DEV Community: GokuScraper悟空爬虫

两个人做的Alook浏览器，凭什么AppStore付费榜第一？

GokuScraper悟空爬虫 — Thu, 03 Sep 2026 06:11:25 +0000

两个人做的Alook浏览器，凭什么AppStore付费榜第一？

一个只有 2 人的团队，在河北安国小县城，做出了一个 iOS 工具付费榜第 1 的浏览器。前身 Take 浏览器上线一个月就被苹果下架，然后换个名字东山再起，反而做成了 iOS 标杆。

年收入大概在 ¥60~80 万，人均三四十万，不比大厂码农差，不过喜马拉雅的一纸诉状也让它知道——踩了版权红线，迟早要还。

这是个啥软件？

Alook 浏览器，一款主打「极简」的移动端浏览器，官网自己贴的标签是「三无产品」——无推送、无新闻、无广告，毫秒级启动。iPhone 和安卓俩平台都上，iOS 端付费，安卓端免费。

它最狠的不是能搜索，而是把一个浏览器塞进了播放器、电子书阅读器、文件管理器三个工具：

音视频全能：网页视频能 0.5~16 倍速播放，支持悬浮窗、后台播放、画中画、投屏，还能嗅探下载网页里的视频和音频，追剧党直接把它当播放器用。
内置 Adblock Plus：默认屏蔽侵入式广告，还能手动标记广告、加载第三方规则，网页瞬间清净。
全格式电子书阅读器 + 文件管理：txt、pdf、epub、mobi、azw3 这些格式全支持，还能解压 zip/rar/7z、Wi-Fi 传文件，一个浏览器顶一个工具箱。

别人做的是「浏览器 + 信息流」，Alook 做的是「浏览器 + 工具箱」，把打扰用户的全砍了，把有用的全塞进来。iOS 端 17 万个评分，平均4.8 分，用户对它的评价很高。

开发者是谁？目前团队规模多大？

开发者是保定乐活网络科技有限公司，法人代表牛荣娜，2014年12月注册，注册资本100万、实缴100万，有限责任公司，注册地址在河北省保定市安国市。

团队规模：天眼查显示参保人数只有 2 人。 俩人在河北安国的小地方，做出了一个 iOS 工具付费榜第 1 的浏览器。另外还查到一家天津乐活网络科技有限公司，2022年成立，同一法人牛荣娜，参保0人——大概率是运营/合规拆分出来的壳，不承担实际开发。

Alook 最早是 iOS 应用，2017年7月上线，一开始就走的付费买断制。后来才出了安卓版，安卓版免费。

开发者从哪来的想法？

开发者没有公开的访谈博客，但可以从产品演变和 V2EX 上的发帖拼出完整图景。

Alook 不是第一个产品，是同一个 App 换了个名字重来。 这家公司先做了一款叫「Take浏览器」的 iOS 浏览器，也是极简风格、也是无推送无广告，但上架后大约一个月就被下架了。证据链很硬：Take 和 Alook 的 App Store ID 都是 1261944766，包名到今天还是 com.ld.TakeBrowser——同一个 App，下架后改名重上，本质是"吃了一堑长一智"。2017年7月苹果在中国区大规模清理 VPN 类应用，时间点正好卡在 Alook 上线（2017-07-29）之前，很可能 Take 也是那波清理中被下架的。

具体想法从哪来？ 开发者自己在 V2EX 说的很清楚——"追剧和看教程的时候有的地方进展太慢，2倍速既可以节省时间，又能保证不丢失剧情和内容。"

他自己就是 iOS 重度用户，受不了主流浏览器又慢又臃肿，又想做视频倍速和悬浮播放——当时 iOS 上几乎没有浏览器能好好做这个功能。再加上"百度首页的信息流屏蔽、智能拼页"这些用户痛点，干脆自己动手。

目标用户是谁？

Alook 没做官方用户画像，但从它反复打磨的功能和用户自发的使用场景，能清楚看出它抓的是哪几拨人：

第一拨：学生党，尤其是考研党。 最典型的使用场景是百度网盘倍速——考研网课存网盘，不开会员不能倍速，Alook 登录网盘网页版切电脑模式就能 2 倍速起跳，还不花钱。当年连医学生考研号都专门写文章推它，宣传点就是"网盘视频免会员倍速看网课"。加上看小说、翻译外文文献，学生党是主力中的主力。

