SKILL.md 里的正文，确实还是写给模型看的提示词。给它换个文件名，不会让模型多长一个脑子。
我用 Skill 做了一段南京新街口的微缩动画。对我有用的部分，是把参考图怎么用、什么时候停下来给我看、哪些脚本可以复用，保存在一套能再次调用的文件里。下次换主题，不用从聊天记录里捞那句“对，就按这个画”。
- 微缩景观动画 Skill传送门： memphis-mini-world
- Skill 设计助手传送门： skill-designer
skill-designer 怎么用，和 Claude 官方的有什么区别？
memphis-mini-world 是拿来做动画的，skill-designer 则帮助设计 Skill 本身。
比如，我知道自己要做微缩动画，却未必知道该怎么写一份能反复使用的说明。哪些内容放入口？哪些另外存？要不要先问需求？能不能让 AI 跳过样帧直接出片？skill-designer 会帮助梳理这些问题。
它介绍了五种常见写法：封装工具或框架规范、套模板生成、按清单审查、先问需求再行动，以及按步骤执行。几种写法可以组合。我这次就用了 Pipeline、Generator 和 Reviewer，分别管制作顺序、可复用的输出结构和检查清单。
我用的是 Pi-Bolt，把解压后的 skill-designer 目录放进 ~/.agents/skills/，用 /reload 刷新，然后就可以调用。下面是这个工具里的示例，换其他工具时，安装位置和命令要看各自说明：
/skill:skill-designer
帮我设计一个微缩景观动画 Skill，支持文字和参考图。
先分析参考，制作同一动画的四张样帧，等我确认后再渲染全片。
请说明适合组合哪些设计模式，以及 SKILL.md、参考文档、模板和脚本怎么分工。
也可以把已有 Skill 交给它，检查触发描述是否清楚、流程有没有漏掉确认环节。它还附了测试提示和预期结果，不过没有配套的自动评测脚本。
Anthropic 官方的 skill-creator 覆盖得更广。我核对了官方当前的说明，它除了帮助写 Skill，还提供测试、评审、修改和触发描述优化的流程与配套工具。
两者有重叠，官方版也讲文件结构和渐进式披露。对我来说，skill-designer 适合先把写法想明白；需要系统比较修改前后的效果时，可以接着用官方 skill-creator。官方部分评测依赖 Claude CLI 或子代理，换个工具运行前还得检查环境。
这次新街口制作里，我用了 skill-designer 的组织方法，没有运行官方 skill-creator 的那套对照评测。两件事别混在一起。
我做出来的，是这样一条片子
20 秒，1080×1080，60fps。镜头跟着一辆珊瑚红的小车，经过中山像，再往另一段街区走。外面留着暖白边，片头从中间慢慢展开。
风格起点是这条微缩景观动画参考帖。我把整理后的规范叫作“孟菲斯微缩世界”：平面色块、靛色描边、硬边投影，窗口边缘散成颗粒，下面的双语图注不跟镜头走。
画面采用 2D 和 2.5D 插画表达。Agent 写 Canvas 绘图程序和声音脚本，我提供参考、提要求、看样帧，再决定要不要改。最后用 ffmpeg 编码成视频。
整个过程改了好几轮。楼画得太像模型，雕像一度有点吓人，车还开进过行人过街区。
提示词能说清楚的事，为什么还要存成 Skill？
做这一条片子时，当然可以一直在对话里补要求。可下次做水乡，我还得重新解释哪些风格不能变、参考图怎么处理，以及“先让我看图”的具体意思。
例如，照片和插画参考承担的任务不同。新街口实景图提供道路、交通岛和地标的位置关系，风格截图决定楼和车怎么画。我只要求雕像保留站姿，没有要求它把照片里的材质和面部阴影一起搬过来。街区密度、楼的体量和立面细节都可以重新组织，坐标也明确标成示意。
如果只说“参考这几张图”，AI 很容易同时参考了我想要和不想要的部分。
还有这条写在主文件里的要求：
Do NOT render the full video until the user explicitly approves the current four-frame preview.
