cyl19970726

@cyl19970726 source repo

48 skills
0 followers
6 hours ago last updated

▌
Object Object 2 · cyl19970726 bundle
{{DESCRIPTION}} 使用场景： (1) 继续执行此策略（查看当前状态和下一步） (2) 了解策略机制（事件-信号-动作映射） (3) 查看回测/实盘结果和经验教训

0
installs
▌
Video Content Reconstruction 2 · cyl19970726 bundle
Reconstruct a video's full content with an adaptive, evidence-backed two-round workflow. Use when Codex must understand, restore, analyze, convert to an article, document, or audit any video whose important information may live across speech, subtitles, on-screen text, interfaces, actions, parameters, before/after states, examples, claims, counterexamples, or visual transitions. First probe the viewer's intended cognitive change, information carriers, meaning changes, relationship structure, and omission risks; then derive and execute a video-specific capture protocol. Do not route by a closed content taxonomy.

0
installs
▌
Goalflow · cyl19970726
个人目标与日程管理系统。当用户提到任务安排、目标管理、日程规划、每日计划、TODO、待办事项、复盘、完成率时自动触发。支持自然语言添加任务、查看日程、智能规划。Use when user mentions tasks, goals, schedule, daily plan, TODO, review, or productivity tracking.

0
installs
▌
Polymarket Knowledge · cyl19970726 bundle
Polymarket CLOB API knowledge base for order management, WebSocket events, and trading operations. Use when working with order lifecycle (place, fill, cancel), debugging WebSocket USER_TRADE/USER_ORDER events, understanding API field mappings, or implementing trading logic.

0
installs
▌
Deep Content Director · cyl19970726 bundle
Evidence-aware content directing for Xiaohongshu and adjacent short-video platforms. Use when the requested deliverable is an account/content strategy, directing brief, topic selection, script/story/tutorial, shot or carousel-panel plan, title/cover/search/CTA package, series design, publishing experiment, or performance diagnosis. Trigger on 起号、选题、内容策划、视频策划、编导、脚本、口播稿、拍摄提纲、分镜、图文笔记、图文轮播、种草笔记、笔记诊断、拉片后的原创重构、博主内容研究、系列规划、内容复盘 and English equivalents. Do not use for pure transcription/reconstruction, pure factual lookup, media downloading, scheduling, or API-level publishing operations; reconstruct source evidence first, then use this skill only when it must become an original content decision.

0
installs
▌
Shared References · cyl19970726 bundle
Shared Hard Invariants referenced by team coordination skills. Do not install directly; load as a cross-reference from collaborate-as-agent-team-member.

0
installs
▌
Collaborate As Agent Team Member · cyl19970726 bundle
The one Agent Team collaboration contract for BOTH roles — the Host who orchestrates a durable Team and the persistent Members who execute shared Work. Use whenever an agent creates or runs a Team, decomposes/assigns/reviews/accepts Work, waits for member progress, or receives, claims, resumes, executes, blocks, submits Work; reads WorkDelivery or the message Inbox; coordinates Host↔Member or peer↔peer; uses provider-native subagents; or must survive review and runtime restart. The shared mental model in Part I is required reading for both roles — asymmetric mental models are the main way multi-agent collaboration fails.

0
installs
▌
Agentfirm Development Loop · cyl19970726 bundle
AgentFirm Development Loop

0
installs
▌
Harness Frontend Product Design · cyl19970726 bundle
通用前端产品设计-实现工作流：文字合同 → agent 出稿图 → 美术风格测量成文字规格 → 对照规格实现 + 机械验证 → Owner 视觉验收。当任务涉及：做/改任何前端页面或组件的视觉与交互、设计图/出稿/mockup、前端美术风格、UI 还原度、设计规格、前端验收，或要判断'这个界面做得对不对/好不好看'时使用。核心纪律：agent 是翻译者与验证者，永远不是视觉裁判。

0
installs
▌
Design Notion Information Architecture · cyl19970726 bundle
Audit, design, review, or migrate governed Notion workspaces and page systems. Use when Codex must decide how pages, databases, wikis, hubs, views, links, and relations should work together; repair navigation or generic "Related pages" patterns; redesign development or canonical-document systems; create an isolated staging copy; plan a reversible migration; or validate a Notion information architecture before cutover.

