@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/hand-drawn-explainer-video-nikola/references/style-guide.md

@@ -1,82 +0,0 @@
-> 范围：本文件服务于原仓库的生图/Flow 提示词模式。程序动画模式以 SKILL.md、DESIGN.md 和用户已确认样片为准；不要套用这里的禁标签卡、静止小关键词规则。
-
-# Q 版蜡笔（Q Doodle）风格与运动参考
-
-这是可选的文字风格规范，当前未附示例图或已验证动画工程。只有选用本风格时，才沿用下面的配色、笔触和运动约束；用户指定新参考图时按新图调整相应视觉约束，不混入相互冲突的固定风格词。如何选择和保存风格见 [style-selection-and-updates.md](style-selection-and-updates.md)。
-
-## 视觉签名
-
-每条生图提示词都重复以下核心约束，避免批次漂移：
-
-```text
-modern Q-version hand-drawn crayon illustration, vertical 9:16, solid warm-white canvas background, exact base color #F8F6EF, only extremely subtle low-contrast paper grain, thick imperfect black hand-drawn marker/crayon outlines, bold flat wax-crayon blocks, sunflower yellow, saturated cobalt blue and vivid tomato red with only a small muted-green accent, natural restrained Q-version proportions, simple hand-drawn composition, 2–4 large readable visual groups, generous breathing room, no glossy rendering, no realistic lighting, no logos, no watermark, no photorealism, no 3D
-```
-
-## 固定背景色
-
-所有静帧和视频尾帧使用同一背景约束：
-
-```text
-Use a flat, uniform warm-white canvas background with exact base color #F8F6EF. Preserve the same background color across every shot. Add only extremely subtle low-contrast paper grain. No black background, no gray-brown tint, no yellow-beige color cast, no gradient, no vignette, no colored paper, no border shadow.
-```
-
-Flow 的自然语言提示不能保证真正的像素级色值。若必须完全一致，使用一张 `#F8F6EF` 的固定 First Frame，或在生成后把画布背景统一替换为该色号。
-
-## 中文关键词模式
-
-当用户希望关键词直接出现在卡片或视频画面中，使用短词而非长句，并指定具体文字、位置和安全区：
-
-```text
-Show the exact Chinese keyword “好奇心” as small, clean, high-contrast hand-drawn lettering directly on the warm-white illustration paper near the upper-left margin, not inside a card, label, sticker, note or text box. Keep it perfectly legible and unchanged throughout the shot, and do not add any other text.
-```
-
-文字优先直接写在暖白纸面、留白处或相关物件旁，禁止使用便签纸、标签卡、贴纸、圆角文字框、标题面板、独立白色卡片或底部字幕框。禁止放在底部字幕区，避免放在人物脸部、快速运动物件或复杂背景上。文字块不超过画面高度的 10%–12%，以“小而看得见”为目标。动态图中关键词本身必须保持静止、清晰、不变形。Flow 仍可能生成错别字，因此 2–8 个汉字是默认上限；公司名、金额、日期和长句应提供无文字安全版备用。
-
-人物头部可略大、身体略小，但不能巨头小身、凸眼或儿童卡通。人物用发型、眼镜、服装、姿势和道具做身份锚点，不要求写实肖像。物件允许比人物夸张，笑点来自动作、比例与物件关系。背景保持干净，不加复杂房间、赛博 HUD、复古报刊、泛黄旧纸、棕灰低饱和、细密交叉排线或假文字。
-
-## 构图规则
-
-- 画面为 9:16 竖版，主体组通常占画面宽度 60%–70%，顶部和左右保留小型关键词安全区，底部预留字幕区且不放大字。
-- 每镜头使用 2–4 个主要视觉组；手机端一秒可读。避免过高的多层堆叠塔、过多小人物、复杂机械细节和海报式信息堆叠。
-- 用明确的左/中/右、前/后、上/下关系写提示词，不只写抽象情绪。
-- 真实 Logo、公司名、人物名、日期、金额和长句优先由后期层处理；短中文关键词可直接手写在画面纸面、留白处或相关物件旁。无文字版本保留干净留白，不生成便签纸、标签卡或文字容器。
-
-## 真实实体锚点
-
-生成涉及真实实体的画面时，使用以下规则：
-
-```text
-For an exact national flag, preserve the official country name, flag colors, geometry, emblem placement and aspect ratio. Do not substitute a similar flag, random flag or abstract color pattern.
-For a real company, use its industry, products, hardware and restrained brand-color cues. Do not invent or deform the exact logo unless an approved reference image is supplied.
-For a famous person, use a clearly stylized non-photorealistic chibi caricature with public identity cues such as hairstyle, glasses, clothing and posture. Do not bypass safety restrictions or attempt a photorealistic face clone.
-```
-
-如果国旗或 Logo 是镜头的核心信息，提示词必须建议使用参考图或后期叠加原始素材；生成模型只负责周围的 Q 版构图和动作。实体准确性不能靠“更详细的形容词”保证。
-
-## Flow 图生视频运动语法
-
-默认提示词可按以下顺序组织：
-
-1. `Start from the supplied completely blank #F8F6EF warm-white paper frame.`
-2. 保持 `locked flat frontal camera`，先让基础物件滑入或以不规则蜡笔线出现。
-3. 人物和次要物件作为 `rigid flat paper cutouts` 依次弹入并轻微 settling bounce。
-4. 关系元素最后出现：例如放大镜对准电表、砝码压住云、硬币接近嘴、手按下开关。
-5. 最后约 0.8 秒 `settle into and hold the exact supplied final illustration`。
-
-统一写明：`tactile 10–12 fps paper stop-motion`、无镜头移动/缩放/视差、无液体变形、无口型、无新增角色、无新增 Logo、无水印、无声音；启用中文关键词时写明关键词直接留在纸面或物件旁、固定、清晰、不变形、不新增其他文字或文字容器。
-
-Flow 使用 First + last 时：First Frame 是完全空的暖白纸，Last Frame 是对应编号完成静帧；提示词必须明确“最终构图与风格参考”。如果只能使用单张完成图，改成：
-
-```text
-Animate the supplied completed illustration in place for about 5 seconds while preserving every drawn shape and the exact composition. Use only restrained paper-cutout micro-motion: one or two small object bounces, a short hinge-like hand movement, subtle wheel rotation or one accent appearing. End on the unchanged supplied illustration and hold it still for the final second.