意思是先给同一动画的四张实际样帧，我明确同意当前版本以后，才出全片。时长、画幅和音乐方向有默认值，不必每次先填一份技术问卷；关键要求冲突、图片打不开时再问。
这个流程用的是 skill-designer 里的 Pipeline。Generator 帮我固定场景配置和项目结构，具体场景仍由 Agent 写代码；Reviewer 负责把造型、运动和输出规格列进检查项。
只把“必须等我确认”写进文档，还不能保证 Agent 永远不跳步。这次也配了程序校验：出片前核对批准记录和对应版本，画面改了就重新看样帧。哪些规则需要代码检查，得在制作时继续补。
渐进式披露，就是用到哪儿读到哪儿
这个词听着像一门管理课。我在 Pi-Bolt 里用到的机制很直白。
模型先看到 Skill 的名称、描述和路径。判断当前任务适合它，再读取 SKILL.md；我也可以用 /skill:名称 明确调用。入口里还会告诉它，进行到某一步时该读哪些详细资料。
以这套动画 Skill 为例，目录大致是：
memphis-mini-world/
├── SKILL.md
├── references/ 画风、参考图处理、运动、音频和验收规范
├── assets/ 配置模板、审阅模板、认可的造型示例
└── scripts/ 校验、导出、断点恢复和媒体检查工具
刚开始看雕像应该怎么画，就读风格说明和造型示例。写人车运动时，再读运行规则和安全检查。等我同意出片，才需要看声音制作、编码和断点恢复的细节。
入口可以这样指路。下面是简化示例：
开始设计时，读取 references/style-guide.md 和 references/visual-calibration.md。
有参考图时，读取 references/reference-images.md，并实际查看图片。
编写绘图程序前，读取 references/canvas-runtime.md。
确认后制作全片时，读取 references/production-workflow.md。
拆文件还不够，入口得写清楚读取时机。把资料全扔进 references，然后期待 AI 自己悟出顺序，多少有点为难它。
这和 Pipeline 管的是两回事。渐进式披露管资料何时加载，Pipeline 管步骤顺序和确认条件。脚本则通过工具运行，负责固定的检查与导出，不必每次把整份源码都读进模型上下文。
所以 Skill 里可以有提示词，也可以带模板、图片和可执行脚本。它仍然依赖宿主工具提供读文件、运行程序等能力，装上一个文件夹，不会自动获得这些权限。
新街口的要求，是边做边补齐的
我把最终要求整理成了下面这版。它合并了多轮反馈，并非最开始一次就输入了这么完整的一段：
/skill:memphis-mini-world
制作南京新街口·中山像的20秒微缩景观动画，1080×1080，60fps。
实景图参考路网和地标关系，所有元素按插画截图的风格重新设计。
镜头从街道一端局部开场，跟随珊瑚红小车经过中山像，再前往另一端。
视角抬高、拉远一些，但不要从整座沙盘的全景开始。
商业中心要有连续多排街区和次级街巷。
雕像保持参考图的站姿，整段不动。
安排红灯礼让行人和白鸽起飞的小剧情。
片头从中心向四周逐渐展开。
先给同一动画的4张样帧和拼图，确认后再生成全片。
Agent 写好场景配置和绘图程序后，从同一套实现里导出样帧。这次的四个审阅时刻是 1.8、6.1、10.7 和 17.6 秒，用来看沿街路线和不同街段。
早期画面像在十字路口边摆了一圈小楼，镜头也接近固定全景。我想要商业中心的密度，后来才补上前后多排建筑和次级道路，让跟车镜头一路能看到新的街段。
楼的画法也走过弯路。小窗、窗台、檐口越加越多，车上还有斜风挡、轮毂和灯组，逐渐长成了另一种低多边形模型。最后删掉这些细节，楼改成大片同色系立面和少量大窗，车保留纯色车身、浅色直立驾驶舱和小实心轮。
雕像保留站姿和长外套轮廓，去掉尖锐侧脸和碎片明暗，才没那么诡异。AI 对细节的热情，有时需要人工劝退。
后来我给 Skill 补了对象级检查。除了看配色，还要把建筑、车和人物放大，分别核对轮廓、比例和细节密度。
动态问题则更具体。主车停下来让行人了，对向车却没接入同一套礼让时序。当时只检查车车间距，漏掉了人车，结果它照常开了过去。
修复后，绘图和安全检查使用同一套角色轨迹，同时检查双向车流、车身占用的范围，以及行人走完后的缓冲时间。20 秒、60fps 对应 1200 个时间点，批准前就能遍历这些时间点做几何计算，不必先渲染整部视频。
片头还有个乌龙。原来的展开在 1.4 秒已经结束，预览却从 1.8 秒开始，我当然只能看到一张完整画面。后来把播放起点改为第 0 帧，并把这条片子的展开延长到约 2.6 秒。
四张静态图可以看构图，播放起点和过街过程还得另外查。我说“车辆，建筑，行人可以了”，也只代表认可这些造型。后面修改片头时，就不用顺手把车和楼再画一遍；整版出片要等另外的明确批准。
从出片到下一次使用
最终画面采用确定性 frame(i)，也就是由帧号决定画面。同一个帧号重复计算，像素应该相同。这样跳着检查、重新加载页面，或者中断后继续，都能核对是不是原来那套画面。
frame(300);
frame(720);
frame(300);
两次第 300 帧需要得到相同的 RGBA 摘要，相当于给像素做一次指纹比对。
正式导出得到 1200 帧，再合成五声音阶轻音乐、街道底声、脚步和鸽翼声，编码为 H.264 和 AAC 的 MP4。导出途中遇到过后台 PNG 回调慢、标签导航中断，后来改用同一现有 Chrome 标签内的同步 PNG 编码，从核验过的 480 帧断点继续。
最终四张批准样帧的原始像素摘要保持一致，成片完成了全部 1200 帧的解码检查。响度、峰值和声音事件同步也能测量。不过这些数值回答不了音乐好不好听、镜头看久了会不会无聊，仍然需要看片和试听。
新街口先用项目内脚本完成，随后我才要求 Agent 把经验写回 Skill，升级到 v1.1.0。新版记录局部确认和整版批准，也把视觉、声音计划、生产版本分开。仅修复导出工具时，可以核验已有帧后继续；改了画面或剧情，就需要重新确认。
新增脚手架通过了 20 项 Python 测试和 10 项 Node 测试，覆盖审批、路径限制、本地接收、断点核验和碰撞反例。但它还没换一个新主题跑完真实 Chrome 端到端制作，具体场景和声音仍需 Agent 编写。不能因为新街口出了片，就宣布所有城市都能一键生成。
我以后可以这样调用它：
/skill:memphis-mini-world 江南水乡的早市，乌篷船从石桥下经过，沿岸有人买花。
画风和制作流程会沿用，新场景仍然要看样帧，不能拿新街口的批准去通行。
回到题目，我愿意把反复做的任务整理成 Skill，因为这次返工留下的规则和工具，下次还能用。只做一次的简单任务，我会直接写提示词。至于一份 Skill 是否真的比原来的办法好，可以像官方 skill-creator 那样拿任务做对照；光看目录多整齐、SKILL.md 写了多少行，我判断不了。
