最近AI圈最实在的消息我觉得不是哪家又发了新模型而是WorkBuddy直接放了个“双模型限免”的操作Hy4 preview开放两周免费使用Hy3直接免费用到9月底。对天天折腾AI工具的人来说这种限免窗口其实比发布会更有含金量——尤其是Hy4 preview这种新模型官方愿意拿出来让你白嫖本身就是一次免费的深度评测机会。如果你还不清楚WorkBuddy能干什么这篇文章我尽量讲透包括它是什么、这次限免怎么用、两个模型怎么选以及我实际跑下来踩过的坑。
1. 先看懂这次双模型限免的含金量1.1 Hy4 preview 和 Hy3 是什么来头WorkBuddy底层接的是Hy系列模型这个系列一直走的是“通用大模型效率场景优化”的路线。Hy3是已经打磨了大半年的主力模型优点在于稳定、响应快、工具的调用链不容易断很多用户拿它做日常的自动化流程基本不会出幺蛾子。而Hy4 preview则是新放出来的预览版从命名就能看出来它是下一代模型的前哨站。预览版和正式版的区别我一般这样理解预览版就像是餐馆推出的隐藏菜单主厨愿意拿出来试做几道新菜给你尝味道有惊喜但偶尔也可能翻车。Hy4 preview在复杂推理、长文本理解、多步骤任务拆解上明显比Hy3强一截特别适合那些需要“想清楚再做”的任务。但因为是preview状态个别场景下响应速度会略慢偶尔也会有输出风格不稳定的情况。这次限免之所以值得关注就是你不用充值就能把这两个模型放在同一个工作台里横向对比。对我来说这种“肩并肩式评测”的机会比官方公布的benchmark有用得多因为实际跑自己的任务比任何跑分都有说服力。
1.2 “两周”和“9月底”为什么值得关注先看清楚规则Hy4 preview限免两周Hy3免费到9月底。时间设计很巧妙——两周的时间正好够你把Hy4 preview的脾气摸透赶在它恢复计费前决定要不要掏钱而Hy3给到9月底说明官方对老模型的定位是“长期饭票”主打一个稳定兜底。这里提个醒限免窗口期的起算时间每个账号可能不一样以你在配置模型时看到的剩余天数为准别想当然以为大家都是同一时间开始。我见过有人以为Hy4 preview还能免费用两周结果一查只剩三天白白浪费了排队配置的时间。拿到账号后第一件事我建议先去模型配置页看一眼倒计时。另一个容易被忽略的点是Hy3的免费到9月底但9月底之后怎么收费、会不会换新模型顶上目前还是未知数。我的建议很简单趁免费期把手头能自动化的流程都尽量跑一遍尤其是那种需要长期定时执行的任务先把效果验证了后面就算收费也能明确值不值得买。
2. WorkBuddy是干什么的值不值得为此装一个2.1 它不是又一个聊天框第一次打开WorkBuddy如果你期望看到ChatGPT那种“一问一答”的界面可能会有点懵因为它的核心思路根本不是聊天而是“搭工作台”。打个比方普通AI工具像一个很聪明但只能坐在工位上等你布置任务的实习生WorkBuddy更像是一个可以给这位实习生配上电话、通讯录、日程表和自动化流水线的“岗位系统”。你能给它定义角色、告诉它什么时候去干什么事、遇到不同情况怎么处理甚至让它跨应用干活。所以你打开WorkBuddy时看到的不是对话气泡而是技能Skill、连接器Connector、自动化流程、以及一个可以自定义指令的指挥中心。这么说吧如果你只是想偶尔让AI写一段文案、解一道题那WorkBuddy对你来说可能有点大材小用但如果你手头有“每天整理聊天记录”“定时把表格数据同步到另一个系统”“每周自动汇总周报”这类固定流程那WorkBuddy就是冲着解决这些问题来的。它适合的人群很明确需要和大量信息打交道、重复劳动多、愿意花一点时间搭建流程的效率控。
2.2 核心能力地图Skill、连接器、自定义指令要真正用起来WorkBuddy先搞清三个核心概念否则很容易转晕。Skill技能是WorkBuddy里最核心的模块相当于给AI预装的专业技能包。一个Skill可以封装一套完整的执行逻辑比如“整理聊天记录”的Skill它会自动读取选中的对话内容提炼要点按模板输出摘要。你不需要每次重新把需求描述一遍调起Skill就等于给AI下达了一套标准作业程序。连接器Connector是WorkBuddy能“伸手够到外部世界”的关键。通过连接器它可以访问钉钉多维表、Obsidian笔记、微信、文件目录等外部应用。这就是我之前说“让它跨应用干活”的基础。没有连接器AI只能在沙盒里自嗨接上连接器它才能把你的业务系统打通。自定义指令则是你给AI下的“最高指导原则”。你可以把常用的工作习惯、输出格式、处理偏好写成指令这样不管跑哪个SkillAI都会先遵循你的规则。比如我习惯所有输出都用Markdown摘要控制在300字以内这些偏好写进自定义指令后就不会每次重复交代。
