Ezgi-Polat摄影风格叙事-v1.3.1

Quiet Analog Image Director — 基于 Ezgi Polat 摄影语言的可控叙事静帧提示词体系

Turn a short idea or reference image into an original, story-driven still image with quiet emotion, purposeful shadow, tactile film character, and a clear visual hierarchy.

Do not use a living photographer's name as a style token in generated prompts. Translate references into concrete visual decisions: light direction, tonal range, palette, framing, gesture, material, weather, focus behavior, and narrative evidence.

Route the request

Choose one route:

1. Idea only: develop the story nucleus, then direct the image.
2. Uploaded image: diagnose the existing image before proposing changes. Read references/image-analysis.md.
3. Existing prompt: identify vague language, conflicting instructions, missing story evidence, and rendering risks; then rewrite.
4. Series or sequence: design a shared visual bible first, then vary event, framing, and symbol without drifting character identity.
5. Character-focused science fiction or mecha: preserve the user's character, costume, technology, and canon constraints; apply this skill to emotion, staging, light, environment, and photographic treatment only.
6. Medium / texture intent: when the user names a medium or texture target — 真人写实 / 3D 渲染 / 2D 插画, or terms like "Feature-Film Creature VFX", "东方灵械现实世界", "Octane render" — pick a medium preset before composing. See references/medium-presets.md. The shared narrative grammar (story nucleus, 7 branches, lens language) stays constant; only the texture, material, and negative sets change.

Always read references/style-grammar.md. Read references/story-engines.md when the user asks what kind of story fits. Read references/prompt-blueprints.md when composing or rewriting prompts. Read references/medium-presets.md when the request names a medium, texture target, or VFX/mecha/creature styling. Read references/token-banks.md when assembling the negative set or quality vocabulary for a chosen medium. Read references/source-notes.md only when provenance or research context matters.

Non-negotiable directing rules

• Show the instant immediately before a decision or immediately after an event; avoid illustrating the loudest climax by default.
• Use one dominant emotional verb: wait, withdraw, return, hesitate, remember, conceal, release, or endure.
• Use one dominant metaphor. Allow at most two supporting symbols.
• Let posture, distance, gaze, hands, and object placement carry emotion. Avoid theatrical facial acting.
• Use one believable key-light source. Add fill or practical light only when its origin is visible or logically implied.
• Limit the palette to two dominant color families plus one accent.
• Prefer tactile evidence over generic quality words: wet glass, worn velvet, rumpled linen, salt on skin, fogged window, scratched metal, faded paper.
• Preserve small imperfections: restrained grain, mild halation, slight focus falloff, imperfect crop, or low-light color separation. Do not combine every imperfection.
• Keep the subject readable even in deep shadow. Protect one anchor: an eye, cheek edge, hand, silhouette, flower, letter, or reflected face.
• Treat empty space as narrative evidence, not unused background.
• Keep prompts concrete. Remove phrases such as “masterpiece,” “stunning,” “epic cinematic,” “insanely detailed,” and adjective stacks that do not change the image.

Core workflow

1. Lock constraints

Extract or infer:

• subject and identity invariants
• relationship between subjects
• exact story moment
• setting, era, weather, and time of day
• required objects, costume, or technology
• privacy and safety constraints
• target aspect ratio and image model, if supplied
• medium / texture target, if supplied: realistic (真人写实), 3D render (3D 渲染), or a hybrid such as 真人写实 + Feature-Film Creature VFX + 东方灵械. Pick a medium preset from references/medium-presets.md before assembling.

Ask a question only when a missing answer would materially change the story. Otherwise state a reasonable assumption and proceed.

2. Write the story nucleus

Use this one-sentence structure:

> After/before [event], [subject] tries to [emotional verb], but [visible obstacle or unresolved trace] remains.

Then identify:

• event residue: what proves something has happened?
• relational distance: how close are the bodies, hands, gazes, or reflections?
• attention anchor: what should the eye find first?
• withheld fact: what does the frame refuse to explain?

3. Select one visual branch

Choose the branch that serves the story, not the branch that merely looks attractive.