-```
-
-## 视觉隐喻速查
-
-| 口播关系 | 优先隐喻 |
-|---|---|
-| 竞争/跑输 | 赛跑、追逐、拔河 |
-| 成本/吞噬 | 漏水、重物下压、吞金兽、计价器 |
-| 验证/财报 | 考试、试金石、体检、放大镜 |
-| 选择/结论 | 岔路、天平、开关、两扇门 |
-| 风险/压力 | 裂缝、警报、摇晃积木、倒计时 |

package/skills/hand-drawn-explainer-video-nikola/references/style-selection-and-updates.md

@@ -1,39 +0,0 @@
-# 参考图、风格选择与动画示例维护
-
-用户指定参考图、切换画风、询问现有风格或要求更新示例时读取本页。只制作本次视频，不等于授权修改以后默认风格。
-
-## 当前实际资产
-
-| 参考 | 说明位置 | 当前具备 | 当前不具备 |
-|---|---|---|---|
-| 已确认程序动画 | [示例工程](../examples/program-animation/skill-demo/README.md) | SVG/HTML/GSAP 工程、旁白、时间轴、PNG 封面和 14 秒预览 MP4，可按需重新渲染 | 完整 36 秒公开 MP4；仓库保留的是 14 秒轻量预览 |
-| 已确认多幕逐笔故事 | [《约法三章》](../examples/stroke-story/yuefa-sanzhang/README.md) | 16:9 手速平滑版 MP4、七幕检查图、时间轴、验证报告和可编辑工程包 | 通用历史人物模板；案例资产不固定新项目的故事和角色 |
-| 已确认双语义岛逐笔故事 | [代表镜头](../examples/stroke-story/steve-jobs/README.md) | 16:9 批准源图、双区标注、区域预览和约 8.9 秒实际样片 | 通用人物模板；该资产不固定故事、角色或时长 |
-| Q 版蜡笔（Q Doodle） | [style-guide.md](style-guide.md) | 粗黑线、蜡笔色块、人物比例、配色、构图和 Flow 动效提示词 | 示例图、素材包、已验证动画工程 |
-
-没有预制的多风格图库。以后增加实际资产时同步更新此表和主文件中的资产状态，不把计划中的文件登记成已经存在。
-
-## 按参考图制作
-
-先实际查看用户提供的图片；图片未附上且没有可访问路径时，请用户上传，不能自行猜画风。区分图片用途：只借用画风、保持指定角色，还是让这张图本身运动。已能从用户要求判断时无需再问。
-
-在当前工程 DESIGN.md 中分别记录：
-- 画风：参考图来源、线条粗细与质感、配色、背景、人物比例、角色锚点和构图留白。
-- 动效：沿用哪份示例，或新的出现顺序、笔迹揭示、关节动作、节奏与转场；一张静图不能证明原作者的运动方式。
-- 声音：沿用当前授权音色；换画风不触发换音色或重做内容一致的旁白。
-
-选择适合素材的制作方式。程序矢量图适合独立物件动画，但不能假装完整保留蜡笔纹理；位图可拆层、补画遮挡部分、准备多姿态，再配合动画。自然连续人物动作可能需要图生视频服务，先核实实际接入与成本。生成或编辑位图时使用可用的图像生成工具并遵守其参考图要求，不将图生视频提示词冒充已经完成的成片。
-
-先制作代表静帧和短动画验证人物一致性、笔触、字幕安全区与动作。不保证一张参考图能自动变出整片所需所有姿态或逐像素一致。用户只要样片则交付后等待；已授权整片则按主流程内部验证后继续，只有实际缺少必要决定时才询问。
-
-## 后续优化动画示例
-
-1. **确定修改范围**：是本条视频的局部调整、为当前默认示例升级，还是新增另一套可选风格。只改本条时保持 Skill 资产不变；“新增一套”不自动替换默认。
-2. **保留可回退版本**：在本次项目工作区复制旧工程并保存输入、版本或摘要，在副本上改。不要先覆盖 Skill 中唯一的已确认示例。
-3. **把效果和方法一起保存**：至少保留实际参考/代表画面、可编辑工程、DESIGN.md 与可播放短样片；动画规范记录独立元素、时间顺序和转场做法。不要只有截图或成片、丢掉可修改的做法。没有制作某类文件就明确标为缺失。
-4. **真实验证**：用新版本渲染短样片，检查代表动作、转场和字幕；同时检查原来正常的主要动作未被改坏。环境和媒体验收分别遵循 [preflight-and-recovery.md](preflight-and-recovery.md) 与 [quality-and-delivery.md](quality-and-delivery.md)。
-5. **按授权落到技能**：已有明确“更新此默认示例/保存为预设”的授权，验证后直接安装；仅展示新样片或讨论可能性不视为替换默认的授权。新增风格可在 assets 下用独立命名目录保存图片、样片与工程，规则放 references。目录名和文件按实际需要创建，不生成空预设。
-6. **同步入口和状态**：更新本页索引和 SKILL.md 的链接；默认配置确有变化时更新 preferences.json。新增可选择的能力才更新 description 和 agents/openai.yaml；每次微调动作不必改简介。修改现有要求，删除被取代的规则，避免叠加“以前作废”的补丁。
-7. **确认实际安装与移交**：验证相对链接、Skill 结构及实际安装文件与通过测试的版本一致；交付可预览样片和回退包，说明哪些风格已经可运行、哪些仍只有规范。
-
-description 是能力和适用场景摘要，不是版本号或更新日志。判断是否更新，需要核对实际安装文件和对应资产，不能只看简介有没有变化。

package/skills/hand-drawn-explainer-video-nikola/references/upstream-whiteboard-review.md

@@ -1,30 +0,0 @@
-# srt-whiteboard-animation 上游借鉴记录
-
-复核来源：<https://github.com/geeklee/srt-whiteboard-animation>，以及该仓库尚未合并的 PR #4。上游代码遵循 MIT License；复制代码时保留原许可证和署名。本记录用于后续维护时避免重复调研，也避免把实验性改动直接当成稳定默认值。
-
-## 已吸收
-
-- 字幕语义驱动的 `sequence` / `startMs` / `protectedRegions` 编排；
-- 持久画布和区域内 stream 连续笔迹，采用 `ink → color`；
-- 渲染前标注预检：画布、坐标、保护区、时序窗口和凝视时间；
-- 空允许掩码仍写满原时间窗，避免后续画面提前；
-- `--hand-height` 与 `--hand-follow` 分开控制，分别解决遮挡和移动观感；
-- 不同尺寸、帧率和时间基准的片段统一重编码后合并。主 Skill 通过 `finalize_stroke_video.py` 实现这一点。
-
-## 已用自己的方案覆盖
-
