Video Shot
Each shot is the core work unit of video generation and editing. This skill owns video/shots/<shot-id>/SHOT.md but does not own model-specific prompt compilation or generation review conclusions.
Before expanding a set of shots from SHOTS.md, determine whether the scene needs a few observation points and whether establishing shots / reactions / inserts / product close-ups are genuinely useful, following references/planning/COVERAGE.md; do not mechanically apply a standard coverage package.
1. Establish Shot ID and Directory
By default use gap numbering, e.g., SC01_SH010, SC01_SH020; for simple projects without scene numbers, use SH010. When inserting shots in between, use legible intermediate numbers without renumbering the entire existing identity.
For complex projects, only add organizational directories (e.g., chapters) inside video/shots/ when flat layout has impaired browsing.
When a shot involves complex cinematography, action, cross-world transitions, or continuity, read references/direction/DIRECTION.md. When specific shot size, focal length/perspective, camera position, push/zoom, depth of field, focus, or product photography is needed, read references/direction/CAMERA-LANGUAGE.md first to define the visible photographic effect, then let the model adapter determine specific terminology. When multiple people touch, hand over objects, hand operations, product use, complex occlusions, or mechanical interactions are involved, read references/direction/INTERACTION.md to clarify the active action character, hands, contact points, and before/after states of the action. For multi-shot projects requiring validation of sequencing, dialogue, total duration, or product readability before expensive generation, references/planning/PREVIS.md allows lightweight animatics using storyboards and temporary audio; not every project requires this. For whether a shot must wait for the actual output of the previous shot, which shots can be parallelized, and how planned/actual handoffs work, see references/continuity/CONTINUITY-HANDOFF.md. When multi-character blocking, product/mechanical interaction, complex camera paths, first/last frame control, or spatial relationships between adjacent shots remain significantly uncertain before formal generation, read references/planning/PREVIS.md to perform minimal necessary previsualization/storyboarding.
2. Write SHOT.md
Complex shots or those to be externally generated can be written directly following references/planning/SHOT-TEMPLATE.md; simple shots keep only truly needed fields. When VIDEO.md explicitly requires multiple frame-aspect versions, read references/planning/MULTI-FORMAT.md to determine whether they can be cropped from the main version or require independent composition/generation.
SHOT.md only stores the current valid shot design, typically including:
- Shot purpose;
- Target duration (from current script / total duration budget, but still subject to the shot's actual action and dialogue);
- Shot size, camera position, camera movement, and necessary focal length information;
- Subject positioning, action, performance, and expression;
- Environment and key props;
- Dialogue, ambient sound, and shot-specific sound;
- Continuity In / Continuity Out;
- Assets currently referenced for reuse;
- Hard constraints;
- Generation notes allowing adjustments by the model adaptation layer.
Not every shot needs all fields. Keep simple shots concise.
3. Continuity
Continuity records only information truly relevant to adjacent shots, e.g., character position, costume, handheld objects, prop state, light direction, motion direction, and actions already performed.
If the actual generated output of the previous shot has acceptable deviations, the next shot continues from the actual visible state; if deviations violate character identity, product's real structure, or other hard constraints, that generated output must not be used as continuity reference.
4. Cinematography and Action
Camera movement must serve information, relationship, space, or emotion, not be mechanical movement for "cinematic feel." Video model action design should prioritize second-level temporal relationships; do not write frame-level control that can be accurately achieved only in post-production as hard model constraints.
A shot can have complex internal action, but must be clearly explainable why it is a single shot in the edit. When a model generates multiple cuts in a single pass, the project still organizes by final shot semantics; do not permanently merge multiple shots into one object because of model capability.
Boundaries for cropping, color grading, sound replacement, multi-aspect derivatives, and genuinely new shots of the same shot follow references/continuity/SHOT-DERIVATIVES.md; post-production derivatives must not overwrite takeNN.mp4 in the shot directory.
5. Pre-Generation Checklist
Before preparing a shot for video-generation, confirm:
- Shot purpose is clear;
- Target duration is compatible with the amount of action;
- Reusable designs (character, scene, product, etc.) are determined;
- Continuity In / Out are sufficiently clear for adjacent shots;
- Reference asset responsibilities are clear;
- Hard constraints and ordinary preferences have been distinguished;
- Subtitles, logos, precise cut points, or other obvious post-production tasks are not mistakenly written as generation hard constraints.
If not satisfied, fix SHOT.md or upstream design first; do not substitute shot design with a "generate and see" approach.
6. Prompt and Generation Result Ownership
Long-term prompt versions and generation results live in the same directory as the shot:
prompt_v01.md
prompt_v02.md
take01.mp4
take02.mp4
Model-specific prompt compilation and one-shot generation packages are owned by video-generation; result selection, regeneration decisions, and post-production fixes are owned by video-review.
Completion Standard
SHOT.md is sufficient for the generation skill to compile prompts without re-guessing the director's intent, and clearly defines what the shot must inherit and what it leaves behind.