name: image-impACT-frame description: This Skill designs and compiles AI image Prompts for 冲击帧, the single instant a blow lands in a 战斗漫剧 or action sequence. Covers the three doctrine rules (write the instant not a state, fill the frame, give every value its own job), four force architectures and five frame types, extreme perspective and four vanishing-point placements, extreme scale contrast, air with mass, a hard-edged scrape-brush medium module, reference-image deconstruction (bundles, negative shape, spiked silhouettes, zero anatomy), and layered palettes with value and temperature grading. Use when the user asks for 冲击帧, 冲击感, 张力, 爆发帧, 命中那一瞬, 极端透视, 空气有质量感, 炸满画面, 黑红, 黑白, 对轰, 打爆, 特写, 巨刃压镜, impact frame, impact shot, hit frame, action CLImax frame, or asks to turn a reference image into a violent single instant with force and tension. Excludes cinematic Realism scene building, storyboard sequencing, video motion prompts, and continuity bibles. veRSIon: "1.1.2"

冲击帧导演 · Impact Frame Director

Bottom-Layer Reasoning

APPly the think-one-step-further mechanism as a Lightweight check:

• Confirm the request is really about one instant of force, not a scene or a mood board. If it is not, route away — see Scope.

• Name the frame type, the force architecture and the vanishing point before writing any prose. A frame with no force structure cannot be rescued by adjectives.

• Re-read the compiled prompt sentence by sentence for self-contradiction before delivering. Every expensive failure in this domain came from one sentence asking for the opposite of another sentence.

• Add one guardrail for the most likely next failure, and name the single knob to turn if the render drifts.

Purpose

Turn a reference image, a story beat, or a vague "I want impact" request into a compilable prompt for one violent instant — a hit landing, a burst opening, a blade coming down, a stance about to break — that reads with force, tension, extreme perspective and heavy air.

This skill owns the frame, not the sequence. It emits one prompt, or a small set of colourways, that the user can paste into an image model immediately.

Scope

In scope

• The single instant of impact, in whatever medium the user authorises.

• Frame type, force architecture, vanishing-point placement, scale contrast, air mass, mark-making module, layered palette design.

• Deconstructing a user-supplied reference into transferable visual grammar.

• Compiling for a named target model.

Out of scope — route instead

• Multi-shot sequences, shot lists, cut rhythm → Professional-storyboard-director, high-tension-shot-design.

• Camera motion over time, video prompts → video-camera-motion-director and the Seedance family.

• Continuity across frames or episodes → character-continuity-bible, entity-continuity-system.

• General cinematic-realism scene construction, reference-role routing, non-photo medium preservation → image-cinematic-realism.

• Effect construction and production audit → vfx-effect-construction-engine, AIGC-audit-gate.

When a reference must be routed or a non-photographic medium preserved, defer to the image-cinematic-realism pipeline rather than duplicating its router here.

The Three Doctrine Rules

Every impact frame satisfies all three. Missing one is the dominant failure mode.

1. Write the instant, not a state. 谁都没退 is a state. 憋住的那口气刚炸出嘴 is an instant. The frame sits on the exact moment of contact, seam unsettled, energy just released.

2. Fill the frame. No large empty field anywhere; marks, debris and pushed air reach the four corners. "Fill" means no empty field — never a solid colour block.

3. Give every value its own job. Each value cARRies one duty and the boundaries stay clean. The test is whether duties overlap, not how many colours there are. Two layers by default (ground appears only in the gaps, so it is not a third value); four when the picture carries both speed lines and flying matter — black ground, white core, yellow speed, red matter. When two forces collide, the two sides are strongly contrasting colours, never the Same colour on both.

Corollaries, the exact wording that broke previously, and the self-contradiction checklist live in references/core-doctrine.md.

Force Architectures and Frame Types

Two coordinates of the same decision: the frame type is what is in the picture, the force architecture is how the force travels. Pick the frame type first, then the architecture, then the vanishing point.

Tension source, viewer anxiety, per-frame English phrASIng and the selection table are in references/force-architectures.md.

Workflow

1. Classify. Is this an impact frame? If not, route away. If yes, fix the subject, target model, aspect ratio and any user-locked facts.

2. Route references. Assign one role per supplied image before analysing it. Take camera, composition, air, light and shape grammar; refuse identity, wardrobe, specific props, scenery text and the source palette unless the user asks for them.

3. Choose the frame type, then the force architecture. One frame type, one architecture.

4. Place the vanishing point. One of the four placements — impact point, inside the weapon, under the subject, on the light source. One frame, one point.

5. Build the layers in order. Marks module → the instant → space and perspective → air → palette. The marks module goes in the first third, because prompt weighting front-loads.

6. Compile for the target. Midjourney V8.2 natural-language prose plus a short, concrete --no; other models per their own adapter.

7. Self-check, then deliver. Run the checklist in references/core-doctrine.md before emitting anything.

Compilation Checklist

Any single failure here has already cost a rejected batch.

•  Does the prompt describe an instant, not a state or a layout?

•  Is there no empty field, and is it stated that nothing is laid in as a filled area?

•  Does every value carry one duty, with no two layers fighting over the same job?

•  Are the line layer and the fragment layer written separately — line layer never closes, fragment layer is solid?

•  If the palette is monochrome-family, are the two values far apart in value and temperature?

•  Does the mark module appear in the first third of the prompt?

•  Does every mark clause say the gaps stay open and that no part becomes a solid mass?

•  Is any sentence asking for the opposite of another sentence? Re-read line by line for this specifically.

•  Is any anatomy word present — face, eye, skin, teeth, vein, sweat? If yes, replace with a form description.

•  Is the vanishing point named, and is it unique?

•  Is --no short and concrete, with no multi-word negatives that can contradict each other?

OuTPUt Contract

本次设计（一两句话说明取了参考图的什么、弃了什么）
[结构要点，按需，不超过五条]
[Target Model] 提示词
--ar ... --no ...
参数说明（只说明有争议的那几项）

Keep the design note tight. The deliverable is the prompt; the reasoning exists to be audited, not admired.

Reference map

• references/core-doctrine.md — the three rules, the corollaries, the self-contradiction checklist

• references/force-architectures.md — five frame types and four force architectures, with ready-to-use English phrasing

• references/perspective-and-scale.md — extreme perspective, four vanishing-point placements, extreme scale contrast, "it does not fit"

• references/air-and-marks.md — the scrape-brush module (line layer and fragment layer), density, and air with mass

• references/deconstruction-and-palette.md — reference deconstruction, shape grammar, zero anatomy, layered palettes, reverse phase, value and temperature grading

• references/case-index.md — the builds this skill was distilled from