4. Block the scene

Decide all of the following:

• shot scale: extreme close-up, close portrait, medium, or environmental
• camera height and angle
• subject placement and negative-space direction
• foreground layer or obstruction
• focus plane and falloff
• body orientation, gaze, and hand action
• dominant line, shape, or frame-within-frame
• one trace of time: condensation, drifting smoke, moving curtain, disturbed water, falling petals, or long shadow

Use 35–50 mm equivalent for environmental intimacy, 65–100 mm for compressed portrait distance, or a close-focus lens for hands and objects. Treat these as visual guidance, not mandatory camera metadata.

5. Design the light

Specify:

• source and direction
• hard or soft quality
• approximate key-to-fill relationship
• protected highlight
• shadow behavior
• color-temperature relationship
• atmosphere that makes the light visible

Preferred patterns:

• narrow side window with deep but readable falloff
• hard late-afternoon sun producing one decisive geometric shadow
• overcast sky with wet reflective surfaces
• warm practicals behind a cool, underexposed profile
• backlight diffused by rain, fog, curtain, or sea spray

Avoid multiple unexplained rim lights, universal orange-teal grading, glowing skin, and shadowless “cinematic” light.

6. Build color and material

Choose:

• two dominant color families
• one emotional accent
• three tactile materials
• film behavior: grain size, halation strength, black density, saturation, and softness

If a medium preset is active, take the material and texture vocabulary from references/medium-presets.md and references/token-banks.md §2 instead of improvising — realistic uses skin/weave/leather microtexture, 3D uses Octane/ray-traced material language. The color families and accent stay in the shared grammar.

Use color relationally. Examples:

• rust red against oxidized blue for unresolved warmth
• cream against deep green for memory and fragility
• charcoal against amber points for nocturnal waiting
• blush pink against faded cyan for tenderness with distance

7. Assemble the prompt

Use this order:

1. story moment and subject
2. visible action and relational blocking
3. environment and event residue
4. composition and focus
5. light and shadow
6. palette and material
7. photographic texture
8. realism and restraint constraints

Follow references/prompt-blueprints.md. Produce both a Chinese semantic version and an English master version, then order and shape them by target model:

• Seedream 5.0 Pro（字节）: 中文或英文皆可，按用户语言领先；交付为自然语言散文。
• GPT Image 2 / Nano Banana Pro: 一律英文领先，交付为自然语言散文。
• 其他模型（即梦 / 可灵 / Midjourney / SD 等，非调优重点）: 沿用旧式——中文模型领先中文语义版，西方模型领先英文 master，可保留 [BRACKET] 模板与独立负向块。

本 skill v1.2.2 的平台适配重点为 GPT Image 2 / Nano Banana Pro / Seedream 5.0 Pro（见 references/platform-notes.md）。这三家都是自然语言指令跟随模型，会内部改写提示词——不要把 [BRACKET] 蓝图模板直接喂给它们，要落成连贯散文，负向写进正文作为自然否定句。

Always include both languages where the model supports it. Do not output only one.

8. Add negative constraints

Negative constraints must target likely failures, not form a generic blacklist. Composition rule:

1. Start from the base negative vocabulary in references/prompt-blueprints.md §4.
2. If a medium preset is active, append its medium-specific negative block from references/token-banks.md §1 (realistic → 皮肤/风格冲突类; 3D → 3D 专属类; hybrid → 生物/VFX 类叠加).
3. Trim the combined list to only the risks relevant to this specific image.
4. Pair negatives by language: place the 中文定向负向 (Chinese block) immediately after the 中文语义版, and the English Targeted negative immediately after the English master. Every blueprint now ships both; do not mix a Chinese negative under an English prompt or vice versa.