第二拨：追剧党和视频重度用户。 网页视频倍速、悬浮窗边聊边看、后台播放、投屏，这套功能打的就是那些不想充视频会员、又要看剧看综艺看课的人。V2EX 老哥一句"alook 看片贼好用，懂得都懂"，懂的都懂。

第三拨：极简控和隐私敏感人群。 受不了 UC、夸克那种满屏信息流和推送的人，把"三无"当信仰。这类用户往往有点技术底子，会自己去写 JS 脚本、自定义规则，是社区的免费自来水。

它的用户画像不是"所有人"，而是"对手机浏览器有明确要求、愿意为纯净体验掏钱/花时间折腾的人"。泛用户用 Safari、用系统浏览器就够了，Alook 要的是那群"受不了将就"的人。

怎么拿到第一批用户？

第一批用户来自两个地方：App Store 自然搜索 + V2EX 社区。

一是上架后的自然流量。 2017年7月 Alook 在 App Store 上线，分类在"工具"，定价 ¥12 买断。只要用户在 App Store 搜"浏览器"或者翻付费榜，就能看到它。它本身功能过硬、评分高，搜到的用户下载后觉得好用就会留下来，这是最基础的底量。

二是开发者亲自去 V2EX 发帖。 上架后没多久，开发者就用账号 alook 在 V2EX 发帖介绍这款浏览器：在线加速音视频、2 倍速看视频、听教程省时间。11792 次浏览，评论区用户直接跟开发者互动。V2EX 的用户多是程序员和技术爱好者，恰好是"对浏览器有要求"的人群，跟 Alook 的目标用户高度重合。这批人是最早的口碑种子。

种子用户运营也没落下。 开发者在 V2EX 上逐条回复评论，连"视频下载"这种苹果不许可的功能，都用 TestFlight 邀请制私下发给老用户。后来 App Store 评论区那些五星自来水，不少就是这批铁杆用户。

怎么做用户增长的？

不买广告，不运营自媒体，增长全靠产品硬扛。

自媒体这块，Alook 几乎等于没做。 个别平台有账号，但主要功能是客服入口，不是内容运营。没有公众号、没有小红书、没有抖音。官网就是一张产品功能页。对于 2 人团队，这反而是合理的：人不够，就省了运营这个岗位。

那用户怎么放量？两大招。

第一招：限免 + 促销，反复收割 App Store 榜单。 免费榜是 App Store 最大的流量入口，付费 App 一限免就能冲榜，冲榜就有自然新增。2018年6月端午节限免撞上世界杯，爱范儿（AppSo）跟着发了评测，少数派、福利吧、壹读一窝蜂跟进，单日下载量逼近 30 万，冲进免费总榜 Top 5。之后逢节点就促销：双十一 ¥18→¥6、2020年3月半价 ¥12→¥6，这套路三年反复用，次次有效。

第二招：安卓免费版，三年憋个大招。 2020 年 7 月安卓版上线，直接免费。iOS 用户憋了三年，终于等到了安卓版，第一时间奔走相告。媒体又是一波跟进报道，安卓用户疯狂涌入。免费版天然自带传播：iOS 用户安利给安卓朋友，安卓用户又安利给其他人。

安卓免费版一上线，才真正把 Alook 从"iOS 小众精品"变成了"跨平台有量产品"。

靠什么赚钱？

Alook 的变现方式不是一眼能看出来的，因为它没有广告、没有会员、没有订阅，甚至连内购都没有。

我看了最新安卓版（10.8）SDK 列表：AndroidX、Glide、Material Design、微信 SDK、RxJava、Kotlin 协程、OkHttp、Chromium WebView——没有 AdMob、没有穿山甲、没有广点通、没有 Google Play 计费。 一个广告 SDK 都没有。iOS 端同样没有植入广告。

那它怎么活？拼起来看，收入来源是这三块：

第一块：iOS 付费买断，基本盘。 定价 ¥6（$0.99），2017 年上线以来一直这个价位。付费用户虽然不多（对比免费浏览器），但胜在稳定——App Store 上架了 9 年，iOS 付费工具榜常年在榜。这是最直接、最干净的收入，每月几百到几千美金的持续性收入。