2.3 和CodeBuddy有什么区别很多人会问WorkBuddy和CodeBuddy到底啥关系简单来说两者是同一个生态里分工不同的产品。CodeBuddy更聚焦在编程场景比如代码补全、代码解释、代码调试WorkBuddy则更偏向通用效率场景核心是“任务自动化信息处理”。在实战中我的感受是如果你要写代码打开CodeBuddy顺手如果你要把代码、文档、聊天记录、重复性杂活统一管起来WorkBuddy更合适。不少人是两个都装让它们各干各的——CodeBuddy管代码WorkBuddy管流程。另外有些操作并不是非此即彼WorkBuddy里也可以调起写代码类的Skill来做数据处理只是术业有专攻复杂工程还是交给CodeBuddy更省心。
3. 快速上手安装、登录、切换到Hy4 preview3.1 下载与安装各平台版本怎么选WorkBuddy的下载与安装整体比较省心官方主要提供Windows、macOS、Linux三套版本。Windows和macOS基本是下一步下一步的操作没什么好讲的。Linux环境下需要留意版本分成基于deb/rpm的发行版安装包和独立压缩包两种如果你用的是非主流发行版优先选压缩包解压后直接运行可执行文件就行。文中提到的麒麟版我实际在国产操作系统上跑过兼容性比预期好基本功能和主版本保持一致只是部分底层组件版本稍旧。安装时有几个容易被忽略的坑我拿出来说下安装路径不要带中文和空格尤其Windows用户我碰到过因为路径带空格导致本地服务起不来的情况。Linux下如果提示缺少依赖常见的是libgtk和libnss3相关库用发行版自带的包管理器补上就可以。首次启动可能会触发防火墙拦截记得放行否则连接器没法正常访问外部服务。
3.2 模型切换与限免确认WorkBuddy装好后第一步是登录账号。登录后进入设置里的“模型配置”页这里会列出当前账号可用的所有模型包括Hy3、Hy4 preview以及你额外接入的OpenAI或其他兼容接口的模型。切换到Hy4 preview的方法很简单在模型选择下拉列表里选中“Hy4 preview”保存即可。这时界面上通常会显示限免到期时间务必看清那个日期。如果列表里没有出现Hy4 preview可能是两个原因一是账号未获得灰度资格旧版客户端不展示新模型更新到最新版本就能看到二是网络同步问题退出账号重新登录刷新一下状态。需要提醒的是模型切换是全局生效的切换后所有Skill和自动化流程默认都会用新模型执行。如果你只想在个别任务上用Hy4 preview其他任务继续用Hy3那就不要改全局默认模型而是在每个Skill或流程的模型参数里单独指定。
3.3 写第一个自定义指令模型切好之后我建议先写一组自定义指令再开始跑实际任务。这是很多人偷懒跳过的一步但恰恰是决定WorkBuddy好不好用的分水岭。没有自定义指令的AI就像一个没有工作手册的新员工你问一句它答一句有了自定义指令它就变成了熟悉你套路的老手。模板直接参考我的# 通用规则 - 所有输出使用Markdown格式 - 中文回答除非内容本身是代码或英文专有名词 - 摘要类任务控制在300字以内 # 信息处理偏好 - 面对聊天记录先提炼结论再按时间线列出关键事件 - 面对表格数据先检查异常值再做统计描述 # 输出风格 - 避免空话套话直接给结论和行动建议 - 任何涉及操作步骤的内容分条列出保存后你的所有Skill都会遵循这套规则。这类“预设人设”的指令写一次后面省的是千百次重复交代的时间。