-- 上游按单条 SRT 合成 edge-tts；本 Skill 保持用户选择的整段配音和真实音频时间轴，不因画面窗口过短强制加速语音；
-- 上游主流程偏 25–35 秒单幅白板；本 Skill 对多事件故事优先拆为 4–8 秒连续横屏场景，以保持人物尺度和故事性；
-- 上游基础合并器负责静音画面；本 Skill 的最终收尾同时处理批准源图补全、旁白、ASS 字幕和统一时间基准。
-
-## 不设为默认
-
-- 小黑风格、极简线稿：可作为用户指定画风，不覆盖已确认的自然肤色 Q 版人物；
-- 每幕强制标题和 2–4 个手写要点：知识讲解可选，历史故事默认由画面和字幕表达；
-- 默认封面：只有发布需要时增加；
-- 橡皮擦幕间过渡：先做样片。它可能引入第二只大手并打断故事连续性；
-- 自动把语音压进字幕窗口：默认延长画面或调整分镜，不牺牲已确认配音的故事感。
-
-## 后续更新判断
-
-查看上游更新时先区分 `main` 与未合并 PR。只吸收能够用真实样片验证、且不改变已确认画风和配音的机制。任何新默认项至少通过：标注错误用例、空掩码用例、复杂群像样片、多幕合并和最终音画检查。

package/skills/hand-drawn-explainer-video-nikola/references/voiceover.md

@@ -1,37 +0,0 @@
-# 配音与声音对齐
-
-## 默认与复用
-
-面向成片时默认使用火山引擎语音合成的刘飞音色 `zh_male_liufei_uranus_bigtts`，资源模型为 `seed-tts-2.0`。优先复用与稿件一致、质量合格的完整音轨；需要新合成时沿用刘飞，只有用户明确指定才改用火山引擎中的其他已授权音色。不因换画风自动换声音。检查完整音轨是否已存在且稿件匹配，避免无谓收费。
-
-电脑系统朗读、edge-tts 或其他低质量本地文本转语音不作为默认成片替代方案。若没有火山引擎授权或用户提供的合格音轨，继续完成不依赖声音的工作并明确标记配音缺口；不要静默换音色，也不要把临时机器朗读交付为已完成旁白。用户明确要求仅做低成本草稿试听时，可以使用临时声音，但必须标为占位音轨且不能覆盖已确认的刘飞资产。
-短讲解一般一次合成整稿，听感会比每句单独请求更一致。超过当前服务限制时按语义自然段拆分，并保持同一声音和设置；1500 字符是本地助手的保守限制，不是声称官方永久限制。
-
-## 安全配音工具
-
-`scripts/volcengine_tts.ps1` 使用固定官方接口，保存输入摘要、音色、音频摘要和完整性元数据；相同输出且输入匹配时复用，不匹配时拒绝覆盖。缺少完整结束标记不保存为成功，不自动重发扣费请求。
-
-```powershell
-pwsh -File <skill>/scripts/volcengine_tts.ps1 -TextFile <project>/narration.txt -OutputFile <project>/assets/narration.mp3 -DryRun
-pwsh -File <skill>/scripts/volcengine_tts.ps1 -TextFile <project>/narration.txt -OutputFile <project>/assets/narration.mp3
-```
-
-可使用 Windows PowerShell 5.1（脚本保存为带 BOM 的 UTF-8）或 PowerShell 7。`-DryRun` 仅检查输入和计划，不读取密钥、不发请求、不收费。省略 `-Speaker` 时读取公开默认刘飞音色；用户明确选择其他火山音色时再传入 `-Speaker <音色ID>`。现场合成仅在用户要求/既有授权范围内执行。
-
-凭据：优先使用已授权的 `VOLCENGINE_TTS_API_KEY` 环境变量；否则同一 Windows 用户的 `%LOCALAPPDATA%/CodexVoice/volcengine-key.dpapi`。不得把密钥放入脚本参数、技能配置、输出 JSON 或 ZIP。若密钥曾在对话公开，提醒用户轮换；不能擅自撤销它。
-随附脚本当前实现 `https://openspeech.bytedance.com/api/v3/tts/unidirectional`、`X-Api-Key`、资源 ID 和请求 UUID，并校验完整结束标记。提供商接口、资源 ID 和音色可能变化；首次调用或报协议错误时查当前官方文档，不猜参数，也不要为验证接口盲目产生付费请求。
-配音缓存必须同时匹配原稿摘要、音色、模型资源、采样参数和音频文件摘要；已有文件不匹配时拒绝覆盖。旧音轨没有缓存元数据时，人工核对稿件、音色与完整性后直接引用，不伪造请求记录或重新收费合成。
-
-## 对齐
-
-独立 ASR 复核前先执行 [预检](preflight-and-recovery.md)：模型文件非空、独立环境加载和短音频推理全部通过后再跑全片。使用服务原生词时间戳时也检查非空、顺序、边界和原稿对应关系；不要因 ASR 不可用而重新收费生成声音。
-
-优先级：真实有效词/句时间戳 → 原稿与音频强制对齐 → 句级 ASR 校对 → 实测停顿辅助人工/音频复核。空 `words` 数组表示没有词时间戳。
-静音检测能找到停顿，不能识别停顿前说了哪句话；只能作为辅助。词时间不可靠时采用完整句字幕，不伪造逐字卡点。
-对原稿中的英文、数字、多音字和最后一句特别核对。听起来漏字/错读时应修配音，不能只把字幕“改对”掩盖声音错误。阅读版“Skill”与合成读音替换要分别记录。
-
-## 自然度与混音
-
-先保持整段上下文、标点和稳定语速，再考虑风格参数。不要用任意“情绪参数”或逐字切割制造自然感。整段仍生硬时只试一小段对比，保留原音色。
-不因补动画而重新生成已通过配音。BGM/音效是可选项，必须来源清楚、版权适用、低于旁白且不遮字；不能把网上可下载当作无版权。
-检查静音、削波、爆音、音轨总长、结尾截断与句间音量。音频与视频末尾的少量编码填充允许在帧级容差内，但不能因此漏掉正文。

package/skills/hbg-life-simulation/SKILL.md

@@ -1,285 +0,0 @@
----
-name: hbg-life-simulation
-description: "Create Chinese HBG “模拟人生 / 人生副本” narrative videos with a consistent comic IP, a rapid multi-life opening, continuous natural-speed narration, synchronized short captions, dense static manga storyboards, audible fixed-level BGM, alternating zoom/pan motion in AICUT, and a consent-gated localhost-only account-backed Images API through CLIProxyAPI as the preferred image channel. Use for 模拟人生、人生副本、统一主角漫画、快闪职业开头、旁白配图、静态漫画分镜、zoom 推拉、左右平移、Edge TTS、字幕压制、本地反代自己的 Codex 账号、CLIProxyAPI，或把长篇中文人生故事组装成视频。 Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "Mr-funny/hbg-life-simulation"