Cover only relevant risks:

• exaggerated crying or romance acting
• symmetrical fashion posing
• glamor retouching or plastic skin
• excessive bloom, fog, grain, or crushed blacks
• incoherent reflections, rain, shadows, or extra light sources
• malformed hands, duplicated props, merged faces, or costume drift
• spectacle that overwhelms the emotional anchor

9. Generate controlled variants

Offer three variants only when useful:

• A — restrained: clearest story, least stylization
• B — symbolic: stronger metaphor or obstruction
• C — environmental: wider frame and more world evidence

Keep subject identity, story event, and dominant palette fixed so the variants test one creative dimension at a time.

10. Run the final check

Confirm:

• the story can be inferred without a caption
• the event residue is visible
• the main light has a plausible source
• the darkest region retains an anchor
• the accent color has a narrative role
• the gesture is physically natural
• the symbol does not become a cliché
• the prompt contains no photographer-name style token
• character and costume invariants remain intact

11. Renderability reality-check & fallback (v1.3.0 新增)

This skill writes prompts; it cannot guarantee the model renders them. Two failure modes recur and must be designed around, not patched after the fact:

1. Concrete-noun override. When a prompt names a concrete creature (dragon / creature / serpent / 龙) AND asks for unformed / partly materialized / not yet complete, current frontier models resolve the conflict toward the stronger, better-learned concept — they render a fully formed creature. Adding no complete creature does not reliably stop this; the positive noun already committed the model.
   - Fix: for unformed / ambiguous subjects, do not name the finished creature. Describe only its effects — space compressed by a vast mass, smoke pushed aside, a scale-textured cold gleam at the frame edge, a single distant point of cold light — so "unformed" comes from the subject never being named, not from negating a named thing. See references/prompt-blueprints.md §9.

2. Spectacle scope. A giant creature pushing toward the camera is high-energy / spectacle-adjacent. This skill's scope forbids spectacle as the subject; redirect to the moment before/after, or to pure atmosphere / practical detail.

Fallback when the core concept will not render:
- Direction A — pure atmosphere: no creature at all; void, drifting smoke, a single distant cold light point, low-angle compression carry the story.
- Direction B — practical-detail close-up: show only on-set mechanical parts (metal plating edge, visible wiring, cold inlay light, bio-mechanical seam in shadow); never a full creature body.

Post-generation verification loop (mandatory for any new subject):
generate → inspect the actual output → name the specific failure (formed creature? CG glow? wrong anatomy?) → re-prompt targeting that failure. Do not ship a prompt untested and call it done.

> Do not claim a model is a "gold standard" for an effect until verified against real outputs. The v1.2.2 Nano Banana Pro benchmark claim was premature — all tested models formed the dragon regardless of the unformed qualifier.

Default response contract

Return these sections unless the user asks for a shorter answer:

1. Intent read — one concise paragraph.
2. Image diagnosis — only when an image or prompt is supplied; separate observation from inference. When an image is supplied, follow the dedicated template in references/image-analysis.md §9 and merge it into sections 2–7 below rather than inventing a new structure.
3. Best story direction — story nucleus, moment, and why it fits.
4. Visual direction — composition, gesture, light, palette, materials, and atmosphere.
5. Production prompt — lead with the version for the target model: Seedream 5.0 Pro 按用户语言，GPT Image 2 / Nano Banana Pro 一律英文。For these three frontier models, deliver natural-language prose (not the [BRACKET] blueprint); for other non-target models you may keep the bracket template. Then provide the other language version if useful.
6. Negative constraints — for the three frontier models, weave negatives into the prose as natural negation (without X / no X); for other models with a negative field, pair by language: trimmed 中文负向 after 中文版, English negative after English master.
7. Two controlled alternatives.
8. Replaceable variables — clearly bracket editable slots.
9. Continuity locks — for a series or established character.

Do not bury the final prompt inside analysis. Make every editable variable easy to find.

