feat: make video prompts scene-led and preserve authored continuity

Separate stable design from evolving action, environment and performance.
Forward shot transitions, lens intent and supported sound without generic
quality defaults; retain legacy audio and transition compatibility.
Add source research, contrasting authoring examples and regression coverage.

Author: Jinjun Gao

CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## Unreleased
 
+## 0.11.0 - 2026-09-09
+
+### Changed
+
+- Reworked video prompt authoring around scene intent, reference roles, stable design and evolving event state across action, performance, environment, product and graphic work. Preserve authored format and craft without a mandatory twelve-part cinematic prefix.
+- Removed unrequested photoreal/skin, pause, palette ratio, composition formula, resolution, frame-rate and shutter defaults. Explicit quality overrides remain supported.
+- Carry shot transitions, local continuity and native sound bridges into compiled prompts; preserve unit scope and locally rebased timing without exposing internal identifiers.
+- Added source-backed prompt research and anonymized failure analysis covering reference-board leakage, held contact poses, state resets and conflicting visual authority. Private media and conversation evidence are not bundled.
+
+### Verification boundary
+
+- Compiler/fixture tests and installed-method exercises verify instruction and text behavior. The inspected negative video informs the diagnosis; no claim is made that a new video-model candidate has proven quality improvement.
+
 ## 0.10.0 - 2026-09-08
 
 - Bring Jingzao photography advice into existing Studio story and shot development before camera choices are locked. Connect composition and environment to story, blocking and shot information; preserve story ownership, Fast scope, existing budgets and separate image execution.

README.md

@@ -34,6 +34,14 @@ Start with the work you need. A small edit stays small; a complete preproduction
 
 Outputs depend on the requested scope and available providers. A prompt is ready for submission only after its actual references and model requirements are checked. Image generation, video generation and final approval remain distinct steps.
 
+## Scene-led video prompts
+
+Prompt structure follows the scene: stable references and design, then evolving
+action, performance, environment, camera and sound. Authored detail is preserved;
+unrequested cinematic prefixes and rhythm formulas are not added. Browse
+[six contrasting prompt examples](examples/prompt-structure-quality/README.md)
+and the [source-backed research and limits](docs/film-preproduction/research/prompt-structure-20260909.md).
+
 ## Product CG styles
 
 Choose from six optional families: dry powder, precision hard surfaces, elastic
@@ -46,7 +54,7 @@ See the [product CG library and test boundaries](examples/product-cg-style-libra
 
 You need a Codex environment with Skills support and Python 3.10 or newer. Image, video, transcription and media-download services have their own requirements.
 
-For a **verified release installation**, download the archive and checksums from [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases/latest), then follow the [exact-tag installation guide](README.technical.md#verified-release-install). The source version is **v0.10.0**; release assets are authoritative for published availability.
+For a **verified release installation**, download the archive and checksums from [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases/latest), then follow the [exact-tag installation guide](README.technical.md#verified-release-install). The source version is **v0.11.0**; release assets are authoritative for published availability.
 
 For a development copy:
 

README.technical.md

@@ -28,7 +28,7 @@ This guide preserves detailed workflow, integration, validation and verified-rel
 
 DIRcreative 不是“输入一句话、吐出一堆提示词”的黑盒。它先判断任务是局部修改、完整开发还是交付审计，再只加载对应合同。局部任务直接交付修改结果；只有真实方向冲突、生成授权或客户交付才停下来询问。
 