4. 实操场景让WorkBuddy真正帮你干活4.1 定时发送微信消息定时发送功能是我用WorkBuddy最频繁的场景之一。很多人一听“定时发微信”就以为是抢红包外挂其实正经用法是工作提醒、日报定时推送、或者给团队定时同步项目进度。实现方式并不复杂新建一个自动化流程触发条件选“定时”设定执行频率和具体时间点动作部分选择“发送微信消息”填入接收人/群和消息内容模板。你需要先在连接器里授权微信扫码登录后WorkBuddy才能以你的名义发消息。实测下来只要微信网页端不掉线发送成功率很高。这里有个很关键的经验消息内容里如果要包含动态数据比如当天的销售数据、待办事项需要通过变量引用数据源而不是写死。WorkBuddy支持在消息模板里注入变量变量值可以来自某个表格、某个Sheet或者一次AI生成的结果。我的做法是早上8点定时触发AI先拉取前一天项目完成情况再生成一段100字左右的项目播报最后通过微信推送给团队。整套流程不到20分钟就搭好了。
4.2 整理聊天记录整理聊天记录是我最初安装WorkBuddy的直接原因。工作群聊一天几百条消息爬楼翻记录真的会让人崩溃。WorkBuddy配合“整理聊天记录”Skill可以自动把海量对话压缩成结构化摘要——谁、什么时候、做了什么、下一步要干什么。实际使用时我一般把聊天记录导出成文本或直接通过连接器读取指定群聊然后调起整理Skill按我自定义指令中偏好的格式输出先给结论再列时间线最后标出待办事项。需要强调的是聊天记录整理这类任务涉及隐私边界建议只处理工作相关的群聊并且确认群内成员都知情。我不是在说教而是这个线上工作场景真的有合规问题别因为方便给自己惹麻烦。另外长聊天记录分段处理的效果会比一次性全量塞给模型好我通常按500行左右切段整理完再合并去重。
4.3 钉钉多维表同步钉钉多维表在不少公司里是项目管理的中枢但跨系统同步一直是痛点。WorkBuddy的连接器可以直接对接钉钉多维表实现定期同步。比如我这边有一个需求是从企业微信的客户跟进表定时把新增记录同步到钉钉多维表里再给相关负责人发通知。搭建步骤大致是先在连接器里授权钉钉应用配置好多维表的访问权限然后新建自动化流程触发条件选“定时”或“数据变更”动作部分选“读取数据源A写入目标表B”。WorkBuddy会按你设定的映射规则把字段一一对应写入。这类跨应用数据同步最怕的是字段类型不匹配。日期字段在A系统里是文本“2025-08-01”到B系统要求是日期类型直接写进去就会失败。我的经验是在流程里加上“数据清洗”中间步骤先用一个小的代码Skill统一字段格式再执行写入。这样稳定性会高很多踩过一次坑之后我就养成习惯了。
5. Hy4 preview 和 Hy3实际用起来差别在哪5.1 能力对比速查我花了一周时间在两个模型之间反复横跳在同样的任务上对比了代码生成、长文总结、多步推理、工具调用稳定性四个维度。直接上结论对比维度Hy3Hy4 preview响应速度快体感接近即时稍慢复杂任务尤其明显代码生成常规任务不错复杂逻辑容易绕明显更稳能处理更多依赖关系长文本总结2000字内效果好再长会丢细节能hold住更长上下文信息保留更完整多步推理偶尔断链需要手动澄清链路完整能自己拆解步骤工具调用稳定高很少出错略有不稳定个别场景会多调用或少调用输出风格稳定统一偶尔有“创造性发挥”需要校对5.2 按场景选择的建议基于上面这些观察我自己总结了一套选模型的口诀日常用Hy3烧脑用Hy4。所谓日常就是定时发送、表格同步、简单信息提取这类流程明确、步骤固定的任务。用Hy3的稳定性和响应速度来解决既不会出岔子执行效率也高。所谓烧脑就是长文档分析、代码重构、涉及复杂判断的决策支持这类任务交给Hy4 preview你能明显感觉到它在“想得更多、更细”。如果你的任务本来就是小步骤重复劳动强行上Hy4 preview唯一能感受到的变化就是响应变慢。反过来如果任务很复杂却还在用Hy3你可能得反复澄清整体体验反而不如直接上Hy4 preview一次说清。限免期间我的建议是所有任务都先并行跑一遍把你的典型任务分别在两个模型里执行用真实输出对比决定后续付费策略。别听别人说哪个好就无脑选哪个自己跑过才靠谱。
