01 / 已上线动画短片

三诗人

Three Poets

导演 / 编剧 / 分镜：许宸扬
原画：廖心悦 雷谨如 仇昕 蔡雨帆
美术 / 绘景：许宸扬 蔡雨帆
3D / 后期 / 制片：余亚婷
声音：许宸扬 Director / Screenplay / Storyboard: Xu Chenyang
Key Animation: Liao Xinyue Lei Jinru Qiu Xin Cai Yufan
Art Direction / Background: Xu Chenyang Cai Yufan
3D / Post-production / Producer: Yu Yating
Sound: Xu Chenyang

漫漫在理想国的大地上奔走寻找三位测量理想国的诗人。Bilibili 27万播放，5.7万点赞。

Manman running across the land of Utopia, searching for three poets. 270k views on Bilibili.

剧照 · Stills

面具

Mask

导演 / 编剧 / 分镜：许宸扬
原画：廖心悦 雷谨如 仇昕 蔡雨帆
美术 / 绘景：许宸扬 杨梓瑞
3D / 后期：余亚婷 杨梓瑞
执行制片：曾胜蓝 Director / Screenplay / Storyboard: Xu Chenyang
Key Animation: Liao Xinyue Lei Jinru Qiu Xin Cai Yufan
Art Direction / Background: Xu Chenyang Yang Zirui
3D / Post-production: Yu Yating Yang Zirui
Line Producer: Zeng Shenglan

上海GAF展映作品。成年漫漫返回过去的时间线试图改写历史。

Screened at Shanghai GAF. Adult Manman returns to the past timeline as the antagonist.

剧照 · Stills

师生

Teacher and Student

导演 / 分镜：许宸扬
编剧 / 原作：舒傲谌
原画：廖心悦
美术 / 绘景：许宸扬 舒傲谌
制片 / 后期：余亚婷 Director / Storyboard: Xu Chenyang
Screenplay / Original Work: Shu Aochen
Key Animation: Liao Xinyue
Art Direction / Background: Xu Chenyang Shu Aochen
Producer / Post-production: Yu Yating

获厦门金海豚最佳短视频动画提名奖。

Nominated for Best Short Video Animation at Xiamen Golden Dolphin Awards.

剧照 · Stills

"请您从我身上卸载真、善、美吧。"

Please Uninstall Truth, Goodness & Beauty from Me

导演 Director / 剪辑 Editing：许宸扬
原画 Key Animation：大史仔仔 · 黑色公害马铃薯 · 混沌卷饼
清上 Clean-up：混沌卷饼 · 杨梅蛐 · 飘来一朵紫云
3D / 合成 3D / Compositing：小牛半邊
绘景 Background Art：咖啡耗子药
配音 Voice Acting：Joeydat3dream · 许宸扬

漫漫手中的漫道平板内置四款软件「真」「善」「美」「愿」，本条动画介绍其中三款的功能与隐喻。Bilibili 7.8万播放。

Manman's tablet houses four metaphorical apps. 78k views on Bilibili.

"必将用劳苦的血肉，击败虚华的理想。"

Flesh and Toil Shall Conquer Vain Ideals

导演 Director：许宸扬
原画 Key Animation：炮 · 青峰 · 许宸扬 · 翘 · 小牛半邊
清上 Clean-up：翘 · panda · 小牛半邊
绘景 Background Art：panda · 许宸扬
3D / 合成 3D / Compositing：小牛半邊
配音 Voice Acting：小牛半邊 · 梁家豪 · 许宸扬
出品 Production：翟家班 Animation

《漫漫：空洞寓言》伪PV，代号「结局条」。主角漫漫直面「不像」的结局，理想与现实在此交汇。Bilibili 21.9万播放，4.8万收藏。

Pseudo-PV for Manman: Hollow Fable. 219k views, 48k favorites on Bilibili.

前司期间主导运营 · Operated during prior tenure

不像-

UID 506499149

上美影确实是我白月光。
但还可以更现代。
这里是不像！

5.7万 粉丝 Fans

294.6万 播放 Views

56.9万 获赞 Likes

36 投稿 Videos

在翟家班任职期间主导运营的原创动画账号，围绕系列《漫漫：空洞寓言》进行内容创作与发布，累计播放近300万。

Animation account operated during tenure at Zhaijiaban. Built around Manman: Hollow Fable. Nearly 3M total views.