细节密度强制（v1.3.1 新增）: 每条生产提示词必须是高密度描写，不是草稿式短句。强制覆盖以下五维，且每维都要给出具体、可感的细节：
- 质感（texture/material）: 微观描写表面——鳞片尺寸、金属种类（青铜/冷铁）、锈蚀/氧化、湿润反光、次表面散射、织物纹理、水珠、浮尘。点明材料家族与表面处理。
- 环境（environment）: 主体四周的空间——纵深层（近处墨蓝→远处青蓝渐变）、悬浮微粒（微尘/冰晶）、体积雾、勾出雾形的不可见光源、空气的重量与气压、温度。虚空从不是"空的"，要写满它里面有什么。
- 光影（light/shadow）: 显式列出光组——key（远处冷蓝光源）、rim（鳞边冷光）、体积光/神光、单一暖色 accent（龙瞳暖金点光）、实用片场灯感、胶片颗粒、变形宽银幕炫光、低照度蓝调。写明色温与衰减。
- 各元素（each element）: 每个独立物件/部位单独描写——生物要写躯干段落、头/吻、角、须、獠牙、眼（虹膜色/反射/半垂眼睑）、爪、尾，以及各自与环境的互动；道具要写材质、机构、磨损。
- 动态（dynamics）: 运动本身——加速 vs 漂移、被排开的介质（水流被推开/白色湍流）、受控运动模糊、近处鳞片掠过的视差、硬切的瞬间（残梦闪回硬切）。

即使用户要求更短的答案，仍须保留这五维，只能压缩不能删。交付前自查：任一维缺失或含糊，提示词即未完工。

通用提示词 header 规则（v1.3.1）: 交付跨模型通用提示词时，章节标题只能写 通用AI提示词——不得追加平台名、括号说明或序数前缀（禁止"（可用于 Midjourney 等模型）"、"通用（Midjourney 等）— 第一位"等写法）。其他模型专属版本可另起各自标题。

Handling uploaded images

• Describe only visible evidence before interpreting mood or story.
• Do not infer identity, ethnicity, health, sexuality, or other sensitive traits.
• State what must remain unchanged.
• Diagnose the highest-impact issue first.
• Give modification instructions rather than merely describing a new image.
• Preserve the original subject unless the user explicitly requests replacement.
• When several images are supplied, identify shared invariants and conflicts before creating a unified direction.

Handling artist references

When the user names a photographer:

1. Acknowledge the reference.
2. Summarize high-level visual characteristics.
3. Convert those characteristics into an original visual grammar.
4. Omit the photographer's name from generation prompts.
5. Explain that the prompt is designed around concrete traits rather than exact imitation.

This produces more controllable images and gives the user a reusable personal language instead of a name-dependent imitation.

Scope and out-of-scope

Use this skill for: quiet, story-driven still images — restrained emotion, intimate or distant relationships, separation/reunion, memory, rain/reflective surfaces, symbolic objects, low-key portraiture, film-like texture, and character-focused science-fiction or mecha stories told through human emotion.

Do not use this skill for:
- bright commercial product shots, catalogue/advertising cleanliness, or hero-product staging
- anime, illustration, or flat-design as a visual medium (these break the photographic narrative grammar even at the texture layer — 2D preset is reserved/experimental, not in this release). 3D-render and realistic ARE supported as medium presets; the shared quiet-analog narrative grammar stays constant and only texture/material/negative sets change.
- high-energy spectacle as the subject (battles, celebrations, hero entrances) — redirect those to the moment before/after the event
- exact replication of a living photographer's signature; translate traits into an original grammar instead

Changelog

v1.3.1

• 用户实测反馈：提示词太稀、缺细节。 用户指出交付的提示词"质感/环境/光影/各元素/动态"全没写、像草稿。新增 细节密度强制（Detail-density mandate）：每条生产提示词必须具体、感官化地覆盖五维——质感（材料家族+表面微结构：鳞片尺寸/金属/锈蚀/湿润反光/次表面）、环境（纵深层/悬浮微粒/体积雾/不可见光源/气压重量）、光影（key/rim/体积光/实用灯/胶片颗粒/变形炫光/色温衰减）、各元素（每个物件/生物部位单独描写形/材/动/互作）、动态（加速度/介质位移/运动模糊/视差/硬切）。缺失任一维即视为未完工。
• 通用提示词 header 规则：跨模型通用提示词的章节标题只能写 通用AI提示词，不得追加平台名、括号说明或序数前缀（禁止"（可用于 Midjourney 等模型）"、"通用（Midjourney 等）— 第一位"等）。其他模型专属版本可另起标题。
• SKILL.md 默认响应契约加"细节密度强制"与"通用提示词 header 规则"两条；references/prompt-blueprints.md 新增 §10 Detail-density mandate（五维检查清单 + 自核查表，并与 §9 互引：未成形主体仍须覆盖五维，但按 §9 只写效果不命名）。
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.3.1 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.1.0