6. 高频问题与排查技巧6.1 常见问题速查表把这段时间看到的、我经历过的典型问题整理成一个速查表方便大家直接对照问题可能原因解决办法模型列表里没有Hy4 preview版本过旧或未获灰度资格更新客户端退出账号重新登录Linux启动报缺少依赖系统缺libgtk、libnss3等库用包管理器安装后重启定时任务到了时间没触发客户端未保持后台运行将WorkBuddy加入开机启动微信消息发送失败微信网页端掉线打开连接器重新扫码授权钉钉字段写入失败字段类型不匹配在流程中加入数据清洗中间步骤本地服务端口被占用默认端口冲突在设置中修改本地服务端口“Claw”面板不显示界面布局里被隐藏去视图/显示设置里勾选显示6.2 三个容易踩的坑第一目录前面有个点别手贱删。WorkBuddy会在用户目录下生成一个点开头的隐藏配置目录用来存放本地配置和缓存。有些人有洁癖看到点开头的目录就想清理结果删掉后配置全丢所有Skill和连接器都要重新绑定。这个目录正常情况下占用不大没必要动。真出问题宁可卸载重装也别去手动删配置目录。第二全局模型切换会影响所有流程改之前想清楚。我自己犯过一次错为了测试Hy4 preview的长文本能力把全局默认模型切了过去结果当天好几个定时任务全部走了新模型有一个报表生成脚本因此输出格式异常。后来所有自动化流程都单独指定模型全局我只保留Hy3避免误伤。第三限免到期前记得截图存档你的常用配置和示例输出。Hy4 preview这种新模型免费期结束后你未必会立刻付费订但你可能还想参考它某次输出。别等到到期了才发现连之前跑出的好结果都没存到期后模型在界面上直接变成灰色不可选只能干瞪眼。最后分享一点个人体会这次双模型限免我的建议是别只当“薅羊毛”把它当成一次低成本的效率实验。哪怕你平时没有明确的自动化需求也可以趁这两周把人工重复最多的那件事交给WorkBuddy跑一遍试试体验一下“早上打开电脑事情已经干完了”的感觉。Hy4 preview和Hy3各有擅长的领域趁免费期找到它们在你实际工作中的边界比任何榜单都实在。我的最后一个小技巧在WorkBuddy里建一个“模型对比笔记”的Skill把同一个问题分别用Hy3和Hy4 preview回答再让AI自己对比两份回答的差异。这样做你会发现很多官方参数看不出来的差别实际一跑就清清楚楚了。限免期结束前记得把有价值的对话记录下来。
本文来源：https://kwkr.cn/news/342900
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
The most practical news in the AI space recently, in my opinion, isn't which company released a new model—it's that WorkBuddy directly dropped a "dual-model limited-time free access" move. Hy4 preview is free to use for two weeks, and Hy3 is free to use until the end of September. For people who tinker with AI tools every day, these free-trial windows are actually more valuable than launch events—especially when a new model like Hy4 preview is offered for free by the official team. That itself is a free in-depth evaluation opportunity. If you're still not sure what WorkBuddy can do, I'll try to explain it thoroughly in this article, including what it is, how to use this limited-time free access, how to choose between the two models, and the pitfalls I actually ran into during my testing.
