Repository navigation
Releases
v0.3.0 · 完整方案与动源 / Complete plan and motion sources
中文
这一版改的是流程本身,不是能力清单。起因是一次真实使用复盘:两支片都没按预期停下来讨论方案——从写意图摘要到整片脚本写完只隔两分半,整片配音和字幕也一并做完了,plan.md 最后才补上去(比脚本晚 43 分钟)。查下来不是模型不听话,是三处文档合起来在推它往前跑。
变更概览:新增 5 项 · 优化 4 项 · 修复 3 项
新增
- 三个必经确认点:意图确认后、完整方案后、小样后。写进 Preset 的 persona——停下来=这一轮结束、交出回合、等用户下一句;用户没回话就不做下一步,也不把没定的事先做完再说。
- 「完整方案」有了定义:
plan.md必须十节齐全——状态行、总述(散文,≤1000 字)、总览、阶段、视觉系统、声音、节奏与转场、标题与封面、制作路径、素材需求表、镜头表。 - 阶段层:片 → 阶段 → 镜。此前只有「段 → 镜」两层,缺的正是用户感受到的那一层。每阶段八列:阶段、起止、时长、主题、风格、视觉变化、镜数、做什么。
- 镜头第 11 字段「动源」:每镜必须声明这一镜的动从哪来——主体位移、运镜、切点、转场、元素动效、整段特效、静止。写了什么,验收就查什么。
plan-check门禁:evidence工具新增动作,也可独立跑tools/plan-check-cli.js。12 项检查,红灯非零退出,有红不许进制作。
优化
video-plan重写:删掉"方案不必四节全出""够用就停"——那是过度工程防护,但它同时也让方案停在了 10% 覆盖。改成十节全出、镜头表必须覆盖全片。video-motion扩写:新增动源一节,讲清七类判据、碎动与整段特效的分界、以及"实测动源与声明不符"这第六种病因。video-motion的触发面改写:它的描述原本只覆盖"用户抱怨像 PPT 时",结果在两支真实片里从未被触发——而片子恰恰就出了这个问题。video-review观察清单新增一节:报"这一段在动"之前,必须同时给出运动面积、最大连通块、连通块数量三个量,并与方案声明的动源对账。
修复
- 禁止清单自相矛盾:原第 3 条把"请用户确认后再继续"列为不合格,使任何暂停写法都会被判死。改为只禁止"每一步都拍板",三个确认点是必经的。
- 旧口径把碎动判成"在动":原口径是「任一块变化即算这一帧有变化」,一个光点飘过就让整帧通过。实测《万年归唐》小样按旧口径
longest_static = 0.0s、全项通过,而同一支片 55.7% 的帧是"高面积碎动"——变化散成 8 团以上、每团不到画面 2%,主体纹丝不动。观众看到的就是"背景在闪"。 video-edit补方案阶段分工:阶段的风格与视觉变化归它,具体帧号归制作阶段。
重要说明
plan-check会拦下按旧规范写的所有plan.md。 在制作品需要补齐十节、动源列与素材需求表,否则过不了门禁。这是有意的:旧规范下那两支片"都合格",其中一支的方案只覆盖了 10% 的镜头。- 动源契约里"爽片主体位移或运镜 ≥30%"是方向性数值,未在真实成片上校准。
- 新的流程本身还没有在真实片子上跑过一遍。它拦得住旧的坏方案(两个测试样本已验证),但"按新流程做出来的片子更好"这件事没有证据。
English
This release changes the workflow itself, not the feature list. It comes from a post-mortem of real use: two films skipped the planning conversation entirely — intent summary to finished script in two and a half minutes, with the full voice-over and subtitles done in the same stretch, and plan.md written last (43 minutes after the script). It was not a disobedient model; three separate parts of the documentation were pushing it forward.
Summary: Added 5 · Improved 4 · Fixed 3
Added
- Three mandatory checkpoints: after intent confirmation, after the complete plan, and after the sample. Written into the preset persona — stopping means the turn ends, the floor is handed over, and nothing further happens until the user replies.