----
-
-# HBG Life Simulation
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-Build a repeatable Chinese narrative-video pipeline: preserve the story, lock character identity, create an HBG opening, align narration, generate sufficiently dense comic frames, add restrained camera motion, mix audible BGM, and deliver a verified MP4.
-
-## Required skill order
-
-1. Read `/imagegen` before generating or editing raster art.
-2. Read `/hyperframes`, then the required HyperFrames domain skills before authoring or rendering.
-3. Read `/media-use` for narration, captions, BGM, or SFX.
-4. Read `/code-video-visual-qa` before the final visual pass.
-
-## Required project artifacts
-
-Keep these files in every project:
-
-- `SCRIPT_SOURCE.md`: verbatim user copy.
-- `PROJECT_SPEC.json`: story-specific title, chapters, corrections, narration settings, opening assets, flash lives, and audio paths; never hard-code these values in builders.
-- `SCRIPT.md`: narration-ready copy; record meaningful corrections at the top.
-- `CHARACTERS.md`: immutable identity, clothing, age, and relationship rules.
-- `STORYBOARD_BASE.json`: semantic beats with narration-relative `cueTime`.
-- `STORYBOARD.json`: final global timing and asset paths.
-- `PROMPTS.md`: exact ImageGen prompts and call identifiers.
-- `HBG_STYLE.json`: single source of truth for orientation, canvas, caption style/safe area, opening motion, approved opening path, and BGM baseline.
-- `audio_meta.json`: real durations and opening/body timing.
-- `assets/generated/anchors/`: approved identity/style anchors.
-- `assets/generated/sheets/`: original generated sheets.
-- `assets/generated/scenes/`: split final scene images.
-- `qa/`: source snapshots, encoded frames, media probes, and mix previews.
-
-Never silently rewrite plot, viewpoint, tone, ending, or character motivation.
-
-## Workflow
-
-### 0. Select the image channel before project work
-
-Read [references/local-account-api.md](references/local-account-api.md) and [references/image-generation.md](references/image-generation.md). Run this read-only readiness check before initializing a new HBG project:
-
-```bash
-scripts/setup_local_cliproxyapi.sh ready
-```
-```
-Prefer the user's own localhost CLIProxyAPI account channel, but require explicit confirmation for each new project before consuming account capacity. Never treat invoking this Skill as consent to use the account, install software, start the proxy, start OAuth, log in, or repair an account session.
-
-- If the result is `ready`, tell the user a local account API is available and ask whether to use it for this project. Select `local-account-api` only after the user agrees; do not manually open, print, or summarize its protected environment or OAuth files.
-- If the result is `not-running`, ask before starting the configured local proxy.