1. First, understand the value of this dual-model limited-time free access. 1.1 Where do Hy4 preview and Hy3 come from? WorkBuddy's underlying model is the Hy series, which has always followed the route of "general large-model efficiency-scenario optimization." Hy3 is the main model that has been polished for over half a year. Its strengths are stability, fast response, and tool-calling chains that rarely break. Many users run their daily automation workflows with it and basically never run into issues. Hy4 preview, on the other hand, is the newly released preview version. As the name suggests, it's the vanguard of the next-generation model. I generally understand the difference between a preview version and a stable version like this: a preview version is like a hidden menu at a restaurant. The chef is willing to test a few new dishes for you to taste—there might be surprises, but sometimes it could also fail. Hy4 preview is noticeably stronger than Hy3 in complex reasoning, long-text comprehension, and multi-step task decomposition. It's especially suited for tasks that require "thinking it through before acting." But because it's in preview status, response speed can be slightly slower in some scenarios, and occasionally the output style can be inconsistent. The reason this limited-time free access is worth paying attention to is that you can put both models on the same workbench and compare them side by side without spending any money. For me, this kind of "side-by-side evaluation" opportunity is far more useful than officially published benchmarks, because running your own tasks is more convincing than any benchmark score.
1.2 Why are "two weeks" and "end of September" worth paying attention to? First, let's be clear about the rules. Hy4 preview is free for two weeks, and Hy3 is free until the end of September. The timing is clever—two weeks is just enough time to figure out Hy4 preview's temperament and decide whether to pay before it goes back to being paid. And Hy3 being free until the end of September shows that the official team's positioning for the older model is as a "long-term meal ticket," mainly providing a stable safety net. A heads-up here: the starting time of the limited-time free window may differ for each account. Refer to the remaining days shown when you configure the model—don't assume everyone started at the same time. I've seen people who thought Hy4 preview was still free for two weeks, only to check and find just three days left, wasting time waiting in line to configure it. After you get your account, the first thing I recommend is to go to the model configuration page and check the countdown. Another easily overlooked point is that Hy3 is free until the end of September, but how it will be priced after that, and whether a new model will replace it, is still unknown. My advice is simple: while it's free, run every workflow you can automate as much as possible, especially tasks that need to be executed regularly over a long period. Verify the results first, so that even if it becomes paid later, you'll know clearly whether it's worth buying.
2. What does WorkBuddy do, and is it worth installing? 2.1 It's not just another chat box. When you first open WorkBuddy, if you expect to see a ChatGPT-style "one question, one answer" interface, you might be a bit confused—because its core idea isn't chat at all, but "building a workbench." To use an analogy, an ordinary AI tool is like a very smart intern who can only sit at a workstation and wait for you to assign tasks. WorkBuddy is more like a "job system" that equips that intern with a phone, contacts, a schedule, and an automated assembly line. You can define roles for it, tell it when to do what, how to handle different situations, and even let it work across applications. So when you open WorkBuddy, you don't see chat bubbles—you see Skills, Connectors, automation workflows, and a command center where you can customize instructions. In other words, if you only want to have AI write a piece of copy or solve a problem occasionally, WorkBuddy might be overkill for you. But if you have fixed workflows like "organize chat records every day," "periodically sync spreadsheet data to another system," or "automatically compile a weekly report every week," then WorkBuddy is built precisely for solving those problems. The target audience is clear: efficiency enthusiasts who deal with large amounts of information, have a lot of repetitive work, and are willing to spend some time setting up workflows.
2.2 Core capability map: Skills, Connectors, and custom instructions. To really use WorkBuddy well, you first need to understand three core concepts, or you'll easily get lost. Skill—Skills are the most core module in WorkBuddy, essentially a pre-installed professional capability package for the AI. A Skill can encapsulate a complete execution logic. For example, a Skill for "organizing chat records" will automatically read the selected conversation content, distill key points, and output a summary according to a template. You don't need to describe the requirements from scratch every time—activating a Skill is equivalent to giving the AI a standard operating procedure. Connector—Connectors are the key to WorkBuddy being able to "reach out to the external world." Through connectors, it can access external applications such as DingTalk multidimensional tables, Obsidian notes, WeChat, and file directories. This is the foundation of what I mentioned earlier about "letting it work across applications." Without connectors, the AI can only entertain itself in a sandbox. Once connected, it can integrate your business systems. Custom instructions are the "highest guiding principles" you give to the AI. You can write your common work habits, output formats, and processing preferences into instructions, so no matter which Skill you run, the AI will first follow your rules. For example, I'm used to all outputs being in Markdown format, with summaries kept under 300 characters. Once these preferences are written into custom instructions, I don't have to repeat them every time.