- 微缩景观动画 Skill传送门： memphis-mini-world
- Skill 设计助手传送门： skill-designer
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The text in SKILL.md is indeed a prompt written for the model. Change it to a file name that doesn't give the model an extra brain.
I used Skill to make a micro-animation at Nanking New Street. The part that is useful to me is how, when, and which scripts can be used again and stored in a set of documents that can be called again. Next time you change the subject, you don't have to pull the phrase "Yes, press this painting."
- Microscape Animation Skyll portal: memphis-mini-world
- Skill Design Assistant portal:
What difference does it make with Claude's official?
Memphos-mini-world is animated, skill-designer helps design Skill itself.
For example, I know I'm going to do microimpress, but I don't know how to write a note that can be used repeatedly. What's at the entrance? Which are the other ones? Should we ask first? Can you get AI to skip the frame and come out? Skill-designer will help comb these problems.
It presents five common writings: encapsulation tools or framework specifications, template generation, inventory review, demand prior action, and step by step. Several writings can be combined. This time I used Pipeline, Generator and Reviewer to control sequencing, reusable output structures and inspection lists.
I'm using Pi-Bolt, and I'm putting the pump-designer directory in ~/.agents/skills/, refreshing with /reload, and then calling. The following is an example of this tool. In exchange for other tools, the location and command of the installation depends on each:
/skill:skill-designer
Help me design a microscape animation, Skill, to support text and reference maps.
We'll analyze and make four frames of the same animation, and we'll rephrase the entire film after I've confirmed it.
Please indicate which design models are suitable for combining and how SKILL.md, reference documents, templates and scripts are divided.
It can also be handed over to the existing Skill to check whether the trigger description is clear and the process missing the confirmation link. It also attached test tips and expected results, although there was no accompanying automatic assessment script.
Anthropic official skill-creator coverage is wider. I checked the official current note, which, in addition to helping write Skill, provides a process and accompanying tool for testing, evaluating, modifying and triggering the description optimization.
The two overlap, and the official version also deals with document structure and progressive disclosure. For me, skill-designer is a good way to start by thinking about it first; if you need a system to compare the effects before and after, you can use the official skill-creator. Official parts of the assessment rely on Claude CLI or sub-agents to check the environment before another tool is run.
