three story beats:

- The caretaker notices the key.
- The caretaker reaches the greenhouse.
- The caretaker commits to opening the door.

Now assign each beat a visible action. “A mysterious greenhouse” describes a setting. “The caretaker stops with the key halfway to the lock” gives you something to direct.

For this exercise, use five shots: discovery, pickup, approach, hesitation, entry. Resist adding a sixth until you can explain what the viewer learns from it.

Put these stills in order before animating them. Set rough durations in an edit and watch the sequence without music. If the decision to enter is unclear, revise the board.
Do not ask another generation to solve a missing story beat.

cutipAI’s published storyboard description supports planning scenes, shots, and frames in one board. That makes it a useful place to organize this exercise; the method does not depend on a particular generator.

2. Give each shot a continuity contract

A continuity contract is a small production note, not a special software feature. Record what must be true when a shot begins and what must be true when it ends.

Discovery
• Entry state: Key on workbench; hand outside frame
• Exit state: Hand enters beside key
• Must remain fixed: Key shape, bench surface

Pickup
• Entry state: Right hand beside key
• Exit state: Key lifted in right hand
• Must remain fixed: Sleeve, grip, key teeth

Approach
• Entry state: Caretaker moving toward greenhouse
• Exit state: Caretaker stops at door
• Must remain fixed: Travel direction, wardrobe

Hesitation
• Entry state: Key near lock
• Exit state: Hand pauses before contact
• Must remain fixed: Right hand, lock position

Entry
• Entry state: Key already turned
• Exit state: Door begins opening inward
• Must remain fixed: Hinges, light direction

Notice the last row. This version deliberately skips the turn inside the lock. That is an editorial choice for the exercise: the viewer can infer it between hesitation and entry.

If the turn itself matters to your story, give it a dedicated shot and acceptance criteria. Do not quietly remove a necessary action because it is difficult to generate.

A contract also separates hard requirements from preferences. “Right hand holds the key” is mandatory here. “A slightly warmer highlight on the knuckles” is optional. Review them in that order.

3. Lock the references that carry the scene

Prepare a small reference set: the caretaker, the greenhouse exterior, and the key. Add a close view of the door if the lock and hinges are important.

For the caretaker, specify a plain charcoal work coat, rolled right cuff, and unadorned hands. For the key, choose one recognizable feature: three teeth and an oval bow. For the greenhouse, fix the door’s location and hinge side.

These are proposed art-direction choices, not claims about what a model will reproduce.

Keep the approved references separate from attractive but rejected versions. Name them clearly enough that tomorrow’s session can identify the intended source without opening every file.

cutipAI describes reusable characters, spaces, and objects that can be referenced across shots. Its character-consistency page also explicitly says references do not guarantee an identical face. Treat the shared source as an aid to direction, not permission to skip review. [2]

4. Separate the frame from the movement

Approve the starting composition before writing a complex motion prompt.

The input image establishes composition, subject matter, lighting, and style; the text prompt primarily describes motion. It also warns that visual defects in an input image can become more pronounced in video.

For the pickup, inspect the still: is the right hand visible, is the key’s silhouette readable, and can the fingers plausibly close around it?

Then try a focused motion instruction:

> Locked camera. The right hand closes around the key and lifts it slowly from the workbench. The wrist settles with the key suspended above the surface.

This is a proposed prompt, not a tested result. Adjust its syntax to the selected model’s guidance.

We recommend simple prompts, positive phrasing, and adding one element at a time during iteration. Apply that advice within its stated model context; do not assume every generator handles instructions identically.

5. Use endpoint controls selectively

For some shots, the ending pose is as important as the opening frame. Hesitation needs to finish with the key near the lock, not with the door already open.

Google documents first-and-last-frame video generation with Veo. That is a verified capability of the documented interfaces, not evidence that every model or every studio exposes it.

Where your selected tool supports endpoints, design both images so they describe a plausible movement. Avoid changing the hand, key, lighting, and camera position all at once unless that transformation is intentional.

Where endpoints are unavailable, write the intended finishing action clearly and inspect the output for a usable exit. You may need to trim earlier, revise the starting frame, or redesign the cut.

cutipAI’s multi-model page says available generation options are shown in the product and change over time. Check the actual shot settings rather than assuming a named model or endpoint mode is present.

6. Review the join before polishing the clip

Place the pickup beside the approach. Watch their boundary repeatedly at normal speed, then inspect it frame by frame.

Use three questions:

1. Action: Does the next shot continue the intended movement?
2. Identity: Are the hand, sleeve, key, and character still recognizably the approved subjects?
3. Space: Can the viewer understand where the character is going?

For this exercise, accept the pickup only if the key remains legible through the chosen trim and the hand has a usable finishing pose. A beautiful middle section does not compensate for an unusable exit.

Log the failure precisely: “key changes silhouette during lift,” not “looks wrong.” Change one relevant variable on the next attempt. Keep the rejected version and note why it failed so you do not accidentally restore it later.

If the join still fails, consider a motivated insert: the greenhouse latch, seen from the caretaker’s position. Use it to show new information, not merely to conceal a broken action.

Your next production starts at the boundary

Before generating a full sequence, write one pair of adjacent shots and their continuity contracts. Approve their references, generate the necessary motion, and see whether the pair tells the intended story.

Then expand the board. Keep the same discipline as the scene grows: a shot enters with known conditions and leaves with something the next shot can use.

The creative control worth protecting is your ability to reject an impressive clip because it belongs to a different film.