前往空间 / Visit Space →

02 / 前期概念设计

Unreleased

我们一定要青史留名

Etched in the Stone

发起者/导演/编剧：许宸扬

通过Bilibili胶囊计划第五季初审。民国艺专的艺术与观念冲突。

Passed preliminary selection of Bilibili Capsule Program Season 5.

Unreleased

马良所画之人

A Woman Drawn by Ma Liang

发起者/导演/编剧：许宸扬

探讨生命意志与艺术再现的边界。

Exploring the boundary between life's will and artistic reproduction.

03 / 纯艺作品

暂未开放

• ❯ node execute.js --project=AnimeTrack [Click to Execute]
  // AnimeTrack · Intro Page Screenshots

  

  

  // Production Overview · Screening Room · Gantt View

  AnimeTrack 以三个核心视图构建完整的制作透视体系。制作概览实时追踪每个镜头在 Layout、原画、清线、上色、BG、3D 及合成各环节的推进状态，并展示人员分配情况，全局 Overview 面板可在同一屏内掌握整组动态。阅片室通过统一时间轴将全员上传素材串联，Layout 粗剪与清线定稿均可即时查看，实现真正动态的进度感知。甘特图在标准功能基础上加入右侧视频预览窗口，任务对应的镜头素材可直接在排期界面内预览，无需跳转确认。

  AnimeTrack delivers full production visibility through three core views: a Production Overview tracking per-shot status across all stages (Layout → Key Animation → Clean-up → Coloring → BG/3D → Compositing) with personnel allocation; a Screening Room that stitches all team uploads along a shared timeline for real-time progress monitoring; and a Gantt Chart extended with an integrated video preview panel, enabling shot review directly within the scheduling interface.

  

  

  // Linear Dependency Scheduling · In-context Sticky Notes

  甘特图的差异化在于跨任务组线性联动排期。传统工具中各任务彼此独立，调整某节点时其他任务不会自动跟进。AnimeTrack 引入"线性连接"机制，将具有前置依赖的任务绑定成链——任意节点的变动自动推移整条链路，多人协作中的排期联动从逻辑层面真正落地。

  便利贴解决了通知与任务界面割裂的痛点。飞书等平台的公告、群置顶本质上属于消息流，与制作监管页面彼此割裂。AnimeTrack 允许便利贴直接悬浮于制作总览界面，全员可见，且可精确定位至具体单元格，而非仅能对整条记录发表评论。

  The Gantt Chart's key differentiator is cross-group linear dependency scheduling. AnimeTrack's "linear link" mechanism binds sequentially dependent tasks into a chain — any change cascades automatically, making multi-person schedule alignment genuinely dynamic rather than a manual operation. The Sticky Note system closes the gap between notifications and the production interface: notes are embedded directly on the dashboard, visible to all, and anchorable to specific cells — not just record-level comments.

  

  

  // Sticky Notes Demo · Achievement System

  本组演示便利贴的完整交互流程，以及由此衍生的成就系统。效率软件的趋势之一是游戏化——在功能之外置入惊喜，让工具本身产生情绪价值。AnimeTrack 的成就系统参照 Steam 成就体系设计：统计用户的操作行为（如创建便利贴、达成进度里程碑），在特定条件触发时颁发分级勋章，让制作过程中每一个"完成"动作都有机会成为值得被标记的时刻。

  This section demonstrates the full sticky note interaction flow and AnimeTrack's Achievement System. Modeled on Steam's achievement framework, the system tracks user actions — note creation, milestone completion — and awards tiered badges when conditions are met, turning routine production checkpoints into moments worth celebrating. The underlying thesis: productivity software's next frontier is gamification, embedding moments of delight beyond pure function.

• ❯ node execute.js --project=CapStoryBoard [Click to Execute]
  // CAP StoryBoard · 视频一

  // CAP StoryBoard · 视频二

  

  

  CAP StoryBoard — CSP × 剪映 动态分镜系统

  在 CSP（Clip Studio Paint）里直接绘制分镜，同步推送至剪映时间线——让静态故事板环节与出版 Layout 环节合二为一。借助剪映的录音、文本、素材库，创作者在画分镜的同时即可感受完整的视听节奏，大幅压缩独立动画的创作周期。

  Draw storyboards in CSP, sync live to CapCut/剪映 timeline. Static storyboard meets dynamic editing — letting animators hear dialogue, music, and SFX while drawing, closing the gap between pre-production and layout.