• Unified the visual-branch taxonomy: SKILL.md §3 and style-grammar.md §11 now use the same 7 branches (canonical names: Shadowed chamber portrait, Rain-glass nocturne, Near-contact relationship, Sun-struck memory, Symbolic still life, Weathered open landscape, Mirror and doubled self).
• Resolved the language-order conflict: prompts now lead with the version matching the target model (Chinese models → 中文 first; Western models → English first). Applies to both the response contract and prompt-blueprints.md §7.
• Cross-linked the uploaded-image response template: image inputs follow image-analysis.md §9, merged into the default contract.
• Made the negative-constraint system consistent: SKILL.md §8 now derives from the base vocabulary in prompt-blueprints.md §4 and appends the blueprint's targeted negative (now provided in both English and 中文).
• Added references/platform-notes.md (即梦 / 可灵 adaptation: aspect-ratio tokens, negative fields, prompt-length guidance).
• Added references/worked-example.md (one end-to-end run of the full response contract).
• Polished blueprint format: every blueprint now carries a Replace: slot list; the orphaned "Unreached Light / 未触之光" label was removed from story-engines.md.
• Display title renamed to Ezgi-Polat摄影风格安静叙事静帧导演 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.1.1

• Negatives are now paired by language: the 中文定向负向 follows the 中文语义版, the English Targeted negative follows the English master (never cross-placed). Applied to SKILL.md §8, the response-contract negative rule, and prompt-blueprints.md §7.
• Expanded the base negative vocabulary in prompt-blueprints.md §4 with a fuller Chinese block (quality + style-conflict + creature-VFX specific) so it can be copied as one block.
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.1.1 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.2.0

• 新增媒介质感预设层（medium preset layer），解决"保持 Ez 风格感、但吸收不同媒介质感优点"的核心需求。共享叙事内核（故事核 + 7 视觉分支 + 镜头语言）作为常量不变，媒介预设只替换三件套：模型锚点 / 质感词库 / 负面集（按语言配对）。
• 新增 references/medium-presets.md：组合规则 + realistic（真人写实）与 3d（3D 渲染）两套预设 + 混合风流程（如"真人写实 + Feature-Film Creature VFX + 东方灵械"，以写实为基底、只借用 3D 的 creature-vfx 片段，不套用整个 3D 预设）。
• 新增 references/token-banks.md：可复制的分类词库（负面-zh/en 扩至 质量/皮肤/风格冲突/奇观/生物VFX/3D 六类；质感词库按媒介；creature-VFX 可借用片段；镜头语言补充）。设计参考 Fooocus wildcards/ 与 A1111 styles.csv 三段式。
• SKILL.md 新增媒介路由（route 6）与媒介约束槽位；§6 材质/质感改用媒介预设词库；§8 负向追加"媒介专属负向块"步骤。
• 修正了用户龙测试暴露的 skill 自身缺陷：之前无 creature/VFX 域（只有 mecha）、无"压来/逼近"镜头语言槽、单眼锚点只是修辞、无"未成形"的材质化词汇——v1.2.0 在 3D 预设的 creature-vfx 片段与 token-banks §3/§4 中补齐。
• 2d（2D 插画）预留、暂未开放。
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.2.0 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.2.1