- "Complete plan" now has a definition:
plan.mdmust contain all ten sections — state line, prose overview (≤1000 characters), at-a-glance table, stages, visual system, sound, rhythm and transitions, title and cover, production path, asset requirements table, and shot table. - A stage layer: film → stage → shot. There used to be only "section → shot"; the missing middle is exactly the unit a viewer actually perceives. Eight columns per stage: stage, time range, duration, theme, style, visual change, shot count, purpose.
- An eleventh shot field, "motion source": every shot must declare where its motion comes from — subject movement, camera move, cut, transition, element animation, full-frame effect, or still. Whatever is declared is what gets checked at review.
- A
plan-checkgate: a new action on theevidencetool, also runnable standalone viatools/plan-check-cli.js. Twelve checks; any red exits non-zero, and red means production cannot start.
Improved
video-planrewritten. Removed "the plan need not have all four sections" and "stop when it is good enough" — that was an anti-over-engineering guard, but it also let plans stall at 10% coverage. The plan is now all ten sections, and the shot table must cover the whole film.video-motionexpanded with a motion-source section covering the seven types, the line between scattered micro-motion and full-frame effects, and a sixth "looks like a slide deck" cause: motion source does not match what was declared.video-motiontriggering rewritten. Its description only fired when a user complained about slide decks, so it was never triggered across both real films — on films that had exactly that problem.video-reviewobservation checklist now requires three motion numbers (area, largest connected component, component count) alongside any claim that a segment "is moving", cross-checked against the declared motion source.
Fixed
- A self-contradicting prohibition. Rule 3 listed "ask the user to confirm before continuing" as unacceptable, which killed any pause instruction. Now only per-step approval is prohibited; the three checkpoints are mandatory.
- The old metric counted scattered micro-motion as motion. It was "any block that changed means this frame changed", so a single drifting light passed an entire frame. Measured on the 《万年归唐》 sample, the old gate reported
longest_static = 0.0sand passed everything, while 55.7% of that film's frames were high-area scattered motion — changes split across 8+ components, none larger than 2% of the frame, with the subject motionless. What the viewer sees is a shimmering background. video-editgained an explicit division of labour: per-stage style and visual change belong to it; frame-level timing belongs to production.
Important Notes
plan-checkwill reject everyplan.mdwritten under the old spec. Works in progress must gain the ten sections, the motion-source column and the asset requirements table, or they will not pass. This is deliberate: under the old spec those two films were both "acceptable", and one of them covered only 10% of its shots in the plan.- The "subject movement or camera move ≥30%" figure in the genre contract is directional and has not been calibrated against real finished films.
- The new workflow itself has not yet been run end to end on a real film. It does catch the old bad plans (verified with two fixtures), but there is no evidence yet that films made this way turn out better.
v0.2.1 · 双语 README 与上手路径 / Bilingual README and getting-started path
中文
这是一个文档版本。不新增能力,修掉的是"打开插件第一眼看得到的地方说的是错的",并把上手路径补全。已发布 0.2.0 里那条最重要的证据边界没有变化:这套东西只被完整跑通过一次。
变更概览:新增 3 项 · 优化 3 项 · 修复 4 项
新增
- 英文
README.md与中文README.zh.md双语互链,配一张 1600×800 封面横幅。封面由那支实测成片的真实静帧合成,生成脚本make-cover.py随包。 - 四张真实成片截图(1280×720),全部出自 5 分 12 秒那次全量实测。仓库此前一张图都没有。
video-library正文补上"库还没建"的判据与话术:不建空目录、不铺占位文件、不为读库而建库。
优化
- README 重写成一条可执行的路径:60 秒上手五步、第一轮会看到什么、你的资产库放在哪、它不做什么。
docs/DETAIL.md新增"你的资产库放在哪",把运行时库与作品文件夹的分工写清。- 已知限制改成更难被误读的写法,明确写出"没有人连看过一遍"与"资产库复用从未被验证"。
修复
locale/zh.json与locale/en.json写着"创作、提示词与声音三个方法技能",那是 0.1.0 的架构。实际是 13 个。docs/DETAIL.md的状态段仍写"尚未实测……下一步是装配后验证",而 0.2.0 早已在真实成片上跑通全量流程。docs/DETAIL.md把按需参考数写成 59 份,实际是 48 份。- 三处死链,指向根本不随包分发的
planning/夜间产出报告.md。
重要说明
- 无能力变更,无需迁移。技能正文升级后即时生效;
cordis.patch.yml未改,不需要重装或重启。 - 资产库与任何素材都不随包分发。你自己的库从第一支片开始时从零长出。
- Node.js 版本要求现在写进了
package.json:22.3 或更高。此前缺失,而 Preset 依赖process.getBuiltinModule。
English
A documentation release. It adds no capability; it fixes places where the plugin's first visible surface said something untrue, and completes the getting-started path. The most important evidence boundary from 0.2.0 is unchanged: this has been run end to end exactly once.