0
installs
▌
Guide Agent Teams · cyl19970726 bundle
Claude Code Agent Teams 完整使用指南：多 session 协作、共享任务、inter-agent 通信。包含 Lead 管理方法论（Spawn 设计原则、协调模式、纠正策略、知识提取时机）。使用场景：(1) 创建和管理 agent team (2) 设计 spawn prompt 和角色分工 (3) 理解 tools API（TeamCreate/SendMessage/Task*/EnterPlanMode） (4) 排查 agent teams 问题（teammate 不出现、权限、shutdown） (5) 对比 subagents vs agent teams 选型 (6) Lead 准备创建 agent team 前参考 spawn prompt 设计清单 (7) 协调并行 agents 时选择合适的协调模式 (8) Agent 偏离任务时应用 3-level 纠正策略 (9) 决定何时提取知识到 guide-*/observe-* skills

0
installs
▌
Guide Meta Thinking · cyl19970726 bundle
元思考（MetaThink）方法论：从第一性原理推导系统行为，而非机械查文档。核心是理解"为什么这样设计"而非"它是什么"，从设计约束推导出不可消除的行为边界，用这些边界作为代码正确性的 eval 标准，并区分"我们的 bug"与"系统的固有限制"。有以下意图时必须加载此 skill： (1) 不知道某个行为是 bug 还是系统正常行为 → 用元思考做 fault attribution (2) 面对新的技术系统，需要快速建立正确的心智模型 → 用元思考推导约束 (3) 对一段代码有多种实现方式，不知道哪种正确 → 用元思考找 eval 标准 (4) reconcile 发现异常，不确定根因 → 用元思考的约束链追溯 (5) 需要理解为什么 guide-polymarket-fundamentals 这样设计 → 读本 skill

0
installs
▌
Lead Daily Practice · cyl19970726
metal-coder 项目内 Lead 的日常操作 SOP —— 把元思考定义 + meta-cli 11 命令 + Tower 蒸馏组合成可执行的 7 步流程。必须加载场景： (1) 在 metal-coder 项目（/home/hhh0x/metal-coder/）内推进任何工作 (2) 不知道某个场景该调哪个 meta-cli 命令 (3) 写完一个 meta 文件，需要自检是否合格元思考 (4) chain 累积 ≥5 metas，考虑是否 Tower distill (5) 决定是否 fork 一个工作为多个 sub-chain 本 skill 是元思考定义（docs/meta-thinking/）和 meta-cli 命令实现的 bridge 层。没有它，Lead 会 default 手写 metas 而不调 cli，工具不被使用价值流失。

0
installs
▌
Skill Creator · cyl19970726 bundle
Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, update or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. Also use for Earning Engine PM v2 skill creation — task-* skills (development tasks), strategy-* skills (trading strategy lifecycle), observe-* skills (analysis reports), guide-*/map-* skills (knowledge docs). When users say "create a task", "plan a new feature", "analyze architecture", "write a guide", or "develop a new strategy", use this skill to route to the appropriate PM v2 template.

0
installs
▌
Guide Meta Agent System · cyl19970726 bundle
v5 入口 — Lead + Sub-agent + Agent-Team-Member + Meta-meta + Skill-manager + Advisor 协作系统宪法（7 角色）。 v4 在 SKILL.md 单文件中累积过满（1638 行），v5 拆分为 13 个聚焦文件 + 1 入口索引，提升 cold-start navigability，并新增 7 角色 (含 skill-manager + advisor) / mermaid 数据流图 / enforcement layer / dashboard 规范。