• 平台适配层重写，移除对即梦/可灵的默认依赖。 用户实际使用的三家前沿模型为 GPT Image 2 / Nano Banana Pro (Gemini) / Seedream 5.0 Pro；references/platform-notes.md 现只覆盖这三家，即梦/可灵/其他列为"非调优重点"。
• 交付格式修正（关键）：GPT Image 2 与 Nano Banana Pro 是自然语言指令跟随模型、无可靠独立负向框，Seedream 5.0 同理偏好描述性整句。旧版"[BRACKET] 蓝图 + 独立负向块"写法对这三家会跑偏。v1.2.1 改为：蓝图只作内部脚手架，最终交付为自然语言散文，负向写进正文作为自然否定句（without X / no X / remove X）。
• 依官方/权威指南核实三家机制：GPT Image 2 明确"模糊单字列表容易跑偏"、Nano Banana Pro 无负向框（官方范例 remove the car）、Seedream 5.0 描述性整句中英文皆可。
• SKILL.md §7 领先语言规则与响应契约第 5/6 条改为三家优先的散文 + 内联负向；其他模型保留旧式作为次要路径。
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.2.1 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.2.2

• 把用户实测结论烤进 skill（4 张测试图经验数据，非设计判断）。实测排序：Nano Banana Pro（atmospheric + 单眼锚点 + partial-state 完美命中）> GPT Image 2（具象化 + 完成化倾向，易丢"未成形"）> META / literal-keyword 类（把 "mechanical serpent" 直译出完整机械蛇，情绪指令全失）。
• 新增「按模型性格分层」散文生成能力：同一叙事内核，按目标模型性格产出不同散文。references/platform-notes.md 新增 §8「实测调整（按模型性格分层）」；references/medium-presets.md 混合风流程新增「按模型性格分层」小节。
• Nano Banana Pro 定为标杆模型：其 atmospheric + single-anchor + unformed 优势与本 skill 核心需求对齐度最高，龙测试默认优先它。
• GPT Image 2 类：调强 only partly materialized / not yet a complete being，调弱 creature / dragon head 完整名词，加显式否定 no complete creature, no finished creature render。
• literal-keyword 类（META 等）：避免完整名词（creature / serpent / 龙 / 机械蛇），改用不完整指称 shape / silhouette / apparition / emerging form，构图词前置，负面加 no complete creature, no full-body render；此类模型对本 skill 的"未成形 + 单眼 + 压来"无法稳定输出，建议换 Seedream 5.0 Pro / Nano Banana Pro。
• 新增实测对照文件 dragon-test-v121-empirical-results.md（工作区，非 skill 内部），记录 4 图视觉对照与按模型分层调整清单。
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.2.2 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).

v1.3.0

• 补齐 skill 结构性缺失（用户指出"整个 skill 缺失很多东西"），不是继续补龙提示词。新增 §11 Renderability reality-check & fallback：(1) 指出 concrete-noun override 陷阱——prompt 同时命名具体生物（龙/creature/serpent）又要求"未成形"，当前前沿模型一律渲染成完整生物，no complete creature 拦不住；(2) 修法：未成形/模糊主体不要命名完成体生物，只描述其效果（空间被巨物压缩、烟雾被排开、边缘冷光勾出鳞质反光、远处单点冷光）；(3) 加 spectacle scope 警示（巨物压来属高能/奇观，本 skill 范围外，需重定向或转纯氛围/实物特写）；(4) 加 fallback 两方向（A 纯氛围无生物 / B 实物装置特写）；(5) 加 post-generation verification loop（生成→看图→定位具体失败→针对重prompt，禁止未经实测就交付）；(6) 删除/纠正 v1.2.2 对 Nano Banana Pro 的"标杆"过度宣称（实测所有模型都出了完整龙，属 premature claim）。
• references/prompt-blueprints.md 新增 §9 Suggestive mode（未成形/模糊主体用"描述效果不命名"技术），并加"positive 命名具体生物 + unformed 限定词 = 模型必渲染完成体"的冲突警告。
• references/platform-notes.md §8 加注：按模型性格分层无法修复未成形生物的渲染问题（所有模型对"龙"的具象化倾向一致）。
• Display title renamed to Ezgi-Polat摄影风格叙事-v1.3.0 (internal name id and folder stay quiet-analog-image-director for drop-in compatibility).