Summary: Added 3 · Improved 3 · Fixed 4
Added
- Linked English
README.mdand ChineseREADME.zh.md, with a 1600×800 cover banner composed from real stills of the delivered test film; the generator scriptmake-cover.pyships with the package. - Four real stills (1280×720), all taken from the 5 min 12 s end-to-end run. The repository previously contained no images at all.
video-librarynow states what to do when the library does not exist yet: no empty directories, no placeholder files, and never build a library just to read it.
Improved
- README rewritten as an executable path: five-step quick start, what the first round returns, where your asset library lives, and what the plugin deliberately does not do.
docs/DETAIL.mdgained a "where your asset library lives" section that separates the runtime library from project folders.- The limitations section is harder to misread: it now says plainly that nobody watched the film straight through, and that library reuse has never been verified.
Fixed
locale/zh.jsonandlocale/en.jsonadvertised "three method skills" — the 0.1.0 architecture. The real number is thirteen.docs/DETAIL.mdstill claimed "not yet tested … next step is post-install verification", while 0.2.0 had long since completed a full workflow on a real film.docs/DETAIL.mdcounted 59 on-demand references; the real number is 48.- Three dead links pointing at
planning/夜间产出报告.md, a file that is not distributed with the package.
Important Notes
- No capability change, so no migration. Skill bodies take effect immediately after upgrade;
cordis.patch.ymlis unchanged, so no reinstall or restart is required. - The asset library and any footage are not distributed with the package. Your own library grows from scratch when your first film starts.
- The Node.js requirement is now declared in
package.json: 22.3 or later. It was missing before, even though the preset relies onprocess.getBuiltinModule.
v0.2.0 · 13 个 Skill 的视频创作 Preset
首个 Skill 形态版本。此前仓库内容是旧工作台实现,本版为产品重写。
内容
- 1 个 Preset(视频创作):装配读写文件、跑命令、长任务、按需加载方法所需的插件;身份与协作原则写在 persona 里
- 13 个 Skill:意图澄清、方案与镜头、文稿与口播、视觉与版式、动效、生成提示词、代码视觉、剪辑与字幕、声音、审片取证、质量与合规、资产库、技能创作
- 48 份按需参考 + 68 个测试用例
evidence工具:抽帧、确定性测量(时长/黑帧/响度/静音)、按镜头与时间点找回证据
本版修复(基于一次 5 分 12 秒长片的全量实测)
- Preset 增加
plugin_manager,装完可自检装配 evidence measure的revision改为缺省r-unset(此前 schema 未标必填而脚本拒绝,首次测量必失败)- 镜头时长以实测为准:规范与
video-plan增加"配音实测后回填镜长并复核动画覆盖"的对账点 - 节奏结论必须先写口径:观察清单规定采样分辨率、分块方式、变化门限、判定规则四项
- 48 处原始资产引用统一为工作区根相对路径,全部可解析
已知限制
- 仅一次全量实测,没有 A/B 对照,不能声称优于直接用提示词
- 听感与连续观看未验证;生成模型、Remotion、Canva 未测
- 平台侧 AI 标识与事实核实不在本包范围
MIT 许可。安装:以本地目录安装本仓库,新开会话选择"视频创作" Preset。