In this new street production, I used the skill-designer organizational method to run the official skill-creator's set of checks. Don't mix the two things.
I made it. It's like this.
20 seconds, 1080 x 1080, 60fps. shot followed a small car of coral red, passing through the Middle Hills and heading for another block. There's a warm white side outside, and it's slowly spreading through the middle.
The starting point for style is this micro-square animation reference. I call it the Memphis world of microcosms: a flat piece, a corset, a hard side projection, a window edge scattered into particles, and a bilingual chart below does not follow shot.
The image is expressed in 2D and 2.5D illustrations. Agent writes about Canvas drawings and sound scripts, and I provide references, requests, frames, and then decides whether to change them. Finally encoded with ffmpeg.
The process was changed several times. The building was so much like a model, the statue was a little scary and the car drove into the street.
prompt can make things clear. Why do we have Skill?
This film, of course, can always be supplemented in the dialogue. But next time I do waterworks, I have to re-explain the styles that can't change, how the reference diagrams are handled, and the specific meaning of “Let me see the map first”.
For example, photographs and illustrations carry different tasks. The New Street Map provides the location of roads, traffic islands and landmarks, and the style screenshot determines how the building and the car are painted. I only asked the statue to remain standing and not to bring the material and facial shadows from the photograph together. The density of the block, the size of the building and the details of the surface can be reorganized and the coordinates clearly indicated.
If we just say, "Refer to these pictures," it's easy for AI to look at what I want and what I don't want.
And this is a request written in the main document:
Do NOT render the full video until the user explicitly approves the current four-frame preview.
It means giving four actual frames to the same animation, and I expressly agree with the current version before coming out. There are default values for length of time, painting and music direction, and it is not necessary to fill out a technical questionnaire each time; the key requirements are conflict and the picture cannot be opened.
This process uses Pipeline from the skill-designer. Generator helps me with the configuration and project structure of the scene, which is still coded by Agent; Reviewer is responsible for placing modeling, motion and output specifications in the check box.
Just write "must wait for my confirmation" in the document, and there is no guarantee that Agent will never jump. This time it was accompanied by a verification of the procedure: check the approval record and the corresponding version before the release, and re-read the image. The rules that require code checks must be continued at the time of production.
Gradual disclosure is where you read it.
That word sounds like a management class. The mechanism I used in Pi-Bolt was straightforward.
The model first sees the name, description and path of Skill. Let's judge the current task fit for it, read SKILL.md; I can also use /skill: name to call clearly. The entrance will also inform it of the details to be read at a certain point in time.
In the case of this set of animations, Skyll, the catalogue is roughly:
memphis-mini-world/
├── SKILL.md
Ideas -References/ Drawings, Reference Map Processing, Exercise, Audio and acceptance check Norms
Ideas -assets/ Configure Templates, View Templates, Approved Model Examples
└ - - scripts/ check, export, breakpoint recovery and media check tool
Just starting with the statue, reading the style description and modeling examples. Read the rules of operation and the security check when writing a car. The details of sound production, coding and breakpoint recovery will be needed when I agree to produce a film.
The entrance can guide the way. The following are examples of simplification:
When you start designing, read access/style-guide.md and access/visual-calibration.md.
Read references/references-images.md and actually view the pictures when reference maps are available.
Read access/canvas-runtime.md before preparing a drawing program.
Make a full film after confirmation and read references/prevention-workflow.md.
The dismantling of documents is not enough, and the entry must be clear about the timing. Throwing all the information into references, and post-production to AI to find out the order itself.
This and Pipeline don't matter. When the progressive disclosure of the tube information is loaded, Pipeline regulates the sequence of steps and the conditions for confirmation. Scripts are run through tools and are responsible for fixed checks and exports without having to read the whole source code into model context each time.
So, Skill can have prompt or a template, a picture and an enforceable script. It continues to rely on host tools to provide the ability to read documents, run programs, etc., and to load a folder without automatically acquiring permission.
The demands at the new street are made while they're done.
I put the final request in the next edition. It combines multiple rounds of feedback and does not enter such a complete paragraph in the first instance:
/skill:memphis-mini-world
A 20-second microfiche animation of the Nanjing New Street Junction, 1080 x 1080,60fps.
Map reference network and landmark relationship, all elements redesigned in the style of illustrations.
shot opened locally from one end of the street and followed the coral red minivan through the middle of the hill to the other.