-- If the result is `not-configured` or `unhealthy`, explicitly ask whether to initialize or repair CLIProxyAPI with the user's own account. State that CLIProxyAPI is a third-party proxy, the user must complete browser/device login personally, the flow remains subject to account permissions, quotas, and provider terms, and the setup includes a models check plus one image smoke test that consumes account capacity.
-- Do not start image generation, run `start`, `install`, `bootstrap`, `login`, or an OAuth-capable installer command until the user clearly agrees in the current conversation. Do not infer consent from the story request or Skill invocation.
-- Never ask the user to paste an account password, Cookie, OAuth record, access token, or API key. Never inspect `auths/`, `client.env`, browser storage, cookies, or provider account data.
-- After consent, let the user complete OAuth in the browser or device-code page. Run only non-secret readiness/models checks and the disclosed single-image test.
-- If the user declines, the account lacks image entitlement, or the local channel cannot be made healthy without further account action, select `built-in-imagegen` and continue.
-
-Use this consent prompt or a concise equivalent:
-```
-```text
-是否使用你自己的账号在本机初始化 CLIProxyAPI，作为本次人生副本的默认生图接口？
-
-这会安装第三方 CLIProxyAPI，并打开浏览器或显示设备码，由你本人完成登录。Agent 不会读取账号、Cookie、OAuth 文件或密钥。初始化后会检查 models，并生成 1 张测试图验证图片权限；这会消耗你的账户额度，且仍受账户权限、配额和服务条款限制。
-
-是否继续？
-```
-When the result is already `ready`, use a shorter prompt that still discloses account usage:
-```
-```text
-检测到你已配置好的本地 CLIProxyAPI。是否使用你自己的账号为本项目生成图片？这会消耗你的账户额度，并仍受账户权限、配额和服务条款限制。
-```
-Record only the selected channel, consent state, provider label, model, and concurrency in `PROJECT_SPEC.json`. Never record an endpoint, key, OAuth path, token, Cookie, or account identifier.
-```
-
-### Initialize the project style
-
-Before step 1, initialize and read the style system:
-
-# Default: landscape 16:9, 1920×1080
-node scripts/init_project_style.mjs PROJECT_DIR
-
-# Only when the user explicitly asks for portrait / 9:16 / vertical video
-node scripts/init_project_style.mjs PROJECT_DIR --orientation portrait
-```
-If the user does not specify orientation, always use landscape. Portrait is supported but never inferred from short-video subject matter, platform, or a previous project. Read [references/style-system.md](references/style-system.md). Do not author separate HTML and ASS style constants, and never copy raw landscape caption/crop coordinates into portrait or vice versa.
-
-Create `PROJECT_SPEC.json` before running any builder. Read [references/project-spec.md](references/project-spec.md). Story-specific data belongs in this file or the semantic storyboard, not in JavaScript. Run the shared builders from the project directory:
-```
-```bash
-node SKILL_DIR/scripts/build_script.mjs
-node SKILL_DIR/scripts/build_storyboard_base.mjs
-node SKILL_DIR/scripts/build_narration.mjs
-node SKILL_DIR/scripts/build_composition.mjs
-```
-
-### 1. Normalize and estimate
-
-Preserve the source first. Correct only unmistakable transcription errors, broken pronouns, punctuation, or domain terms. Keep second-person narration unless asked otherwise.
-
-Write correction `from` values as the smallest unique source fragment. Do not span optional punctuation or paragraph line breaks when two smaller replacements are safer; source copy often contains transcription-inserted newlines, so long literal replacements create avoidable validation retries.
-
-Estimate duration before generating art. Use 3.4–4.2 Chinese characters per second as an early estimate, then replace estimates with real TTS/VTT timing.
-
-### 2. Lock the visual IP
-
-Read [references/visual-system.md](references/visual-system.md). Derive project-specific identities in `CHARACTERS.md`; never reuse characters from an example story. Generate and approve the required character anchor or anchors before story scenes. Repeat immutable facial traits in every later prompt; never rely on names alone.
-
-### 3. Build the HBG opening
-
-Read [references/opening-system.md](references/opening-system.md). Treat the opening as a fixed module inside the full video, not as the whole skill.
-
-Keep the mandatory order:
-
-1. Narrate only `今天体验的人生副本是……`.
-2. Immediately run ratchet SFX and rapid alternate-life flash together.
-3. Stop on the selected-life image.
-4. Narrate the complete selected-life title while that image remains visible.
-5. Start body narration only after the reveal finishes.
-
-Render Chinese titles in HTML, never inside generated images.
-
-After the opening preview is approved, bind its exact path in `audio_meta.opening.previewVideo` or adopt it to the canonical `HBG_STYLE.json` path. Full rendering must fail when that approved file is absent; never silently reuse an older preview or the raw flash/SFX video.
-
-### 4. Generate narration and timing
-
-Default to one continuous full-story Edge TTS asset at `+0%` rate. Keep only the opening lead and selected-life reveal as separate TTS assets. Preserve the VTT and treat it as the master timeline.
-
-Treat line wrapping in `SCRIPT.md` as editorial formatting, not speech direction. Before sending the chapter body to Edge TTS, join wrapped lines without inserting a pause and rely on sentence punctuation for phrasing. Never let a newline divide a word, amount, date, measurement, or sentence. Add missing punctuation such as a colon before a spoken list as a recorded source correction rather than letting a line break imitate punctuation.
-
-Derive scene starts, chapter starts, and captions from real cue times. Never distribute scenes evenly after audio exists.