0
installs
▌
Guide Metal Coder Essence · cyl19970726 bundle
metal-coder 的北极星定义 —— 8 个元思考组成的演化序列，构成项目本质。从"我们解决什么问题"开始，到"系统怎么自己进化"结束，每个元思考都 derives_from 前一个，本身就是一条 derivation chain。 **所有其他 skill / cli 命令 / docs 都是这 8 个元思考的具体化或优化。** 读完本 skill 才算真正理解 metal-coder。必须加载场景： (1) 第一次接触 metal-coder（在 docs/ 之前先读本 skill） (2) 设计任何新功能 / 新命令前 — 先确认它服务哪个元思考 (3) 评估"这个改动是不是核心"— 用 8 元思考序列对照 (4) 想理解 metal-coder 与 v5 / docs/ / lead-daily-practice 的关系 (5) 跨 work chain 反思时 — 检查 8 元思考是否真的在被实施

0
installs
▌
Object Object · cyl19970726 bundle
{{DESCRIPTION}} 使用场景： (1) 继续执行此任务（查看当前状态和下一步） (2) 了解{{TOPIC}}的设计方案 (3) 从执行记录中提取经验教训

0
installs
▌
Tripo 3d · cyl19970726 bundle
Tripo v3 在 Blender Harness 中的可恢复 provider adapter。用于 Tripo P1 低模、多视图生 3D、Smart Retopology、文件上传、task poll/fetch、AutoRig、retarget、与混元候选对比，或处理 Keychain、短期 URL、SUBMIT_UNKNOWN 和 provider DONE 不等于资产通过等边界。

0
installs
▌
Hunyuan 3d · cyl19970726 bundle
腾讯混元生3D在 Blender Harness v1 中的 provider adapter。涉及混元图/文生3D、多视图、人物画像、AutoRigging、文生动作、纹理、减面、部件、UV、格式转换、10类能力/19 Actions、JobHandle、COS 下载或 Blender retarget 时使用。混元只产 candidate；DONE 永不等于 Blender 资产通过。

0
installs
▌
Jygc Ar Assets · cyl19970726 bundle
Use for organizing, classifying, moving, packaging, or publishing Jieyang Gucheng AR assets across the repo, mini program package, CDN, or CloudBase Storage.

0
installs
▌
Jygc Ar Lapian · cyl19970726 bundle
Use for project-specific AR reference breakdowns, frame analysis, and storing lapian cards for Jieyang Gucheng mini program XR decisions.

0
installs
▌
Blender Harness · cyl19970726 bundle
揭阳古城 Blender 长程资产与动画任务的 adaptive harness。当任务涉及 Harness、路线选择、执行中改路线、RouteHypothesis、ProbeRun、Quicklook、多视图证据、独立评审、Blender headless 作业、Hunyuan 候选进入 Blender、揭小贤素体/绑骨/服装/动作，或需要判断哪些内容该做 machine gate、agent review、casebook 时使用。核心不是固定 Phase 流水线，而是目标画面→可证伪路线→小探针→证据→继续/修改/放弃/请求 owner。

0
installs
▌
Jygc Spot Animation · cyl19970726 bundle
揭阳古城真实景点 AR 效果层的领域合同、路线选择与探针菜单。用于扫真楼、景点打卡动画、触发图、photo-match/fSpy、固定效果相机、长出点、黑底加法光、SBS alpha、实时 3D、composite preview、史实和宗教表达检查。它帮助选择配准、可分解性、合成、内容或真机 probe，不规定固定六拍、G0 四件套、300 帧或 tournament 流水线。

0
installs
▌
Jygc Magnet Animation · cyl19970726 bundle
揭阳古城冰箱贴、印章卡和文创卡的卡面 AR 动画领域合同、路线选择与探针菜单。用于微缩世界、破框立起、虚拟运镜、擦镜/扑镜、SBS alpha、预渲染视频、磁贴 SKU 创意或判断候选是否符合卡面产品体验。它帮助选择最低成本的定帧、motion、edge-owner、A/V 或 runtime probe，不规定固定 Phase/Gate、固定帧数或评审编排。

0
installs
▌
Jygc Dragon Rig Pipeline · cyl19970726 bundle
揭阳古城 AR 中东方长身龙的领域合同、路线选择与高信息增益探针。用于龙、Loong、盘龙、龙绕楼、扑镜、龙三视图、Hunyuan 龙 source、组件拼接、retopo、UV、Spline IK/B-Bones、龙绑定或动画路线判断。它帮助判断“当前最危险的龙资产假设是什么、用哪个最便宜的 Blender/Hunyuan 探针证伪”，不规定固定 Phase/Gate 流水线。

0
installs
▌
Compare Creators · cyl19970726 bundle
Compare two or more accepted Creator Dossiers across positioning, audiences, values, content portfolios, performance tiers, proof systems, directing and visual patterns, evolution, demand, and commercial evidence. Use for multi-creator research; do not turn comparison into copying advice.