2.3 What's the difference from CodeBuddy? Many people ask what the relationship is between WorkBuddy and CodeBuddy. Simply put, they are two products with different roles in the same ecosystem. CodeBuddy is more focused on programming scenarios, such as code completion, code explanation, and code debugging. WorkBuddy leans more toward general efficiency scenarios, with its core being "task automation and information processing." In actual use, my feeling is that if you want to write code, opening CodeBuddy is more natural. If you want to manage code, documents, chat records, and repetitive miscellaneous tasks together, WorkBuddy is more suitable. Many people install both and let them do their own things—CodeBuddy handles code, WorkBuddy handles workflows. Also, some operations aren't strictly either-or. WorkBuddy can also invoke coding-type Skills for data processing, but each tool has its own expertise. For complex engineering, it's more reliable to leave it to CodeBuddy.
3. Quick start: installation, login, and switching to Hy4 preview. 3.1 Download and installation—how to choose versions for each platform. WorkBuddy's download and installation are generally straightforward. The official team mainly provides three versions: Windows, macOS, and Linux. Windows and macOS are basically "next, next, finish" operations—nothing much to say. On Linux, you need to pay attention to versions: there are both packages based on deb/rpm distributions and standalone tarballs. If you're using a less mainstream distribution, prioritize the tarball—just extract it and run the executable directly. As for the Kylin version mentioned in the article, I've actually run it on a domestic operating system, and compatibility was better than expected. Basic functionality remained consistent with the main version, though some underlying component versions were slightly older. There are a few pitfalls during installation that are easy to overlook, so let me mention them. The installation path should not contain Chinese characters or spaces—especially for Windows users. I've encountered cases where the local service failed to start because the path contained a space. On Linux, if you get a prompt about missing dependencies, the common ones are libgtk and libnss3 related libraries. You can install them using your distribution's package manager. On first launch, the firewall might prompt you for permission. Remember to allow it, or connectors won't be able to access external services normally.
3.2 Model switching and confirming the free access. After installing WorkBuddy, the first step is to log in. After logging in, go to the "Model Configuration" page in settings. This page lists all models available to your account, including Hy3, Hy4 preview, and any additional OpenAI or other compatible API models you've connected. Switching to Hy4 preview is simple: select "Hy4 preview" from the model dropdown list and save. At this point, the interface usually shows the free-access expiration date—make sure to read that date carefully. If Hy4 preview doesn't appear in the list, there are two possible reasons. First, your account may not have received gray-release access; old client versions don't show new models, so update to the latest version and you'll see it. Second, it could be a network sync issue—log out and log back in to refresh the status. One thing to note: model switching takes effect globally. After switching, all Skills and automation workflows will use the new model by default. If you only want to use Hy4 preview for certain tasks while keeping Hy3 for others, don't change the global default model. Instead, specify the model individually in the model parameters of each Skill or workflow.
3.3 Writing your first custom instructions. After switching the model, I recommend writing a set of custom instructions before running actual tasks. This is a step many people skip out of laziness, but it's precisely the dividing line that determines whether WorkBuddy is useful or not. An AI without custom instructions is like a new employee without a work manual—you ask, and it answers. With custom instructions, it becomes a veteran who already knows your ways. Feel free to reference my template: # General rules - All output uses Markdown format - Answer in Chinese, unless the content itself is code or English proper nouns - Summary tasks should be kept within 300 characters # Information processing preferences - When dealing with chat records, first extract conclusions, then list key events chronologically - When dealing with tabular data, first check for outliers, then provide statistical descriptions # Output style - Avoid empty talk and clichés; give conclusions and actionable suggestions directly - Any content involving operational steps should be listed item by item. After saving, all your Skills will follow these rules. This kind of "preset persona" instruction—once written—saves you hundreds or thousands of times of repeated explanations.
