layer-video-continuity

Layer Video Continuity

Overview

A single shot holds its own continuity. A sequence does not: every generation re-invents the face, the costume, the light and the set unless something carries them across the cut. This skill is about what carries them. Single-shot motion and camera are layer-video and layer-video-direction; cutting the finished shots together is layer-video-timeline; the storyboard comes from layer-art-direction.

Three things carry continuity, strongest first:

1. The last frame of the previous shot as the first frame of the next.
2. Character elements: reference images or a clip of the character, attached to every shot.
3. One approved still that every shot is drawn from.

If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add layerai/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Need	Route
Shot N+1 starts where shot N ended	Shot N's video file_id in the first_frame slot with video_frame: "last"
A frame at an exact time	video_frame_seconds on the same guidance entry; it overrides video_frame
Same character in every shot	element_frontal_image, element_reference_image or element_video guidance files
Match the motion of an existing clip	reference_video
Models that accept all of the above	list_base_models with filter.guidance_file_types listing every type the sequence will use
What a slot accepts, and its limits	get_base_model on the chosen model
Explore the whole sequence cheaply	draft: true on every shot, complete only the keepers

A guidance entry is {file_id, type} plus optional video_frame, video_frame_seconds and weight. A video placed in an image slot (first_frame, last_frame, reference_image, init_image) contributes one frame: its first frame by default, its last for last_frame. So to start a shot on the previous shot's ending you must say so: video_frame: "last".

Pick a model by what it accepts

Continuity is a capability question before it is a prompt question. Ask list_base_models for filter.use_case: "image_to_video" and filter.guidance_file_types set to every slot the sequence needs, for example ["first_frame", "element_frontal_image"]. Every listed type must be accepted, so an empty result means the combination is too ambitious: drop the least important slot rather than inventing a model. Families such as Seedance, Kling or Veo differ in what they take, which is why the filter, not the family name, decides.

Then get_base_model on the first result. It gives each slot's file and length limits and the family's own prompt guide, which is the right phrasing for that model. Use one model for the whole sequence: two families render the same element differently, and the cut shows it.

Elements hold the character

Attach the same element files to every shot that shows the character, unchanged. Use a frontal image for the face, reference images for the costume from other sides, and an element video where the model accepts one. Produce these once from the approved design (layer-game-assets for a character sheet) and reuse their file_id values: never regenerate an element mid-sequence.

Elements hold identity, not position. The prompt still states where the character is, what they do and which way they face.

Write the sequence as a shot list

From the storyboard, write one line per shot before generating anything:

• One action per shot. A shot that has to show a door opening, a reaction and a sword draw will do none of them well. Split it.
• Screen direction. If the hero exits frame right, they enter the next shot from frame left. State it in both prompts.
• Eyeline and light. Name the light direction once and repeat it verbatim in every shot of the scene.
• Cut type. A continuous action chains on the last frame. A cut to a new angle does not chain: start it from the approved still or the elements, or the new angle inherits the old composition.

Draft the whole sequence first

Where get_base_model reports capabilities.draft, render every shot with draft: true, review the sequence as a cut, then complete only the shots you keep. A completion is a second execute_forge with base_model_id set to the draft's draft.completion_base_model_id, batch_size: 1, and the draft's file_id as a guidance file of type draft.completion_guidance_type, with no prompt: it keeps the draft's prompt, motion and length. Estimate a completion with the completion model and the draft's duration_seconds. Drafts can expire (draft.status, expires_at), so complete promptly.

Chaining drafts previews the cut cheaply, but a completion re-renders the whole shot, so its last frame can drift from the draft's and the join shifts. When a join must be exact, complete shot N first and generate shot N+1 from the completed take; say this costs a second run of each chained shot. Set length with duration_seconds.

Worked example

"A 4-shot cutscene: our knight walks to the gate, pushes it open, a dragon looks up, the knight draws her sword."

1. The approved knight design and front view already have file_id values. No upload.
2. list_base_models with filter.use_case: "image_to_video" and filter.guidance_file_types: ["first_frame", "element_frontal_image", "element_reference_image"]. Take the first result; get_base_model for slot limits, durations, capabilities.draft and the prompt guide.
3. Shot list: (1) wide, knight walks left to right toward the gate; (2) continuous, she pushes it open, chained; (3) new angle, the dragon lifts its head, from an approved dragon still, no knight elements; (4) medium, knight draws sword, facing right, elements attached.
4. Price the job: one estimate per kind (a shot with elements, a chained shot, the dragon shot), then the four completions at the drafts' durations. Present the total; above 20 CUs wait for an explicit yes.
5. Draft all four, each with the knight elements as {file_id, type: "element_frontal_image"} and {file_id, type: "element_reference_image"}. Shot 2 adds {file_id: <shot 1 video>, type: "first_frame", video_frame: "last"}.
6. Poll each at poll_interval_seconds. Review the four as a sequence, redraft any shot that breaks screen direction, then complete the keepers and re-chain shot 2 from the completed shot 1.
7. Hand the completed file_id values to layer-video-timeline for the cut.

Common mistakes

• Putting the previous clip in first_frame without video_frame: "last", so the shot restarts at the beginning of the last one.
• Choosing the model by family name instead of by guidance_file_types.
• Changing models partway through a sequence.
• Regenerating the character's element images between shots.
• Chaining across a cut to a new angle, so the new shot inherits the old composition.
• Packing several actions into one shot.
• Paying for full-quality takes before the sequence has been reviewed as drafts.
• Chaining from a draft when the join must be exact.