第二块：搜索引擎分成，藏在暗处。 用户在 Alook 里用百度搜索时，Alook 会在链接后面自动添加追踪标识符 ?from=1015011a。开发者 V2EX 上亲口承认："确实是为了赚钱，加追踪符给我们带来一些收益的前提下也不会影响用户体验。" 不止百度，所有搜索引擎除了国外那几个，全都加了追踪。每有一个用户通过 Alook 搜索，百度就要给 Alook 一笔渠道分成。虽然单个用户不值钱，但累积起来是笔细水长流的收入。

第三块：导航页流量变现。首页搜索框下方能看到"秘塔 AI"这类入口。

克制极简的 Alook 为什么在寸土寸金的首页留这个位置？合理的推测是收了坑位费——Alook手里有高密度的年轻用户，正是 AI 公司最想要的精准流量。

三块加起来，够 2 个人活，但不够暴富。 没有广告 SDK、没有内购、没有订阅、没有会员——这在iOS浏览器里几乎是独一份。代价是团队规模一直卡在 2 人，没有余力做大版本迭代和运营。

一年赚多少？

先说数据底子，所有数字都来自第三方数据平台七麦（2026 年 9 月），不是苹果官方公布的数字——苹果从不公开单个 App 的下载量，所以这里的数据要分清两种性质：

排名是真的： 中国区 App Store 工具付费榜第 1、应用付费总榜实时第 7，这是七麦直接抓 App Store 公开榜单拿到的，是真实数据。
下载量是预估的： 昨日下载量 432（iPhone）、单价 ¥6、无内购，这些是七麦通过"排名反推下载量"的模型估算出来的，行业误差一般在 10%~20%，能看方向，别当精确数。

更关键的一点：下载量 ≠ 购买量。 七麦的下载量是"总下载次数"，里面混着不花钱的部分——已购用户换手机、重装 App 时重新下载，不算新购买（Apple 官方定义：同一账号换机下载和重装不计入销量）。所以不能拿"下载量 × 单价"直接当收入。

但有一个更硬的指标：App Store 付费榜排的是"付费下载"，每一次上榜背后都是一笔真实购买。 工具付费榜第 1，说明近期中国区所有付费工具 App 里，买 Alook 的人最多，这个排名比下载量更能说明收入规模。

算账（iOS 中国区为主，用量级不钻精确数）：

昨日下载 432（iPhone）+ 128（iPad）= 560 次，近 30 日日均约 611 次，近一年合计约 15.5 万次（以上均为七麦预估）
按单价 ¥6：15.5 万 × ¥6 = ¥93 万（税前，这是"全算新购"的上限）
苹果抽成 30% → 上限约 ¥65 万到手
但实际新购占比低于 100%（已购用户重装占了部分），所以保守取 ¥50~70 万，更符合实际

零散收入还有两块，估个量级：

海外区零星下载：那些上榜的小国市场，一年估几万块
搜索引擎渠道分成：百度小渠道号那套，一年估几万块
安卓免费版：不直接赚钱，但养口碑、引流

结论按三档给，每档把三块收入拆开算：

悲观： 日下载掉到 200，中国区 iOS 到手约 ¥30 万 + 海外零星 ≈ ¥1~2 万 + 渠道分成 ≈ ¥1~2 万 → 合计 ¥30~40 万
最有可能： 中国区 iOS 到手 ¥50~70 万（核心基本盘） + 海外零星 ≈ ¥3~5 万 + 渠道分成 ≈ ¥3~5 万 → 合计 ¥60~80 万
乐观： 日下载 600+ 且新购占比高，中国区 iOS 到手 ¥80 万以上 + 海外零星 ≈ ¥5~10 万 + 渠道分成 ≈ ¥5~10 万 → 合计 ¥90~110 万

一个 2 人团队，一年大几十万，人均三四十万，比打工强，但也没暴富。这就是"小而美"的真实样子——体面，但不神话。

彪哥的总结

Alook 这个软件，三个字总结：小而美，但踩过线。

定位克制是最大的武器。 2 个人、河北安国小地方，九年就磨一个"无推送、无新闻、无广告"的浏览器，把视频倍速、悬浮、下载做到 iOS 第一。不碰信息流、不塞广告、不做会员，克制本身就成了护城河。
冷启动靠杠杆，不靠买量。 限免冲榜、产品口碑、开发者亲自下场做客服，一分钱广告费没花，做成了 iOS 工具付费榜第 1。
但踩版权红线，迟早要还。 视频下载、嗅探、去广告，这些让用户叫好的功能，也是版权方最恨的"侵犯版权行为"。喜马拉雅 2022 年告它，同类判例赔几十上百万。安卓版后来砍视频下载，就是付学费。做工具可以，算清法律账再动手。

