git:20260927.e78c966 to git:20261003.b038f50

13 added, 4 removed. Audit A to A.

---

name: zy-cinematic-realism

description: Compile scene ideas, visual references, or existing prompts into model-native image prompts. Use for cinematic creation or Dream Decode visual analysis and transfer; when a reference has a non-photographic medium, preserve that medium rather than defaulting to cinematic realism. Supports ChatGPT Images 2.5 / GPT Image 2.5 (with GPT Image 2 legacy compatibility), Midjourney, Seedream 5.0 Pro, Nano Banana, or a model-neutral workflow, plus model routing, continuity, prompt diagnosis, and result repair.

---

<!--

Project: 造梦师：AI时代电影视觉指南

SPDX-License-Identifier: CC-BY-NC-4.0

Source: https://github.com/popopo-99/zy-cinematic-realism

-->

# 造梦师 · ZY Cinematic Realism

- Public edition: 《造梦师 v2.3.0 — 解梦 / DREAM DIRECTOR v2.3.0 — Dream Decode》.

+ Public edition: 《造梦师 v2.4.0 — 创作控制与继续使用 / DREAM DIRECTOR v2.4.0 — Creator Control & Continuity》.

Build every result in this order:

`Scene Master → Creative Grammar → Model Compiler → Result Repair`

First decide what the image is, then how it is witnessed, and only then how to express it to a target model. One Scene Master may compile into multiple native model prompts.

When visual references are used for analysis or transfer:

`User Intent + Reference Image(s) → Reference Role Router → Per-Reference Medium Analysis → Dream Decode (Expression Mechanism only when observed) → Core Visual Rules → Transfer Scope → Compiler Priority Gate`

Then compile `Scene Master + Decoded Visual Grammar → target-specific Model Compiler`. Scene Master remains the source of scene facts; reference-derived visual grammar remains a separate input.

## Mode Selection

Infer the mode from the request. Do not print a menu.

- **Create** — build a new cinematic scene and compile it for one target model.

- **Model Router** — recommend a model when the user is unsure.

- **Transcode** — preserve an existing scene while changing model-native expression.

- **Multi-model Pack** — compile one locked Scene Master for several models.

- **Continuity** — create a Continuity Bible, shot list, and controlled Shot Deltas.

- **Repair** — diagnose a generated result and produce a model-native repair instruction.

- **Prompt Check** — inspect an existing prompt without rewriting unless requested.

- **Director / Style / Cinematography** — apply a selected visual grammar.

- **Dream Decode** — analyze why references look the way they do, transfer their visual grammar, synthesize multiple references, create a Decode Card, or repair reference drift. Internally select Analyze, Transfer, Multi-Reference, or Repair; do not print a submenu. Trigger for requests such as `分析这张图画风`, `模仿图一风格`, `复刻这种美术风格`, `为什么这张图看起来这样`, `提取视觉语言`, `图一风格做图二内容`, `参考这几张 moodboard`, `生成结果还是不像参考`, or `帮我对比哪里漂了`.

- **Remix** — change one major variable while preserving every other invariant.

- **Creative Shuffle** — combine one causal story card, camera card, and style card.

+ - **Project Handoff** — summarize accepted project state for reuse, or restore from a user-supplied card; no automatic storage.

## Target Model Gate

When the user wants a prompt ready for generation and the target model is unknown, ask only:

> 你准备在哪个模型里生成：GPT Image 2.5、Midjourney、Seedream 5.0 Pro、Nano Banana，还是其他？

Ask once only when the choice materially changes prompt structure. Do not ask when the user already named a model or the active conversation establishes it. If the user says to proceed, skip questions, use a model-neutral Scene Master prompt, and mention model-specific transcoding only when extra explanation is allowed.

If the user requests prompt-only output, return only the prompt requested: no interpretation, menu, follow-up, or unrelated adapter question.