0
installs
▌
Analyze Creator Videos · cyl19970726 bundle
Orchestrate a creator's complete public portfolio research from profile URL through acquisition, full-corpus annotation, high/median/mean-near/low selection, deep video reconstruction, creator synthesis, optional independent evaluation, and one evidence-backed Dashboard. Use for neutral creator research; keep the user's own copying or publishing strategy in content-strategy-workbench.

0
installs
▌
Creator Sample Selection · cyl19970726 bundle
Select reproducible high, median, arithmetic-mean-near, and low creator samples for comparison and deep reconstruction. Use after corpus metrics and portfolio annotations exist; do not use for subjective cherry-picking or for claiming selected posts represent the full account.

0
installs
▌
Content Strategy Workbench · cyl19970726 bundle
Design the user's own content strategy from their positioning, constraints, and explicitly selected creator or post evidence. Use for the separate creation workspace—content engines, 10/30-post plans, next-post briefs, titles, covers, hooks, evidence, CTA, and experiments—not for neutral creator research.

0
installs
▌
Creator Research Evaluator · cyl19970726 bundle
Independently evaluate a creator research run, its acquisition honesty, corpus reproducibility, selection coverage, video CR/DL/VE gates, synthesis evidence, and Dashboard parity. Use as a fail-closed reviewer; do not evaluate work with the same reasoning context that produced it.

0
installs
▌
Creator Research Synthesis · cyl19970726 bundle
Synthesize a creator's positioning, audiences, values, trust, lifecycle, content system, performance differences, publishing evolution, public demand, and commercial evidence from a full corpus plus validated deep videos. Use for Creator Dossiers, not for recommending what the user should copy or publish.

0
installs
▌
Creator Portfolio Annotation · cyl19970726 bundle
Annotate every observed creator post at portfolio depth using titles, covers, post copy, public metadata, and available evidence. Use to build open-ended topic, format, promise, audience-value, proof, and visual-language fields without pretending surface evidence is a full video reconstruction.

0
installs
▌
Video Content Reconstruction · cyl19970726 bundle
Reconstruct a video's full content with an adaptive, evidence-backed two-round workflow. Use when Codex must understand, restore, analyze, convert to an article, document, or audit any video whose important information may live across speech, subtitles, on-screen text, interfaces, actions, parameters, before/after states, examples, claims, counterexamples, or visual transitions. First probe the viewer's intended cognitive change, information carriers, meaning changes, relationship structure, and omission risks; then derive and execute a video-specific capture protocol. Do not route by a closed content taxonomy.

0
installs
▌
Xiaohongshu Creator Acquisition · cyl19970726 bundle
Acquire and resume a Xiaohongshu creator's public profile, visible post inventory, post details, covers, public metrics, bounded comments, and selected media with auditable coverage. Use for creator research collection, not for interpreting content mechanisms or evading platform controls.

0
installs
▌
Ego Multiplatform Publish · cyl19970726 bundle
Use Ego Browser to prepare, save drafts, verify, or publish content through the real creator forms of Xiaohongshu, Douyin, WeChat Channels, Bilibili, and WeChat Official Accounts. Use when platform-specific fields and reliable platform-side evidence matter in this self-media project.