-
-Burn in short synchronized captions, but preserve complete semantic phrases. Use Chinese word segmentation and punctuation boundaries before considering character count. Landscape captions may normally use 8–18 Chinese characters; portrait normally uses 8–14. A complete short utterance such as `你点头` may remain short, but a split operation must never leave a one-to-five-character suffix such as `起来`, `柜上`, `栗子`, `清楚`, or `一步`. Keep one line where possible and no more than two lines. Remove sentence-ending punctuation when the spoken pause already supplies it. Read the values from `HBG_STYLE.json`; do not duplicate them in a composition builder. Use the orientation-specific caption baseline from [references/style-system.md](references/style-system.md): landscape defaults to 48 px with a 58 px bottom margin at 1920×1080; the validated portrait baseline uses 48 px with a 280 px bottom margin at 1080×1920. In ASS, `BorderStyle=3` requires non-zero box padding/outline; zero makes the encoded background effectively disappear. Keep the complete caption box inside the safe area for the selected orientation.
-
-Run `node scripts/audit_caption_semantics.mjs PROJECT_DIR` after narration generation. Do not build the composition when it reports a short split tail, an undersized generated fragment, or a narration line that ends without sentence punctuation.
-
-### 5. Enforce storyboard density
-
-Read [references/storyboarding.md](references/storyboarding.md). Select semantic beats, then audit every calculated scene duration.
-
-- Target ordinary still duration: 4–8 seconds.
-- Allow 8–12 seconds only for a deliberate emotional hold or complex visual reading.
-- Split ordinary scenes longer than 12 seconds.
-- Treat any scene longer than 16 seconds as a mandatory missing-image defect.
-- For an 8–12 minute narration, expect roughly 70–110 stills unless the user explicitly chooses a sparse style.
-- For other lengths, size the first storyboard from real TTS duration instead of reusing the 8–12 minute count: start near `ceil(bodyDuration / 6.5)` stills, with a practical range of `bodyDuration / 8` to `bodyDuration / 5`. A 14–15 minute story normally needs about 125–165 stills. Place extra beats proactively inside long explanatory paragraphs, repeated financial details, internal-identity conflict, and denouement passages before generating art.
-
-Add a new still when location, time, dominant action, emotional power, focal object, or memory state changes. Do not hide sparse coverage with an extra-long zoom.
-
-Before building API prompts, write `participants.count` and `participants.allowed` explicitly on every semantic scene. Do not infer participant count later from a keyword heuristic: pronouns such as `他`, `对面`, or `当着他的面` are easy to miss and can turn a two-person action into an incorrect solo frame. Treat missing participant metadata as a planning defect.
-
-### 6. Generate scene art
-
-Use the image channel selected in step 0. Generate strict 2×2 sheets only when four related, low-risk beats share characters and setting. Require equal panels, uniform black gutters, no text, no captions, no speech bubbles, no logos, and no watermark. Generate high-risk frames as standalone images rather than grid panels: hand close-ups, overlapping limbs, phone front/back orientation, typing, smoking, chopsticks, soldering or other tool contact, and identity-critical hero shots. If a grid panel fails one of these checks, regenerate that beat alone; never rescue it by cropping around the defect.
-
-Opening flash sheets must preserve the target orientation at the panel level. For a landscape project, use standalone landscape images or strict 2×2 sheets whose individual panels are landscape; never split a 5×2 landscape sheet into ten portrait cells and then enlarge them with `object-fit: cover`. Resolve every final flash asset to the exact canvas size before composition building: 1920×1080 for landscape or 1080×1920 for portrait. `validate_style_system.mjs` treats a mismatched flash asset as a hard failure.
-
-The user-owned localhost CLIProxyAPI path is the preferred default after the consent and readiness gate in step 0. Built-in ImageGen is the fallback after decline, failed entitlement, or an unavailable local channel. A user-approved CLIProxyAPI localhost endpoint may supply the bundled imagegen CLI path only after `/v1/models` and one disclosed image smoke test pass. Never store an API key in the project, skill, prompt JSONL, shell history file, or report. After the anchor is approved, independent sheets may run concurrently; default to 5 and cap concurrency at 10. Treat one sheet or standalone frame as one generation job. API output must still be displayed visibly in chat, saved project-locally, inspected, and counted.
-
-For API batches, create `assets/generated/prompts/SHEET_MAP.json` before generation. Map every job to final scene IDs in panel order and prove that every storyboard scene is covered exactly once. Generate into a dedicated staging directory. Some compatible providers or CLI versions flatten `out` subdirectories to basenames when `--out-dir` is used; never assume `sheets/g01.png` will remain nested. Resolve staged files by verified basename, reject collisions, then copy sheets and singles into canonical project directories.
-
-Do not trust the requested API size as the delivered pixel size. Probe every output and normalize every final scene to the exact canvas (`1920×1080` landscape or `1080×1920` portrait) before composition building. Split sheets only after confirming four independent panels and uniform gutters.
-
-Every prompt must declare the exact number of visible people and which characters are allowed. Do not paste the entire supporting-character roster into every prompt: models may insert every described character. Add an explicit prohibition such as `No other person may appear` for identity-critical and high-risk frames.
-
-After every call:
-
-1. Show the generated sheet visibly in chat.
-2. Save it under `assets/generated/sheets/`.
-3. Inspect identity, anatomy, real-world object orientation and use, gutters, and panel order.
-4. Split it with `scripts/split_2x2.sh`.
-5. Inspect the split images before adding them to the storyboard.