Generic `GPT Image`, `OpenAI image`, `ChatGPT 生图`, `GPT Image 2.5`, or `Images 2.5` requests use the ChatGPT Images 2.5-compatible adapter. Explicit `GPT Image 2`, `GPT-Image-2`, or `gpt-image-2` requests retain legacy compatibility. For migration, test validated prompts unchanged first; repair only evaluated failures. The Skill does not pin a ChatGPT / Codex image model or invoke an API model ID. Keep API details out of ordinary prompt output.

Unqualified `Midjourney` or `MJ` routes to the current Midjourney V8.2 adapter. Treat an explicitly requested V6, V6.1, or V7 target as legacy and verify its compatibility rather than silently applying V8.2 behavior.

## Scene Master Invariants

Build or recover the Scene Master before applying style or model syntax. Preserve all user-fixed facts. A model adapter may change structure, order, density, native syntax, exclusions, reference-image wording, parameter form, or edit-instruction form. It may not silently change:

- character identity, scene, time, weather, story beat, or current action;

- visual center, camera witness position, spatial layers, or composition logic;

- primary light sources, core props, aspect ratio, or explicit restrictions.

For Dream Decode visual decisions (camera/viewpoint, composition, subject scale, light, framing, perspective, blocking, aspect ratio, visual center, and similar fields), mark an explicit user choice **USER-LOCKED** and an unspecified system-completed value **OPEN**. Priority is `USER-LOCKED > reference-derived Hybrid Decision > system-generated default`. Resolve OPEN values using Hybrid Decisions and Transfer Scope (Preserve / Adapt / Fill / Release) before the final Scene Master lock; never overwrite USER-LOCKED choices. The resolved Scene Master remains the sole source of scene facts.

For Transcode and Multi-model Pack, apply a Transcode Lock before writing any target prompt. `MODEL SYNTAX MAY CHANGE. SCENE LOGIC MAY NOT.`

For Dream Decode, separate reference **Scene Facts**, transferable **Visual Grammar**, and **Hybrid Decisions** such as composition, viewpoint, subject scale, occlusion, spatial layering, and light direction. Route each reference's Primary Medium after assigning its role; identify an Expression Mechanism only when a distinct, observable one exists. Transfer visual logic without importing unassigned people, wardrobe, objects, locations, text, brands, or story events. USER-LOCKED Scene Master facts outrank reference-derived grammar; an authorized non-photo medium outranks default cinematic realism unless the user requests reinterpretation. For non-photo media, use viewpoint, value, edges, marks, surface, and rendering behavior rather than default camera/capture instructions.

## Core Create Workflow

These camera and source-light steps apply to photographic / cinematic Create work. When Dream Decode authorizes a non-photographic Primary Medium, express the corresponding viewpoint, value, surface, marks, or stylized rendering behavior instead of forcing photographic capture.

1. Parse fixed facts: era, place, characters, event, mood, time, weather, aspect ratio, target model, references, and restrictions.

- 2. Choose a specific story beat with a narrative before, concrete current action, and implied next event. Prefer observed transition, waiting, aftermath, departure, private observation, or interrupted routine unless the user requests a climax.

+ 2. Choose a specific story beat with a narrative before, concrete current action, and implied next event. Use the emotional direction and timing requested by the user, including joy, public celebration, bright daylight, or climax. Observed transition, waiting, aftermath, departure, and interrupted routine are optional approaches when the brief leaves the moment open.

3. Give each character physical blocking and a small action. Express emotion through posture, distance, gaze, silence, and object handling, not posing.

4. Construct foreground, midground, background, usable topology, causal traces of use, and materially consistent surfaces.

5. Place the camera at a physically possible witness position. Choose distance, height, focal behavior, boundary, and movement only as the story requires.

6. Map motivated sources, protected shadows, highlight surfaces, exposure behavior, and restrained color relationships.

7. Apply only the requested creative grammar. Do not let a card or director overwrite fixed facts.

8. Compile through the selected model adapter. Select scene-specific exclusions rather than pasting a generic negative list.

9. Apply anti-AI cleanup and run the quality checklist before answering.

## Reference Routing

Load only the branch required for the active mode.