Keep your perspective up and out, but don't start with the panorama.
Business centres need to have consecutive multiple blocks and secondary blocks.
The statue is standing on the reference map, and the entire section remains intact.
Make arrangements for a red light to let pedestrians and pigeons take off.
The head is gradually spreading from the centre to the surroundings.
Four frames and puzzles of the same animated, confirmed and then produced the entire piece.
Agent produces a frame from the same set of realizations when you have written the scene configuration and drawing program. The four readings, 1.8, 6.1, 10.7 and 17.6 seconds, were used to see the route along the street and the different blocks.
The early images were like a small building at the intersection, and shot was close to the fixed landscape. I wanted the density of the business centre, which was later filled with a series of buildings and secondary roads so that the new block could be seen along the way with shot.
The way the building is painted is also through the curve. The more small windows, windows, the more the roofs, the more wind, the more the wheels and the more the lights are, the more another low polygon model grows. The details were finally deleted, and the building was replaced by large monochromic surfaces and small windows, with the vehicle retaining a pure colour, a light, straight cockpit and a small solid wheel.
It's not so weird that the statues keep standing and a long coat outline and remove the sharp side face and the fragments. AI enthusiasm for detail sometimes requires manual dissuasion.
And then I filled Skill in for a level check. In addition to colouring, buildings, vehicles and characters will be scaled up to check profiles, scales and detail density separately.
The issue of dynamics is more specific. The main car stopped to allow pedestrians to enter the same comity sequence. Only the workshop was inspected and the vehicle was missing, and as a result it passed through.
Upon repair, mapping and security checks used the same set of role tracks, as well as checking the two-way traffic, the area occupied by the vehicle and the buffer time after the pedestrians had left. 20 seconds, 60fps corresponds to 1,200 time points, which are geometrically calculated before approval without having to render the whole video.
There's a dragon on the table. The original spread was over in 1.4 seconds, but the preview started in 1.8 seconds, and of course I could only see a complete picture. The starting point was changed to frame 0 and the expansion of the film was extended to approximately 2.6 seconds.
Four static maps can be viewed, and the starting point and crossing are subject to additional checks. I said, “Vehicles, architecture, pedestrians are ready”, and that means only approval of the models. When later changes are made, the car and the building are not drawn again by hand; the entire version is produced for further express approval.
From film to next use
The final picture is defined by certainty, i.e. by frame. pixels should be the same with the same frame. This way, jumps, reloads, or continues after interruption, and checks if the original image is not the same.
frame(300);
frame(720);
frame(300);
The same RGBA summary is required for the 300 frames twice, which is equivalent to a fingerprint matching of pixels.
Officially exported 1,200 frames that synthesize five sound steps of light music, street bottoms, footsteps and dove wings, coded H. 264 and MP4 for AAC. During the export, backstage PNG slows and the tag navigation is interrupted, and the synchronised PNG code in the same existing Chrome label continues from the verified 480 frame breakpoint.
Finally, four original pixels of the approved frame, summary, were aligned and completed all 1200 frames decoded in pieces. The impact, peaks and sound events are also measured. However, these numbers do not answer the sound of the music, and shot is bored with it and still needs to be watched and tested.
New Street first finished with the project script, and then I asked Agent to write back to Skill and upgrade it to v1.1.0. The new records are partially validated and approved in their entirety, and separate visual, sound and production versions. Only when the export tool is repaired can the existing frame be verified and continued; a change in the image or the plot would require re-confirmation.
The new scaffold has passed 20 Python tests and 10 Node tests, covering approvals, path limits, local receivers, breakpoint verifications and collision reversals. But it's not finished with a new theme, the actual Chrome-to-end production, and the specific scene and sound still need to be written by Agent. It is not possible to announce that all cities can be created by a single key because of a new street.
I can call it later:
/skill: memphos-mini-world, early market in Gangnam Water Village, boated under the Stone Bridge, and flowers bought off the coast.
The painting and production processes will continue, and the new scene will still need to be sampled and not accessible with the approval of the new street.
Turning back to the subject, I would like to organize the repeated tasks into Skill, because the rules and tools left by this return to work will be useful next time. Just do a simple one-time assignment, and I'll write prompt. As for whether a Skill is really better than the original approach, it is possible to contrast the mission with the official skill-creator; I can't judge just how well the catalogue is and how many lines SKILL.md is writing.
- Microscape Animation Skyll portal: memphis-mini-world
- Skill Design Assistant portal:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.