4. Practical scenarios: let WorkBuddy actually do work for you. 4.1 Scheduled WeChat message sending. Scheduled sending is one of the scenarios where I use WorkBuddy most frequently. Many people, when they hear "scheduled WeChat sending," immediately think of a red-envelope-grabbing cheat. Actually, the legitimate uses are work reminders, scheduled daily report pushes, or periodically syncing project progress to the team. The implementation isn't complicated. Create a new automation workflow, set the trigger condition to "scheduled," configure the execution frequency and specific time point, select "Send WeChat message" as the action, and fill in the recipient/group and message content template. You'll first need to authorize WeChat in the connector. After scanning the QR code to log in, WorkBuddy can send messages on your behalf. Based on my testing, as long as the WeChat web client doesn't go offline, the sending success rate is very high. Here's a crucial piece of experience: if the message content needs to include dynamic data—such as today's sales data or to-do items—you need to reference variables from a data source rather than hardcoding them. WorkBuddy supports injecting variables into message templates. Variable values can come from a table, a sheet, or the result of an AI generation. My approach is to set a scheduled trigger at 8 a.m. The AI first pulls the previous day's project completion status, generates a project brief of around 100 words, and then pushes it to the team via WeChat. The whole workflow took less than 20 minutes to set up.
4.2 Organizing chat records. Organizing chat records is the original reason I installed WorkBuddy. When work groups have hundreds of messages a day, scrolling through the history to find things is genuinely maddening. WorkBuddy, combined with the "Organize Chat Records" Skill, can automatically compress massive conversations into structured summaries—who, when, what they did, and what the next step is. In practice, I usually export chat records as text or directly read a specific group chat through a connector, then activate the organization Skill to output according to the preferences in my custom instructions—conclusions first, then a timeline, and finally to-do items highlighted. It's important to emphasize that tasks like organizing chat records involve privacy boundaries. I recommend only processing work-related group chats and confirming that group members are aware. I'm not being preachy—this online work scenario genuinely has compliance issues. Don't get yourself into trouble for the sake of convenience. Additionally, processing long chat records in segments works better than feeding the entire thing to the model at once. I usually split them into chunks of about 500 lines, organize each chunk, then merge and deduplicate.
4.3 DingTalk multidimensional table sync. In many companies, DingTalk multidimensional tables serve as the hub for project management, but cross-system synchronization has always been a pain point. WorkBuddy's connector can directly integrate with DingTalk multidimensional tables to achieve regular sync. For example, on my end, I had a need to periodically sync new records from a customer follow-up table in WeCom to a DingTalk multidimensional table, and then notify the relevant person in charge. The setup steps are roughly as follows: first, authorize the DingTalk app in the connector and configure access permissions for the multidimensional table; then create a new automation workflow, set the trigger condition to "scheduled" or "data change"; in the actions, select "read data source A and write to target table B." WorkBuddy will map fields one by one according to the mapping rules you set. The thing this kind of cross-application data sync fears most is field type mismatches. A date field might be text like "2025-08-01" in system A, but system B requires a date type. Writing it directly will fail. My experience is to add an intermediate "data cleaning" step to the workflow—first use a small code Skill to normalize the field format, then perform the write. This makes stability much higher. After stepping on that landmine once, I got into the habit.