- **Create:** [cinematic-principles.md](references/cinematic-principles.md), [camera-and-light.md](references/camera-and-light.md), [anti-ai-cleanup.md](references/anti-ai-cleanup.md), [model-routing.md](references/model-routing.md), the selected adapter, and [quality-checklist.md](references/quality-checklist.md). Read [negative-prompts.md](references/negative-prompts.md) when exclusions are needed.

- **Model Router:** [model-routing.md](references/model-routing.md) and [model-capability-matrix.md](references/model-capability-matrix.md).

- **Transcode / Multi-model Pack:** [prompt-compiler.md](references/prompt-compiler.md), [model-routing.md](references/model-routing.md), and only the source and target adapters needed.

- **Prompt Check:** [prompt-check.md](references/prompt-check.md), the target adapter when known, and [anti-ai-cleanup.md](references/anti-ai-cleanup.md) when cinematic realism is relevant.

- **Continuity:** [continuity-cards.md](references/continuity-cards.md), [prompt-compiler.md](references/prompt-compiler.md), and the target adapter.

- **Repair:** [result-repair.md](references/result-repair.md), the target adapter, and [continuity-cards.md](references/continuity-cards.md) for a series.

- **Dream Decode:** [reference-role-router.md](references/reference-role-router.md) → [medium-router.md](references/medium-router.md) → [dream-decode.md](references/dream-decode.md); read [prompt-compiler.md](references/prompt-compiler.md) and only the selected adapter when compilation is required. Also read [result-repair.md](references/result-repair.md) for Decode Repair and [decode-card.md](references/decode-card.md) when the user requests a reusable card or series-level visual continuity. Do not load all Style Cards by default.

- **Director:** [director-routing.md](references/director-routing.md), [directors/index.md](references/directors/index.md), and one matching director file. For recommendations, also read [directors/recommendation-matrix.md](references/directors/recommendation-matrix.md) and load only two or three candidates.

- **Style:** [creative-cards.md](references/creative-cards.md) and [style-cards.md](references/style-cards.md).

- **Cinematography:** [cinematography-cards.md](references/cinematography-cards.md).

- **Remix:** [remix.md](references/remix.md), [prompt-compiler.md](references/prompt-compiler.md), and the target adapter.

- **Creative Shuffle:** [creative-shuffle.md](references/creative-shuffle.md), [style-cards.md](references/style-cards.md), and [cinematography-cards.md](references/cinematography-cards.md) when useful.

+ - **Project Handoff:** [project-handoff.md](references/project-handoff.md) only when the user requests handoff, saving a card, or restoring a project; add [continuity-cards.md](references/continuity-cards.md) for a series.

- **Examples:** read [examples.md](references/examples.md) only when calibration is genuinely useful.

- **Scaffold:** use [basic-prompt-template.md](assets/basic-prompt-template.md) only when a compact model-neutral Scene Master writing order is useful; the selected adapter still controls final syntax.

Model adapters:

- [ChatGPT Images 2.5 / GPT Image 2.5, plus GPT Image 2 legacy compatibility](references/models/gpt-image-2.md)

- [Midjourney V8.2](references/models/midjourney.md)

- [Seedream 5.0 Pro](references/models/seedream-5-pro.md)

- [Nano Banana family](references/models/nano-banana.md)

## Director Compatibility

- Preserve the existing Director Four-Axis system. Use it only when the user names a director, requests a director method or comparison, or asks for a recommendation. Supported named directors remain mandatory strong/iconic mode. Follow `director-routing.md`, translate all four axes into scene-specific decisions, and never substitute a name or film title for camera, light, color/exposure, composition, space, and blocking. Do not copy a specific film shot.

+ Preserve the existing Director Four-Axis system. Use it only when the user names a director, requests a director method or comparison, or asks for a recommendation. Respect subtle / clear / strong director strength; when unspecified use clear. USER-LOCKED facts and visual decisions always outrank differentiation. Keep four-axis decisions internal unless explanation helps or the user requests a director breakdown; the target adapter controls final prompt syntax. Follow `director-routing.md`, translate all four axes into scene-specific decisions, and never substitute a name or film title for camera, light, color/exposure, composition, space, and blocking. Do not copy a specific film shot.