-当前源码版本为 `v0.10.0`；已发布版本与下载以 [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases) 为准。本版用公共入口完成资产准备与一次批次审阅，复用未变化的候选质量证据，减少生产任务手写协议和重复看图；保留五视图、真实依赖与最终采纳边界。成人体型和服装剪裁按明确设计传递，不以固定 A-pose 或默认保守服装覆盖；视频提示词在镜头正文前统一采用十二项电影质感前缀。真实生成和输出审核仍有服务与模型限制，不保证每次出图或普遍一次成功。
+当前源码版本为 `v0.11.0`；已发布版本与下载以 [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases) 为准。本版用公共入口完成资产准备与一次批次审阅，复用未变化的候选质量证据，减少生产任务手写协议和重复看图；保留五视图、真实依赖与最终采纳边界。成人体型和服装剪裁按明确设计传递，不以固定 A-pose 或默认保守服装覆盖；视频提示词按场景组织稳定事实与逐拍变化，保留实际参考、环境、表演、摄影和声音；仅转发已写明的质量方向，不自动补十二项电影质感前缀。真实生成和输出审核仍有服务与模型限制，不保证每次出图或普遍一次成功。
 
 源码仓库包含 DIRcreative 根 Skill、19 个 `skills/dircreative/*` 内部子 Skill，以及 `ai-film-asset-stress-test`、`ai-film-production-ledger` 两个 P0 能力入口。正式 DIRcreative 安装包按安全设计只暴露根 `$dircreative`，其余入口会内部化后由 selector 路由；Skill Stack 还会发现宿主中已安装的外部专业 provider。许可明确的依赖可通过锁定 bundle 随包提供，镜造通过显式更新入口安装；其他依赖仍由宿主发现。具体范围与命令见 [依赖安装指南](docs/dependencies.md)，不能把“宿主可调用”表述成“GitHub 已内置”。ADCO 始终是独立外部编排方。
 
@@ -307,12 +307,12 @@ Skill; see [review trust host configuration](docs/film-preproduction/review-trus
 正式安装源是同一 GitHub Release 中的归档和 `SHA256SUMS`，再由该 tag 的精确、
 干净源码执行同进程验证与安装；不能运行归档内的 installer，也不能用 metadata
 自证。下面的信任链从 `v0.5.0` 起适用；更早版本不满足这条正式安装门。
-以下命令在 `v0.10.0` tag 与 Release 实际发布后生效。
+以下命令在 `v0.11.0` tag 与 Release 实际发布后生效。
 
 ```bash
-gh release download v0.10.0 \
+gh release download v0.11.0 \
   --repo papperrollinggery/Paperrolling-DIRcreative-SKILL \
-  --pattern 'dircreative-0.10.0.tar.gz' \
+  --pattern 'dircreative-0.11.0.tar.gz' \
   --pattern 'SHA256SUMS'
 ```
 