• ❯ node execute.js --project=StoryCut [Click to Execute]
  // StoryCut · Intro

  StoryCut — 故事板切片导入工具

  在短视频竖屏动画的制作中，分镜本的导入一直是个痛点。StoryCut 通过半自动化的方式，一键将 PDF 拆解并按镜头号重命名，直接对接 Premiere 或 AE，让创作者告别手动切图的机械劳动。

  StoryCut automates the extraction of storyboard PDFs into sequentially named PNG files, ready for direct import into Premiere or AE. From 1 PDF to n frames in seconds.

  Live story.work ↗
• ❯ pnpm orchestrate --agent=BaziLifeCurves AI agent [Click to Execute]
  // 方法论 ① · /dev/roadmap 截图

  

  // 方法论 ① · 自己写 flowmap 来调自己的 agent

  开发 agent 的第一件事不是写提示词，而是把整套编排骨架画出来。/dev/roadmap 是项目内置的开发态工具——反射 prompts.ts + DELIVER_TURNS，自动生成一张可点击、可跳源、可在线打补丁的有向图，覆盖 21+ 节点 × 14 个解读轮次（含子段）× pipeline 6 个流式阶段 × 8 道审计闸。

  右侧还挂了一张 SVG 动态流程图：主轨竖线走主流程，虚线侧轨画分支——「三派分歧才走的 open_phase」「love_letter 才触发的 turn 7」「母题侧记」一眼分清；点节点跳详情，左侧时间轴自动滚到对应行。详情面板能看 system prompt 全文、一键 vscode:// 跳源，也能在线编辑——保存前先看 diff 预览，保存后自动写回源文件并留 .bak.<ts> 备份。

  大部分团队靠 LangSmith / LangFuse 这类外部观测工具。我选择反射自己的源码画图——不依赖第三方、改提示词时不用切窗、调试时不是 grep 加 console.log，而是看着流程图找出哪个节点没跑或跑错。这是一种工程哲学：承认 agent 编排会复杂到肉眼无法维护，所以在写 agent 的同时，专门为自己造自省工具。

  Step one of building any agent: draw the orchestration map. /dev/roadmap reflects prompts.ts + DELIVER_TURNS into a clickable, source-jumping, in-browser prompt-patcher — covering 21+ nodes × 14 deliver turns (with sub-segments) × pipeline 6 streaming stages × 8 audit gates. A live SVG flow chart on the right shows main track + dashed side branches (open_phase, love_letter turn 7, motif side notes); the detail panel shows full prompts, one-click vscode:// jump-to-source, and inline editing with diff preview and automatic .bak.<ts> backup on save. Most teams reach for LangSmith / LangFuse; I built it from reflection on my own source — agent orchestration grows too complex for eyeballs alone, so the introspection tool is a first-class citizen.

  // 方法论 ② · chart.html 出图 + 流式解读

  

  

  

  // 方法论 ② · 决策权留给数学，叙事权留给 LLM

  所有"未知量"都不允许问 LLM 怎么办。必须先用识别器算出先验——14 个候选命格的初始概率分布，再用证据（用户答题）一题一轮更新成后验，最后用置信带决定下一步：≥ 0.80 直接采纳 / 0.60-0.80 采纳但加保留 / 0.40-0.60 触发追问轮 / < 0.40 拒绝出图。LLM 看到的提示词永远是"已锁死的事实 + 已计算的后验"——它只负责把数字翻译成大白话。

  正因为这条，几个看似零散的设计连成一条线：--answer-source agent_inferred 永久退出（agent 自答会让似然 = 先验，贝叶斯框架失效）；后验和信息增益写到点开头隐藏的状态文件里不让 LLM 看见（看见就会编故事自洽）；算法不收敛时让用户补事件，LLM 只做"事件 → 结构化候选"的转换，重算决策的活还给 multi_school_vote.vote()。决策权留给数学，叙事权留给 LLM。

  No unknown quantity is ever delegated to the LLM. Every decision starts with a prior (a distribution over 14 candidate phases from the detector), updated turn-by-turn into a posterior via user answers, and gated on confidence: ≥ 0.80 adopts; 0.60-0.80 adopts with caveat; 0.40-0.60 triggers another round; < 0.40 refuses to render. The LLM only ever sees locked facts plus computed posteriors — it translates, never judges. This makes the rest fall into place: --answer-source agent_inferred hard-exits because LLM-generated answers collapse the Bayesian update; posterior and information gain live in dot-prefixed state files the LLM never sees, severing the channel through which it might rationalize its own guesses; when consensus collapses, the LLM only converts user-supplied events into structured candidates and hands control back to multi_school_vote.vote(). Math owns judgment. The LLM owns voice.