+ ## Output Depth and Follow-up

+ Infer depth from the request; do not print a menu or add a required selection turn. Prompt-only means only the requested prompt. Ordinary Create work may use a brief `画面决定` / `Visual decisions` of two to four concrete choices (moment, viewpoint, light, or relevant locks), followed by the prompt. Do not repeat the prompt as an explanation. Detailed Scene Master, four-axis breakdown, cards, and series plans are shown only when requested or genuinely needed by that mode.

+ On follow-up edits, recover the latest accepted state, change only the requested variable, and retain earlier accepted changes. A suggested choice is not a user-approved image. State image-generation or inspection status accurately; prompt completion is not image generation. Project Handoff is opt-in and does not add a card to ordinary output.

## Output Contracts

- - **Create:** `画面理解` → `[Target Model] Prompt` → target-native constraints or settings only when useful.

+ - **Create:** brief `画面决定` / `Visual decisions` when useful → `[Target Model] Prompt` → target-native constraints or settings only when useful.

- **Transcode:** short `Scene Lock Summary` → `[Target Model] Prompt`; do not re-explain the concept.

- **Multi-model Pack:** one short Scene Lock → clearly different native prompts sharing the same invariants.

- **Prompt Check:** three to five high-impact `PASS / WEAK / RISK / CONFLICT` findings → Rewrite, Surgical Fix, or No Rewrite as requested.

- **Repair:** dominant failures → `CHANGE ONLY` → `PRESERVE EXACTLY` → target-native repair prompt.

- **Dream Decode — Analyze:** `画面理解` → `媒介判断` → optional `表达机制` (only when observed) → `解梦分析` → five to eight `核心梦律`; omit an absent mechanism rather than printing an empty block. No prompt when the user asks only for analysis.

- **Dream Decode — Transfer:** `解梦摘要` → `核心视觉规则` → `[Target Model] Prompt`. Prompt-only requests return only the prompt.

- **Dream Decode — Multi-Reference:** compact `Reference Roles` → `Shared / Assigned Visual Grammar` → Core Rules → prompt when requested.

- **Decode Card:** only when requested or needed for reuse or series continuity; include five to eight `Core Visual Rules / 核心梦律`, explicit `Transfer Scope`, `Allowed Variation`, `Source Residue`, and `Drift Warnings`; optionally include evidence-backed Primary Medium, Medium Constraints, and Expression Mechanism. Use a compact card unless deeper archival grammar is useful, and never claim automatic persistence.

- **Decode Repair:** classify `Valid Adaptation` before `Actual Drift`; only actual drift enters at most three dominant failures → `CHANGE ONLY` → `PRESERVE EXACTLY` → target-native repair prompt.

- **Continuity:** Continuity Bible → Shot List → Shot Deltas → model-native prompts.

- **Remix:** Preserved → Changed axis → New Prompt.

- **Shuffle:** Combination → Story Card → Camera Card → Style Card → Final Prompt.

+ - **Project Handoff:** a compact, complete Project Handoff Card when requested; on restore, a short recovered-state summary and the requested next output. Request only a missing input that materially blocks the next task.

Do not expose hidden reasoning or internal locks unless the user requests them. Do not leave placeholders.

## Non-Negotiable Rules

- The frame must still work after removing `cinematic`, camera brands, film stock, focal-length badges, resolution claims, and grain.

- Prefer specific nouns, physical actions, causal traces, and source-based light over `masterpiece`, `epic`, `stunning`, `award-winning`, `8K`, or `highly detailed`.

- Do not invent decorative rim light, neon, fog, flare, shallow focus, damage, dirt, obstruction, or imperfection without a physical or narrative cause.

- Avoid poster, fashion-campaign, game-render, concept-art, and advertising logic unless requested or intrinsic to an explicitly authorized Dream Decode reference medium. Never let this default override a paper illustration, print, collage, stylized 3D, or other non-photographic primary medium.

- Do not invent model versions, API fields, parameter values, reference counts, strength ranges, or resolution limits. Frontend behavior may differ from the underlying model.

- Respect the user's requested language, format, model, aspect ratio, and output brevity.