-
-Reject physically impossible anatomy or props even when the image is otherwise attractive. Before judging pose quality, count each visible person's arms, hands, palms, and fingers; confirm every hand belongs to one arm, has the expected number of digits, and remains spatially distinct under overlap. For every hand close-up, inspect the source at 100%: trace each fingertip backward through one palm, one wrist, and one forearm. If overlap makes the number or ownership of hands impossible to prove, reject the frame instead of assuming the anatomy is hidden correctly. A two-hand contact shot must show exactly two wrist-to-palm chains and must not contain a third palm-shaped mass, duplicated finger fan, or finger cluster emerging from beneath another hand. Reject extra, fused, detached, duplicated, ownership-ambiguous, or impossible hands. Phones must have the screen facing the viewer/user, cameras and buttons on plausible sides, fingers wrapping the device naturally without passing through it, and the gaze aligned with the visible screen. Apply the same reality check to cigarettes, chopsticks, bowls, soldering irons, circuit boards, doors, chairs, bedding, tools, and every hand-object contact.
-
-Use separate calls for unrelated settings or identity-sensitive hero frames.
-
-### 7. Animate static frames
-
-Use deterministic linear camera motion and alternate directions:
-
-- zoom in: `scale 1.00 → 1.10–1.13` with slight x/y drift;
-- zoom out: `scale 1.12 → 1.00`;
-- pan left/right: retain `scale 1.07–1.11`, move no more than 4% of frame width;
-- emotional hold: `scale 1.00 → 1.025`;
-- use hard cuts for memory fragments and short dissolves only for emotional continuity.
-
-Animate an image child inside a full-bleed clipped wrapper. Never overlap transform tweens on the same element. Keep every image covering the canvas throughout its motion.
-
-### 8. Mix narration, SFX, and BGM
-
-Keep narration, ratchet SFX, BGM, and images on distinct tracks. Default to a constant BGM baseline with no narration-triggered ducking or repeated volume automation.
-
-Create a 15–20 second opening mix preview before rendering the full video. Start with an audible bed around 8–12 dB below narration, then adjust in 3–4 dB steps by ear. Preserve at least 3 dB true-peak headroom in the encoded mix. Use [references/opening-system.md](references/opening-system.md) for the validated HBG example.
-
-### 9. Assemble in HyperFrames
-
-Use one clip per beat or one sub-composition per chapter. Derive clip timing from aligned narration. Use HTML for titles, act labels, and captions. Generate `第一幕`, `第二幕`, and later act labels from the configured chapter count; do not assume exactly ten chapters. Avoid game HUD counters unless explicitly requested.
-
-### 10. Render long videos safely
-
-Read [references/render-and-qa.md](references/render-and-qa.md) before rendering any video longer than four minutes.
-
-1. Run `scripts/preflight_long_render.sh PROJECT_DIR`; keep at least 12 GiB free for a 1080p high-quality 8–12 minute render unless a measured project-specific estimate proves less is safe.
-2. Run `node scripts/validate_style_system.mjs PROJECT_DIR`; do not start the long render unless the approved opening, safe caption ratios, flash zoom, and selected-life zoom pass.
-3. Run `scripts/render_long_video.sh PROJECT_DIR OUTPUT.mp4`. It enables chunked encoding, raises the FFmpeg timeout, and uses VideoToolbox on macOS by default.
-4. If measured HyperFrames capture would materialize the full timeline and exceed safe disk space, validate and use `scripts/render_streaming_ffmpeg.mjs PROJECT_DIR OUTPUT.mp4`. Keep scene timing, motion, captions, chapter labels, opening, and audio mix identical; change only the rendering transport.
-5. Always render to a new revisioned filename. Never overwrite the last verified MP4.
-6. On `ENOSPC`, stop and inspect exact `renders/work-<renderJobId>-*` or `renders/ffmpeg-work-<output-name>` directories. Delete only abandoned render work directories after confirming no matching render process is alive; preserve source assets and completed MP4 files.
-7. On `ffmpegEncodeTimeout`, do not recapture repeatedly with the same settings. Keep chunked encoding enabled, use a supported hardware encoder, and raise `FFMPEG_ENCODE_TIMEOUT_MS`.
-
-### 11. Quality gate
-
-Do not deliver until:
-
-- the anchor and generated sheets are visible in chat;
-- every referenced asset is project-local;
-- the duration-density audit reports no unexplained scene over 16 seconds;
-- `hyperframes check --strict` passes;
-- source hero frames and transition midpoints are inspected;
-- encoded opening, body transitions, long holds, and final frame are inspected;
-- the HTML preview and final ASS caption path both use `HBG_STYLE.json`, and at least one encoded ASS caption frame has a visible background box inside the selected orientation's safe area;
-- every visible person has the correct count and ownership of limbs, hands, palms, and fingers, including during overlaps;
-- every hand close-up has been opened at full resolution and each fingertip has been traced to exactly one palm, wrist, and forearm; unverifiable occlusion is a rejection, not a pass;
-- every important prop is oriented and used plausibly, with special attention to phone front/back direction and hand-object contact;
-- resolution, FPS, duration, frame count, audio streams, true peak, and black-frame detection are verified.
-- the final encoded MP4, not only source snapshots, has been checked with `scripts/verify_final_video.sh` at the opening, every corrected/high-risk shot, one phone/tool shot, one body transition, and the ending.
-
-Regenerate identity-drifted sheets. Never conceal bad faces or sparse storyboarding with motion blur or speed.
-
-## Reusable helpers
-
-- `scripts/audit_storyboard_density.mjs`: report or gate storyboard holds longer than 12/16 seconds.
-- Run the density audit with the JSON path, for example `node scripts/audit_storyboard_density.mjs PROJECT_DIR/STORYBOARD.json`; passing the project directory itself is invalid.
-- `scripts/audit_caption_semantics.mjs`: reject semantic split tails, undersized generated caption fragments, and narration lines without terminal punctuation.
-- `scripts/build_script.mjs`: turn `SCRIPT_SOURCE.md` plus `PROJECT_SPEC.json` into a corrected, chaptered `SCRIPT.md`.