  

  

  // 工程实现 — 两条方法论落到代码里的样子

  ① 双层校验 + 重试反馈环：每段 LLM 输出先过 JS 快校验，再过 Python 重校验。违规理由直接拼到下一轮提示词末尾让模型自己改，最多 3 次，重试期间不污染半成品。两边校验同源，不会漂移。

  ② 反系统化铁律：同一母题在多个位置出现时，相似度必须 < 0.6——逼 LLM 换角度、换动词、换比喻，而不是复制原句。把"读起来不像模板"这种主观感觉直接写成可校验的数字。

  ③ 真·流式 agent，不是流式显示：脚本算一段就吐一段、agent 立即写 + 推送 + 暂停一轮，下次再续。deliver 阶段更狠——当前大运、其它大运、关键流年全拆成 N 个独立 LLM 调用（当前大运 1 综述 + 4 维分述 = 5 个），单图一次跑 20-30+ 次 LLM，前端能实时看见"正在写第 3 段大运"「下一个：2031 年」。

  ④ 8 道出图时审计闸：节点顺序、流式 emit、母题连续性、收尾段计数、过早决策……任一闸不过，整图拒出。

  ⑤ 8 阶段对话状态机：intake → pipeline → elicit → deliver → … → locked，只允许向前推进，所有外部请求走单一入口闸——比把对话历史一股脑塞进 messages 数组干净得多。

  Five consequences of the two methodologies above: (1) Two-layer validation with retry feedback — every LLM turn passes a JS fast-check and a Python deep-check; violations get appended to the next prompt for the model to fix, up to 3 retries, without polluting partial state. (2) Anti-systematization rule — the same motif appearing in multiple anchors must score below 0.6 paraphrase similarity, forcing the model to vary angle, verb, and metaphor instead of repeating the source sentence. (3) Real streaming agent, not streaming UI — Python yields stage-by-stage; agent writes, pushes, pauses; resumes on the next turn. Deliver stages further unroll into N independent LLM calls (current dayun = 1 overview + 4 dimensions; each other dayun; each key year), yielding 20-30+ calls per chart, with progress chips so users see "writing dayun 3 of 5" or "next key year: 2031" live. (4) Eight render-time audits — node order, streamed emit, motif continuity, closing-section counts, premature decision, etc. Any gate fails, no chart. (5) Eight-phase conversation state machine — intake → pipeline → elicit → deliver → … → locked, forward-only, single input gate. Cleaner than dumping history into a messages array.

  Live yourbazi.online ↗
• ❯ python3 run_ai_games.py --engine=Gemini3 [Click to Execute]
  // ① Sunhike

  

  

  // Sunhike · Scarab's Solar Odyssey

  灵感来自西西弗斯的故事——一只圣甲虫每天要将一颗太阳从山脚推上山顶。大山的全部光阴由这颗太阳决定：推得太快会灼伤大地、蒸干水源；推得太慢则夜晚过长，树木枯萎、万物凋零。玩家需要以恰当的节奏将太阳升起、悬停、再缓缓落下，维持整座山一天的完整日照周期。太阳坠落太快还可能摔碎——在"让世界活着"与"不把太阳弄丢"之间，找到那条微妙的平衡线，就是这个游戏的全部乐趣。

  Inspired by the myth of Sisyphus: a scarab beetle must push the sun from the base of a mountain to its peak each day. The entire mountain's cycle of light depends on this sun — push too fast and the land scorches dry; too slow and endless night starves the forest. The player must find the delicate rhythm of ascent, zenith, and controlled descent to sustain one full day of sunlight without shattering the sun on the way down.