5. What's the real difference between Hy4 preview and Hy3 in practice? 5.1 Capability comparison quick reference. I spent a week bouncing back and forth between the two models, comparing them on the same tasks across four dimensions: code generation, long-text summarization, multi-step reasoning, and tool-calling stability. Here are the conclusions directly. Comparison dimension: Hy3 vs Hy4 preview. Response speed: Hy3 is fast, feeling near-instant; Hy4 preview is slightly slower, especially noticeable in complex tasks. Code generation: Hy3 handles routine tasks well, but complex logic easily gets convoluted; Hy4 preview is noticeably more stable and can handle more dependencies. Long-text summarization: Hy3 works well within 2,000 characters, but loses details beyond that; Hy4 preview can handle longer contexts and preserves information more completely. Multi-step reasoning: Hy3 occasionally breaks the chain and needs manual clarification; Hy4 preview maintains a complete chain and can decompose steps on its own. Tool calling: Hy3 is highly stable with rare errors; Hy4 preview is slightly unstable, occasionally calling too many or too few times in certain scenarios. Output style: Hy3 is stable and consistent; Hy4 preview occasionally has "creative flourishes" that need proofreading. 5.2 Scenario-based selection advice. Based on these observations, I've developed my own formula for choosing models: use Hy3 for daily tasks, use Hy4 for brain-intensive tasks. "Daily" means tasks like scheduled sending, table sync, and simple information extraction—tasks with clear processes and fixed steps. Hy3's stability and response speed handle these well without hiccups, and execution efficiency is high. "Brain-intensive" means tasks like long-document analysis, code refactoring, and decision support involving complex judgments. With Hy4 preview, you can clearly feel that it's "thinking more and in more detail." If your task is inherently small-step repetitive labor, the only change you'll notice by forcing Hy4 preview is slower responses. Conversely, if the task is complex but you're still using Hy3, you'll likely need to repeatedly clarify, and the overall experience will actually be worse than just using Hy4 preview and explaining everything once. During the free-access period, my advice is to run all your tasks through both models first. Execute your typical tasks on each model separately and compare the real outputs to decide your future payment strategy. Don't blindly choose whichever one others say is better—only what you've tested yourself counts.
6. Frequently asked questions and troubleshooting tips. 6.1 Common issues quick reference table. Let me organize the typical problems I've seen and experienced during this period into a quick reference table so everyone can check directly. Issue: Hy4 preview not in the model list. Possible cause: version too old or no gray-release access. Solution: update the client, log out and log back in. Issue: Linux fails to start with missing dependencies. Possible cause: system missing libgtk, libnss3 libraries. Solution: install via package manager and restart. Issue: scheduled task didn't trigger at the scheduled time. Possible cause: client not running in the background. Solution: add WorkBuddy to startup items. Issue: WeChat message sending failed. Possible cause: WeChat web client went offline. Solution: open the connector and re-scan the QR code to reauthorize. Issue: DingTalk field write failure. Possible cause: field type mismatch. Solution: add a data cleaning intermediate step to the workflow. Issue: local service port occupied. Possible cause: default port conflict. Solution: change the local service port in settings. Issue: "Claw" panel not displayed. Possible cause: hidden in interface layout. Solution: go to view/display settings and check the option to show it. 6.2 Three easy pitfalls. First: the directory starting with a dot—don't delete it out of habit. WorkBuddy creates a hidden configuration directory starting with a dot in the user directory, used to store local configuration and cache. Some people with a tidiness compulsion see dot-directories and want to clean them up, but after deleting it, all configurations are lost, and all Skills and connectors need to be re-bound. This directory normally doesn't take up much space, so there's no need to touch it. If there's really a problem, it's better to uninstall and reinstall than to manually delete the configuration directory. Second: global model switching affects all workflows—think carefully before changing it. I made this mistake myself. To test Hy4 preview's long-text capability, I switched the global default model, and as a result, several scheduled tasks that day all went through the new model. One report generation script had abnormal output formatting because of this. After that, I specified models individually for all automation workflows and kept only Hy3 as the global default to avoid collateral damage. Third: before the free-access period expires, remember to take screenshots and archive your common configurations and example outputs. After Hy4 preview's free period ends, you might not immediately subscribe to it, but you may still want to reference a particular output it produced. Don't wait until it expires to realize you didn't save the good results it produced earlier. After expiration, the model will appear grayed out and unselectable in the interface, and you can only stare at it. Finally, let me share a personal thought. For this dual-model free access, my advice is: don't just treat it as "getting free stuff"—treat it as a low-cost efficiency experiment. Even if you don't have clear automation needs in daily life, you can still use these two weeks to hand your most repetitive manual task to WorkBuddy and try it once, to experience the feeling of "opening your computer in the morning and finding the work is already done." Hy4 preview and Hy3 each have their own areas of strength. Using the free period to find their boundaries in your actual work is more practical than any ranking list. My final tip: create a "Model Comparison Notes" Skill inside WorkBuddy. Have Hy3 and Hy4 preview answer the same question separately, then let the AI compare the differences between the two responses. Doing this, you'll discover many differences that official parameters can't reveal—they become crystal clear as soon as you actually run them. Before the free-access period ends, remember to save the valuable conversations.