0
installs
▌
Tripo API · cyl19970726 bundle
Tripo3D REST API 实战参考与统一驱动 —— 图生3D / 文生3D / 多视图重建 / 自动绑骨 / 预制动画重定向的确切端点、参数、模型版本、积分、以及大量实测踩坑。当任务涉及:用 Tripo 把图/文生成 3D 网格、给模型自动绑骨 (Mixamo 骨架)、套 Tripo 预制动画、选 Tripo 模型版本(v2.5/v3.0/v3.1/P1)、控制贴图质量/四边面/风格化、排查 Tripo 任务失败(1004/2006)、估算 Tripo 积分消耗 —— 时务必使用本 skill。出现 tripo3d、image_to_model、 multiview_to_model、animate_rig、animate_retarget、tripo api key、图生3D、绑骨、tripo 积分等关键词也应触发, 即使用户没说"skill"。配套驱动脚本 `scripts/tripo_gen.py`。注:做完整 AR/动画管线请配合 `ar-3d-animation-pipeline`;本 skill 只聚焦 Tripo API 本身。

0
installs
▌
Blender MCP · cyl19970726 bundle
用 Blender MCP(blender-mcp 插件 + socket)程序化驱动 Blender 做 3D 动画管线的实战方法论 + evaluator + 踩坑。当任务涉及:用 Blender 给角色做/改动画、重定向 Mixamo 动作、IK 摆姿势、NLA 合成多 clip、导出 glTF/GLB、在 Blender 里渲染验证、清理/重绑骨、把 Tripo/FBX 模型加工成带多动作的 AR 模型 —— 时使用本 skill。出现 blender、blender-mcp、execute_code、bpy、armature、retarget、IK、glTF导出、骨骼动画、动画评估等关键词也应触发。配套:动画来源/管线的全局地图见 `ar-3d-animation-pipeline`;Tripo 建模绑骨见 `tripo-api`。 ⚠️ 活文档:每次实践有新经验就往"踩坑/经验"追加。

0
installs
▌
Xhs Crawler · cyl19970726 bundle
用 rnote.dev API 采集小红书(RED)数据,做市场调研/选题/竞品扫描。支持搜索笔记/商品/用户、笔记与商品详情、评论、话题、用户主页、热点灵感 feed,并能把笔记还原成可点击原文链接。触发词:"小红书爬虫"、"采集小红书"、"小红书调研"、"xhs"、"rnote"、"文创选题数据"、"竞品笔记扫描"。

0
installs
▌
Codex Imagegen · cyl19970726
用 codex 生成/编辑高质量位图(照片、插画、产品图、UI mockup、概念图、贴图、IP 形象、透明抠图)的实战方法论 —— codex 自带一个 `image_gen` 内置工具(底层 gpt-image,无需 OPENAI_API_KEY、免联网配置),你用自然语言让 codex "生成一张图/改这张图",它会【自动触发】系统 imagegen skill 调用该工具。本 skill 教你怎么把它的质量拉满 + 在 harness 多 agent workflow 里批量驱动 codex 生图。当任务涉及:用 codex/AI 生成图片、出概念图/海报/产品图/封面/hero、给图换背景/改光影/抠透明、做一批资产或变体、 IP 形象/吉祥物出图、给 [[ai-ip-series-design]] 真正"跑图"、文生图/图生图/图改图、gpt-image/透明背景/4K 出图、在 .star workflow 里 fan-out 多个 codex 并行生图 —— 时使用本 skill。出现 codex 生图/image_gen/gpt-image-2/ gpt-image-1.5/imagegen/抠图/chroma key/批量生图/generate-batch 等关键词也应触发,即使没说"skill"。 ⚠️ codex 专属:claude/kimi 没有这个内置工具。真要做可绑骨的 3D/AR 见 [[ar-3d-animation-pipeline]]/[[tripo-api]]; 多 agent 编排见 [[author-workflow]]。⚠️ 活文档:新踩的坑/新验证的参数往 §7 追加。

0
installs
▌
Author Workflow · cyl19970726 bundle
Use when an agent needs to author a dynamic multi-agent workflow at runtime: write a Starlark program (loops, conditionals, data-driven fan-out) that calls agent()/parallel()/phase()/log() over the harness runtime, declare a mandatory workflow(name, design_intent) header, run it with `harness workflow run-script`, and read the resulting WorkflowRun/WorkflowStep records back from the dashboard snapshot or store.