  // ② 攀岩模拟器 · Crux Simulator

  

  

  

  // Crux Simulator · Four-Limb Independent Control

  灵感来自一次跟朋友去攀岩的经历。攀岩最迷人的地方在于四肢的独立控制——每一只手、每一只脚都需要单独发力，同时还要维持躯体平衡。于是我设计了一套极为直觉的操控方案：左键 = 左手，右键 = 右手，A = 左脚，D = 右脚，Q / E = 身体摆动。这套操作非常生动地还原了真实攀岩中"四肢各自为战、躯干居中协调"的感受。

  为了增加挑战性，我还专门开发了一个定线器，用来设计各种复杂的攀岩路线关卡，让每一面岩壁都有不同的破解方式。

  Born from a real climbing session: the game maps each limb to a separate key — LMB/RMB for left/right hand, A/D for left/right foot, Q/E for body sway — faithfully recreating the sensation of independent four-limb control and core balance. A built-in route setter lets players design arbitrarily complex wall layouts, ensuring every climb is a unique puzzle.

  AI 游戏开发 — Sunhike & 攀岩模拟器

  在 AI 爆发的时代，我尝试使用 Gemini 和 Cursor 突破代码能力的限制，将脑海中的游戏机制直接落地。这不仅是技术实验，更是对叙事媒介的拓宽——《Sunhike》用一只圣甲虫推太阳的隐喻重述了西西弗斯的故事；《攀岩模拟器》则把真实攀岩的四肢独立控制感忠实地搬进了键鼠操作。两款游戏都由 AI 辅助编程完成。

  In the era of AI, I used Gemini + Cursor to bypass traditional coding barriers and bring game concepts directly to life. Sunhike retells the Sisyphus myth through a sun-pushing scarab; Crux Simulator faithfully translates real climbing's four-limb independence into keyboard + mouse controls. Both powered by AI pair-programming.

• ❯ node execute.js --project=LittlePlan [Click to Execute]
  // LittlePlan · Intro

  // Little Plan · 全年规划工具

  市面上几乎所有日历和规划工具都在回答同一个问题：今天做什么？但我想要的是另一个维度——这一整年，我在哪里？飞书可以做到，但我不打算每天开机就登飞书翻自己的私人待办。所以我自己做了一个。Little Plan 的逻辑分两层：年度和月度用甘特图，不用日历看板——甘特图能让你看见时间的分量，而不只是一格格方块。日视图更极端：只有昨天、今天、明天，就这三天。它挂在桌面上，每天开机自动启动。每次打开电脑的第一眼，不是消息，而是自己这一年还剩下多少宏图伟业。

  Almost every calendar app answers the same question: what do I do today? I wanted a different question answered — where am I in the full year? Little Plan works in two layers: annual and monthly views use Gantt charts (not calendar boards), making the weight of time visible rather than abstract. The day view shows exactly three things: Yesterday, Today, Tomorrow. It mounts to the desktop and launches on startup — every morning, before the noise starts, you see how the year is going and what today's work means inside the larger arc.

• ❯ ssh root@feishu_workflow [Click to Execute]
  // 飞书制片表 v4.1 · AnimeTrack 的前身

  在 AI 工具尚不成熟、还没有能力直接 Wire Coding 的阶段，我选择在飞书多维表格中搭建整套制片工作流。它后来演化成了 AnimeTrack。感兴趣的话，可以下载完整的多维表格文件。

  Before AI coding tools were mature enough for direct application development, I built a full production workflow inside Feishu's multi-dimensional tables — the early prototype that later evolved into AnimeTrack.

  ⬇ 制片表 v4.1 （公版）.base // 飞书多维表格 · 可直接导入

  // 业务流程改造 · 翟家班数字化转型

  

  

  

  

  业务流程改造 — 翟家班数字化转型

  这四张截图展示了我之前在一家大型传统艺术机构工作时的成果。当时公司的运作方式非常传统，基本靠纸面和人力沟通。

  我作为战略总监负责公司整体运营，期间对全公司的工作流进行了大幅度改造：将所有业务流程搬到线上，统一在飞书上运行；为每个部门建立了标准 SOP；规范了绩效考核体系；整合了销售端与企业微信的数据对接。通过将外部数据抓取到表格中，每月进行核对并根据数据进行复盘，这一整套数字化的工作方式都是由我引入并落地的。

  As Strategic Director at a traditional art institution, I led a full digital transformation: migrated all workflows to Feishu, established department-level SOPs, standardized performance review systems, and integrated sales data with WeCom. Monthly data reviews replaced paper-based management entirely — all systems designed and implemented from scratch.