@@ -322,7 +322,7 @@ gh release download v0.10.0 \
 ```bash
 set -euo pipefail
 REPO_URL="https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL.git"
-TAG="v0.10.0"
+TAG="v0.11.0"
 EXPECTED_COMMIT="$(
   git ls-remote --exit-code --tags "$REPO_URL" \
     "refs/tags/$TAG" "refs/tags/$TAG^{}" |
@@ -334,7 +334,7 @@ EXPECTED_COMMIT="$(
     }
   '
 )"
-ARTIFACT="$(pwd)/dircreative-0.10.0.tar.gz"
+ARTIFACT="$(pwd)/dircreative-0.11.0.tar.gz"
 CHECKSUMS="$(pwd)/SHA256SUMS"
 VERIFY_ROOT="$(mktemp -d)"
 trap 'rm -rf "$VERIFY_ROOT"' EXIT

README.zh-CN.md

@@ -34,6 +34,13 @@ DIRcreative 是面向 **AI 影视前期制作的 Codex Skill**。从一句创意
 
 产出取决于请求范围和可用工具。提示词提交前需要核对实际参考资产与模型要求；图片生成、视频生成和最终采纳分别处理。
 
+## 按场景组织视频提示词
+
+先明确参考职责与稳定设计，再写逐拍变化，让动作、表演、环境、摄影与声音互相衔接。
+保留具体创作要求，取消自动附加的电影质感前缀与固定节奏模板。
+可查看[六类对照提示词](examples/prompt-structure-quality/README.md)和
+[研究依据与验证边界](docs/film-preproduction/research/prompt-structure-20260909.md)。
+
 ## 产品 CG 风格与分析
 
 可按需调用柔光干粉、精密硬表面、弹性与织构、流变触感、雕塑材质、图形模块六类风格。
@@ -45,7 +52,7 @@ DIRcreative 是面向 **AI 影视前期制作的 Codex Skill**。从一句创意
 
 需要支持 Skills 的 Codex 环境与 Python 3.10 及以上。图像、视频、转录和媒体下载服务各有自己的运行要求。
 
-**正式版本安装**：从 [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases/latest) 下载归档与校验文件，按[精确 tag 安装说明](README.technical.md#verified-release-install)完成验证。当前源码版本为 **v0.10.0**；已发布版本以 Release 中的实际资产为准。
+**正式版本安装**：从 [GitHub Releases](https://github.com/papperrollinggery/Paperrolling-DIRcreative-SKILL/releases/latest) 下载归档与校验文件，按[精确 tag 安装说明](README.technical.md#verified-release-install)完成验证。当前源码版本为 **v0.11.0**；已发布版本以 Release 中的实际资产为准。
 
 安装开发副本：
 

VERSION

@@ -1 +1 @@
-0.10.0
+0.11.0

docs/film-preproduction/production-prompt-discipline.md

@@ -59,45 +59,24 @@ If any check fails, stop before generation or external handoff and route to the
 
 ## Video Prompt Structure
 
-Use a model-agnostic five-part discipline for every video prompt:
-
-```text
-Model policy -> Camera -> Subject -> Look -> Action
-```
-
-- Model: which target model or prompt-only target is being prepared.
-- Camera: one motivated camera behavior with start and end target.
-- Subject: visible actor, product, prop, or scene owner inherited from source truth.
-- Look: lighting, material behavior, palette, style, and surface constraints.
-- Action: one primary action that fits the shot duration.
-
-Each model adapter may expand that into its own shape. A six-slot adapter is valid when the model benefits from explicit separation:
-
-```text
-camera + subject + action + setting + style + lighting
-```
-
-Rules:
-
-- camera must name physical start and end targets or explicitly stay locked-off,
-- subject must inherit identity/product locks instead of redesigning them,
-- action must fit the shot duration and input frame,
-- setting must inherit the scene/FOV source,
-- style and lighting must be concrete production constraints, not praise words,
-- audio is separate from visual prompting unless the target model supports native audio.
-
-Before finalizing a video prompt, reduce the shot to a micro-scene beat sheet:
-
-- initial visible state,
-- trigger or pressure,
-- subject action path,
-- camera start target,
-- camera end target,
-- timing beat or pause,
-- final visible state,
-- sound or silence policy when relevant.
-
-Reject prompts that combine several unrelated actions, multiple camera moves, and a slogan into one shot.
+Use `skills/dircreative/references/prompt-structure.md` as the maintained writing
+method. Model policy is resolved internally; copyable text carries the actual
+scene and supported references. Organize relevant intent, reference jobs and
+stable design once, then causal events with camera, environment, performance,
+sound and outgoing state. The order and heading count follow the scene and the
+user's format, not a universal Model-Camera-Subject-Look-Action template.
+
+Preserve the original creative scope. One dominant event may contain a connected
+exchange; do not force it into one verb, slow movement or a held ending. An
+actual start frame differs from identity/style/motion reference. Environment
+changes and object ownership persist across cuts. A written avoid list cannot
+remove unwanted pixels from a reference: review actual inputs and resolve their
+jobs before export. Model-native audio and finishing cues keep distinct routes.
+
+The internal micro-scene check is current relation, trigger, owned path,
+consequence and next condition. It detects causal gaps, not required prose
+headings. Quality is judged on the scene, then on inspected generated output;
+field counts and prompt length do not prove fidelity or cinematic effect.
 
 ## Image Prompt Structure