0
installs
▌
Decision Closure · cyl19970726
决策闭包方法论:面对"没做过的复杂任务",在执行前重建任务的决策闭包(必须被做出的全部决策),用决策覆盖表暴露"被意外决定/无人决定"的决策类,把延期机制(闭包债逐次碰撞偿还)换成纸面成本。当出现:复杂任务反复返工/重复延期、同一交付物验收多次不过、用户反复纠偏、"看起来有设计但仍然混乱"、需要复盘"为什么一直失败"、或要给反思/复盘本身立验收标准(回溯测试)——时使用本 skill。配套 workflow:closure-retro.star(三路闭包源重建覆盖表+盲集回溯验证)。⚠️ 活文档:每次覆盖表被新失败证伪,把漏掉的决策类追加进 §6。

0
installs
▌
Video Understand · cyl19970726 bundle
拉片——把一条对标视频拆成可复刻 SOP。管线:下载→whisper逐字稿→抽帧(粗览+按需密抽)→判型→逐段填表→产出①逐段拉片报告(它怎么做的,工具链带置信度)+②复现SOP(你怎么抄)。用于"学习/复刻优秀博主的爆款视频"。触发词:"拉片"、"拉这条视频"、"拆解爆款视频"、"复刻对标博主"、"理解这条视频"、"这个视频在讲什么"、"学习博主视频"、"video understand"、"看懂小红书视频"。

0
installs
▌
Webar Mindar IOS · cyl19970726 bundle
Build, debug, and ship browser-based AR image-tracking experiences (MindAR + A-Frame + Three.js inside a Next.js App Router H5), and especially diagnose the iOS-Safari/WebKit-only failures that desktop Chromium never reproduces — blank or black camera after "start AR", scenes collapsing to width:0, next/image not rendering, gyro/DeviceOrientation not firing. Use this whenever the task involves WebAR / MindAR / A-Frame / image-tracking / "扫描识别" AR on phones, an AR page that works in Chrome but is blank/broken on iPhone, reproducing iOS WebKit bugs without a physical device, or deploying such an app to a GCE VM behind a cloudflare quick tunnel for on-phone testing. Reach for it even when the user just says "the camera is blank on iPhone after I tap start" or "make the AR camera fullscreen" — the iOS-WebKit gotchas and the Playwright-WebKit verification combo here save hours of guess-and-redeploy.

0
installs
▌
Session Forensics · cyl19970726 bundle
给 agent 会话配「观察者」：读完整会话历史（Codex rollout JSONL、Claude Code transcript、 Kimi wire、DeepSeek Harness / dsh 的 zstd 会话，含 subagent/team 日志），用「物证流 vs 叙事流」对质，找出执行 agent 自己原理上看不见的问题。四种能力：① audit——审计求解过程是否病态（空转、卡住、目标漂移、 gate 膨胀、修工具循环、假通过）；② handoff——交接包（压缩/中断/换人/跨 CLI 时， --check 闸保证完成度）；③ progress probe——运行中探针（silent worker / subagent 现在在干什么，秒级判定，防空等）；④ harvest——沉淀（从历史挖出该固化成 skill/gate/harness 的东西）。当出现：读 session 历史、复盘会话、审计另一个 agent、为什么进度这么慢、为什么一直失败、agent 卡住了、subagent 在干嘛、compaction 之后丢了什么、要交接给下一个 agent、~/.codex/sessions、~/.dsh/sessions、rollout jsonl、session.jsonl.zstd、 dsh 会话、deepseek harness 历史、agent 复盘 —— 时使用本 skill。 ⚠️ 不要用向量检索/RAG/agent-memory 读 session（实证判死，见 §禁止事项）。

