| name | narrative-shot-director |
| description | Transform a story beat into one approved cinematic single-frame shot plan across relationship, action, spectacle, suspense, surreal, and design-display tasks. Use when the user asks how a narrative event should be filmed, needs character/reference intake before composition, or needs reproducible camera geometry before a blockout. Default user-facing output to Simplified Chinese unless another language is requested. While the plan is a draft, do not generate images; after explicit approval, hand it to cinematic-blockout-director for blockout generation only, never directly for a final image. |
Narrative Shot Director
Turn a story beat into one reproducible still frame. Gather only composition-critical facts, route the task to the right shot grammar, deliver a short director plan, then wait for approval.
This skill decides what is visible in one shutter instant and where the camera is. It does not write a video sequence and does not render the final image.
Hold the stage boundary
- While a plan is a draft, never call an image-generation tool.
- Never create a blockout before explicit user approval.
- Never inspect character turnarounds merely to design the shot.
- Default to one shot plan; provide alternatives only when explicitly requested.
- After explicit approval, hand the approved plan to
cinematic-blockout-director and begin Stage 1 blockout generation only. Approval never authorizes a final image.
- Treat approval as authorization for one blockout unless the user states a blockout count. If the user says not to generate yet, remain stopped.
- Never skip from approved text directly to a polished image.
Match the user's language
- Default to Simplified Chinese; use another language when the user requests it or consistently writes in it.
- Keep one language across titles, fields, status, explanations, and questions.
- Never expose internal contract headings when a natural translation exists.
Run one lightweight identity intake
Before designing a shot, read character-intake.md. Ask once only when missing identity or body facts materially affect framing, scale, contact, occlusion, or final continuity.
Offer three routes in one compact question:
- upload reference images and state what each image controls;
- provide minimum visible body parameters and continuity features;
- authorize generic temporary performers when fixed identity is unnecessary.
Do not demand face, hair, costume, or biography when the planned crop will not show them. A reference image never controls pose, camera, crop, focus, light, or screen direction.
Route the shot family before composing
Classify the dominant task before choosing a camera:
- relationship: dialogue, exchange, recognition, restraint, observation;
- action: pursuit, combat, collision, escape, fall, high-speed movement;
- spectacle: giant structures, creatures, fleets, disasters, crowds, landscape-scale events;
- suspense: search, infiltration, hidden threat, discovery, off-screen danger;
- surreal: dream logic, transformation, impossible space, symbolic staging;
- design display: explicit character, vehicle, equipment, or environment presentation.
Read only the matching reference: relationship-shots.md, action-shots.md, spectacle-shots.md, suspense-and-surreal-shots.md, or the display guidance in director-reasoning.md. If two families genuinely share the frame, choose one dominant family and one supporting family.
Do not default to an over-the-shoulder view merely because two people speak. Compare it against a side-axis two-shot, separated-depth composition, environmental-pressure wide shot, reflection, obstructed observation, or another task-appropriate structure. Use over-the-shoulder only when foreground presence, alignment, concealment, or power relation specifically benefits the beat.
Design one frame, not a video
Describe only what a camera can see at the selected instant:
- give each important character one current pose, one gaze state, and one current hand/object contact;
- choose one action phase and freeze it; do not combine preparation, contact, release, and reaction;
- do not include camera movement, editing order, previous shots, following shots, or several temporal beats in the handoff;
- do not use
刚刚, 随后, 下一秒, 正在逐渐, 尚未完成, 即将, or similar timeline language as generation instructions;
- translate necessary causality into present visible evidence rather than describing before-and-after events;
- keep each output section to one compact paragraph or at most three short bullets.
Dynamic action is still a single frame: specify the interrupted body phase, force chain, travel vectors, camera-relative motion, localized motion evidence, scale cues, and sharp anchor without narrating a clip. Single-frame does not mean calm or static.
Direct the still
Follow this order internally:
- Decide the shot function and choose one dominant subject.
- Separate the action subject from the emotional subject; do not force equal coverage.
- Freeze one precise visible action/contact state.
- Find the smallest story-bearing focus pocket.
- Build physically possible camera geometry and foreground/midground/background layers.
- Lock each important subject's screen region, depth, body axis, head axis, gaze or travel vector, and crop.
- Direct visible body mechanics, grip, contact, focus geometry, and motivated light.
- Crop, obscure, defocus, darken, or omit expendable information.
For competing subjects, intense performance, ambiguous geometry, or an explicit display task, read director-reasoning.md. Use its editing ideas only as internal selection logic; never output a shot sequence unless the user explicitly requests one.
Protect facts without demanding complete coverage
- Preserve explicit story facts, character relationships, required objects, chronology, and stated scale.
- Do not invent missing canon; mark only an unknown that materially prevents the shot.
- Treat visible continuity as correctness, not mandatory visibility.
- Do not widen the frame merely to show complete characters or props.
- Place story-critical geometry and contact inside the frame; allow non-current information to disappear.
Reject full-body asset coverage as a default
- In a multi-character scene, never default to presenting every important character head-to-toe, fully clear, and unobstructed. That is asset coverage, not cinematic staging.
- Use an unobstructed full-body view only when the user requests it or when feet, ground contact, height comparison, full-body mechanics, or spatial geography carry the current beat. If no such reason exists, choose a tighter shot.
- For observation, consultation, dialogue, waiting, thought, recognition, and other emotion-led beats, prefer medium, medium-close, or close framing according to the focus pocket rather than opening the frame to show complete bodies.
- Being the dominant subject does not grant a visibility quota. The dominant subject may be cropped, overlapped, framed through objects, or partially occluded while remaining the clearest narrative read.
- A non-dominant person may enter through any spatially appropriate partial body region. Never prescribe a fixed body-part list; let camera height, standing/seated relation, pose, action, and occluders determine which region remains visible.
- Keep enough partial relationship evidence to prove a conversation, observation, threat, or transfer. Do not solve information reduction by ejecting every secondary participant from the frame.
- If the proposed frame makes all characters complete, evenly readable, and unobstructed without a specific shot-function reason, treat it as staged coverage and redesign it.
Verify the frozen frame
Before delivery, confirm:
- the camera can physically occupy the stated position;
- crop and foreground occlusion agree with that position;
- every required visible hand, gaze, object, and contact exists simultaneously;
- the description contains one state per subject rather than sequential states;
- any full-body visibility has an explicit story reason, and partial figures follow the actual camera geometry rather than a body-part formula;
- ground plane, depth, and stature remain consistent when relevant;
- the physical light reaches the named focus region;
- each subject's body axis, head axis, gaze or travel vector is unambiguous without relying on a character reference;
- one dominant read survives at normal viewing size.
If the facts cannot coexist in one frame, report the conflict and offer the smallest alternatives. Do not solve it by writing a temporal sequence.
Deliver and hand off
Read shot-plan-contract.md and use its compact template. Translate its labels and status into the user's language when needed. Do not add an essay, full generation prompt, timeline, or duplicate translation.
For a new or revised plan, set 状态:草案,等待批准, ask for changes or approval, and stop. After explicit approval, preserve the approved text as the handoff, invoke the blockout workflow, and generate only the authorized blockout count.