-- `scripts/build_storyboard_base.mjs`: validate and normalize the agent-authored semantic storyboard without embedding story beats in code.
-- `scripts/build_narration.mjs`: generate config-driven continuous Edge TTS, captions, audio timing, and final storyboard timing.
-- `scripts/build_composition.mjs`: build a config-driven HyperFrames project with dynamic chapters, opening lives, titles, BGM, and orientation layout.
-- `scripts/init_project_style.mjs`: create a locked landscape `HBG_STYLE.json` by default, or a portrait one only with `--orientation portrait`, without overwriting an existing project style.
-- `scripts/validate_style_system.mjs`: fail fast on wrong opening dimensions/duration, unsafe caption ratios, missing ASS box padding, or absent opening zoom.
-- `scripts/split_2x2.sh`: split a generated sheet into four landscape 1920×1080 PNGs by default, or portrait 1080×1920 PNGs when passed `portrait`.
-- `scripts/make_contact_sheet.sh`: create a contact sheet for review.
-- `scripts/preflight_long_render.sh`: check free disk space and list abandoned render work directories before a long render.
-- `scripts/render_long_video.sh`: run revision-safe long-form HyperFrames rendering with chunked encoding and an extended timeout.
-- `scripts/render_streaming_ffmpeg.mjs`: low-disk fallback that renders each still to a short zoom/pan segment, concatenates, burns captions, and mixes the final audio without writing every timeline frame.
-- `scripts/verify_final_video.sh`: probe the encoded MP4, scan black/silent intervals and loudness, extract named QA frames, and build a contact sheet.
-- `scripts/setup_local_cliproxyapi.sh`: check local readiness without exposing secrets, then—with explicit consent—install a checksum-verified CLIProxyAPI release outside projects, bind a user-owned account to localhost, and create a protected OpenAI-compatible client environment for batch work.
-
-## Source
-
-Rewritten for AICUT from [`Mr-funny/hbg-life-simulation`](https://github.com/Mr-funny/hbg-life-simulation)
-(MIT), SKILL.md. The method is that skill's own; the way it runs is AICUT's.

package/skills/hbg-life-simulation/references/image-generation.md

@@ -1,53 +0,0 @@
-# Image generation channels
-
-## Selection
-
-- Prefer the user's own localhost CLIProxyAPI account channel after the Skill's per-project explicit consent and readiness gate.
-- Run `scripts/setup_local_cliproxyapi.sh ready` before project initialization. This check must not print or inspect credentials.
-- Ask before using account capacity even when the result is already `ready`. If setup, start, repair, or login is needed, disclose that separately and ask before doing it. Skill invocation is not consent. Let the user complete OAuth personally.
-- Fall back to the built-in `image_gen` tool when the user declines, the account lacks image entitlement, or the local channel remains unavailable.
-- Use another OpenAI-compatible endpoint only when the user explicitly provides or selects it.
-- Read the system `imagegen` skill before either path.
-
-## Built-in fallback
-
-Issue one built-in call per sheet or standalone frame. Recover the returned image into the project, display it in chat, inspect it, then split or adopt it.
-
-## Compatible API batch path
-
-Use the bundled imagegen CLI rather than writing a new API client:
-
-```bash
-IMAGE_GEN="${CODEX_HOME:-$HOME/.codex}/skills/.system/imagegen/scripts/image_gen.py"
-OPENAI_API_KEY="..." OPENAI_BASE_URL="https://example.invalid/v1" \
-  python "$IMAGE_GEN" generate-batch \
-  --input tmp/imagegen/prompts.jsonl \
-  --out-dir assets/generated \
-  --concurrency 5 \
-  --no-augment
-```
-
-Rules:
-
-- Inject credentials only through the current process environment. Never write them into repository files, scripts, prompt JSONL, skill files, logs, or final reports.
-- For the consented CLIProxyAPI path, source its protected `client.env` outside the project. The generated `OPENAI_API_KEY` is only a localhost client credential; never substitute or expose the provider OAuth record.
-- Before a large local-account batch, require a successful authenticated `/v1/models` check and one visible single-image test. Fall back when the account lacks image entitlement.
-- Use one JSONL job per distinct 2×2 sheet or standalone frame. Default concurrency is 5; do not exceed 10.
-- Write `assets/generated/prompts/SHEET_MAP.json` before the call. It must map each job to final scene IDs in top-left, top-right, bottom-left, bottom-right order and cover every scene exactly once.
-- Specify the exact output filename, orientation, quality, panel order, anchor traits, exact participant count, allowed characters, and forbidden extra characters in every job.
-- For a 2×2 sheet, require four equal landscape panels and uniform black gutters. For a standalone frame, require one 16:9 composition.
-- Do not describe the younger brother, parents, coworkers, or other supporting roles in a prompt unless that exact frame needs them.
-- After the batch finishes, build and display contact sheets in chat. Inspect identity, anatomy, phone direction, hands, tools, panel count, and unexpected people.
-- Generate into a staging directory. Resolve outputs by verified basename because compatible endpoints or CLI versions may flatten nested `out` paths under `--out-dir`; reject duplicate basenames instead of overwriting.
-- Probe delivered dimensions. Normalize every final scene to the project's exact canvas before it enters `STORYBOARD.json`; the provider may return a nearby size even when a larger size was requested.
-- Regenerate only failed jobs with corrected participant counts or physical constraints; preserve good outputs.
-- Record successful jobs, failed attempts, regenerations, model, endpoint class, size, and concurrency in `PROMPTS.md`. Do not record the key.
-
-## Counting
-
-Report both:
-
-- successful generated assets/jobs;
-- failed or rejected attempts separately.
-
-A 2×2 sheet is one generation job and four final scene images after splitting.