0
installs
▌
AI Ip Series Design · cyl19970726 bundle
用 AI 做「可控的 IP 系列形象」的完整方法论与实战提示词 —— 把一个吉祥物/IP 角色扩展成一整套风格统一、角色一致、可延展(换装/动作/入景/周边)的系列，而不是只出一张惊艳的单图。当任务涉及:设计一套 IP/吉祥物/虚拟人系列、做 IP 家族(多角色同风格)、AI 出图"第一张惊艳但成不了系列/ 一致性崩"、角色三视图/转身/表情包/换装/动作设计/把角色融进场景、黏土风/羊毛毡/盲盒手办质感、多图参考锁角色一致性、2D 形象转 3D 质感、给 IP 做周边延展、品牌吉祥物落地 —— 时使用本 skill。出现 IP设计/IP系列/吉祥物/虚拟人/角色一致性/character consistency/IP family/换装设计/动作设计/ 黏土风/羊毛毡/即梦/Lovart/Nano Banana/Seedream/多图参考、找IP设计参考/对标成熟IP/世界级IP范例/moodboard/benchmark 等关键词也应触发，即使用户没说"skill"。它管的是【二维形象设计层】:产出一套锁死的一致角色 + 转身/姿势/场景图。要把这些图真正做成可绑骨的 3D 角色/AR 资产,接 [[ar-3d-animation-pipeline]] / [[tripo-api]] / [[hunyuan-3d]];多版本横向择优用 [[parallel-concepts-tournament]];给一致性/美感立评审用 [[evaluator-design]]。 ⚠️ 活文档:每做一个新 IP,把新踩的坑/新验证的提示词追加进 §6、§7。

0
installs
▌
Exploration Judgment · cyl19970726
一套关于"探索的判断力"的思维方式:在沉成本之前判断哪个节点值得深挖、哪个够用即走, 并把一切知识当作可证伪、会过时的东西来持有。核心命题是「思想比 gate 重要」—— 失败几乎从不是"没有深挖的方法",而是"没认出这里该深挖"。当你开始调研/上手一个工具/API/库/技术、要在多个方案间选型、打算"就用默认值"、在同一步反复返工/重抽、搭建知识库或文档、或隐隐觉得"我对这块的理解很浅"时,务必调用本 skill。也适用于复盘"为什么我们总是漏掉关键东西"。即使用户没说"深挖/探索/判断力"也应触发。

0
installs
▌
Ar 3d Animation Pipeline · cyl19970726 bundle
从 0 到 1 制作浏览器 WebAR 的 3D 动画角色/场景的完整流程、技术原理与已验证工具链。当任务涉及"把一个 IP 形象/吉祥物做成「扫描识别卡或冰箱贴 → 3D 角色临现并做动作(倒茶/挥手/提灯等)」的 AR 体验", 或需要:概念图(art direction)定调、AI 图生3D(Tripo/TRELLIS)、给 Q版角色自动绑骨(Mixamo)、移动端模型优化(gltf-transform meshopt)、react-three-fiber 场景装配、手作骨骼动画、给程序化动画做客观精度评估(几何 + 视觉 agent 双评估)、MindAR 图像追踪接入、AR 资产部署真机 —— 时, 务必使用本 skill。涉及 codex/FLUX 生成对版参照图、Tripo API 自动建模绑骨、TRELLIS、rembg 抠图、 GLB 减面压缩、骨骼蒙皮、扫贴/识别卡 AR、3D 吉祥物动起来、AR 冰箱贴等关键词时也应触发, 即使用户没明说"流程"或"skill"。

0
installs
▌
Beat Performance Pipeline · cyl19970726
把"3D角色一段连续表演"(如揭小贤走到茶桌坐下冲茶讲解)拆成 Beat(与配音cue逐句对齐的最小动作单元), 每个 Beat 走设计→实现→多维evaluator(workflow)→loop 的生产管线。核心物料=【坐标契约 scene-layout.json】 (角色/道具/落座点/抓取点共用一份坐标,authoring与runtime一致=手到壶到)。当任务涉及:分镜化角色表演、与配音同步的动作序列、道具拿放(re-parent)、走路/坐下/倒茶类动作、茶汤/重力类物理合理性检查、逐beat评审 —— 时使用本 skill。配套 workflow:项目 workflows/beat-eval-loop.star。 ⚠️ 活文档:揭小贤茶艺表演的每个 beat 实作经验都往本文件追加。

0
installs

48 published skills