我是爬虫彪哥，如果这篇拆解对你有启发，点个赞或者转发给身边的兄弟们。

要是觉得哪块没拆透，或者有啥补充，欢迎评论区接着聊。

下期预告：咱们拆阿里P8做的Mole，从大厂高管到独立开发者，看看他能整出啥花样……敬请期待。

投稿或合作，请联系邮箱：chai@gokuscraper.com

主打拦截广告的X浏览器，靠什么吃饭？

GokuScraper悟空爬虫 — Tue, 01 Sep 2026 23:08:30 +0000

一个主打帮用户拦截广告的浏览器，自己靠什么赚钱呢？

答案有点黑色幽默：靠广告。 只不过不是弹窗广告，不是信息流广告，而是藏在首页预置图标里的渠道号。

你每次搜一个词、点一个热搜、从一个预置图标跳出去，X 浏览器就赚一笔。

但是——这些广告位，用户全都可以一键删除。 它把广告位摆在你面前，删不删你自己定——这已经比很多浏览器强多了。

这就是 X 浏览器——拦广告起家，靠广告吃饭，但赚得坦荡。

这是个啥软件？

X浏览器，一款安卓手机浏览器，主打一个"小"。安装包就 2 M 出头，连大厂浏览器一个零头都不到，却把浏览器该干的活全干了——上网、看视频、下载、看书签，一样不落。

名字里的"X"来自英文单词 eXtreme（极致），开发者崇尚极简主义，产品也照着这个思路做。它还有第二层意思：X 是数学方程里的未知数，代表变化和迭代，意思是这软件是跟着用户反馈一点点长起来的。

三个特征，各说一句：

极速。 2 M 大小，极简设计交互，只占极少资源，启动基本秒开，没有新闻流、没有推送、没有后台进程，省电省流量。

干净。 注重隐私，只申请极少权限。官方说法是"无新闻，无推送，无后台进程"，把商业浏览器那些硬广告的套路全砍了。

能扩展。 内建油猴脚本支持，不用装第三方脚本管理器，直接装脚本就能给浏览器加功能。还支持集成第三方下载器、播放器，按需扩展。

说穿了，这就是个"回归浏览器本质"的小工具——功能做到最少，把能折腾的空间留给用户。

开发者是谁？目前团队规模多大？

开发者叫程凯，一个北京程序员，现在是一个人全职维护 X浏览器，不折不扣的独立开发者。官网 about 页原话："我是一名独立开发者，目前在全职维护X浏览器。"公司主体是北京心往科技有限公司，天眼查显示成立于 2015-06-30，注册资本 102 万，法定代表人程坤，说白了就是自己给自己挂的一家公司壳，方便上应用商店、走经营手续。

程凯这人的履历，放程序员堆里算"根正苗红"：

2010 年前后： 自己闷头做了个塞班（Symbian）手机浏览器，做出来发现 UC 已经把市场占了，没干成。
傲游浏览器 5 年： 被傲游创始人陈明杰挖去当手机版技术负责人，一干就是 5 年。后来因为产品理念跟空降的产品总监闹掰，走人。
竞业协议： 离职被竞业协议封印，几年内不能碰浏览器开发。中间去一家医美 O2O 创业公司当 CTO，干了一个月就后悔，觉得"为了钱搭进去几年青春不值"，交接完走人。
2015 年： 赶上"大众创业"，跟朋友搞出海浏览器项目，拿到天使投资，租办公室招人，3 个月做出第一版"X"。结果投资人想撤资，他把钱退了、遣散团队，一个人把项目继续做完。
之后： 就一直是一个人，在咖啡馆、图书馆写代码，全职维护 X浏览器至今。

所以团队规模：就程凯一个人，从写代码、维护网站到回用户消息，全是他。

软件解决了什么痛点？

2015年前后，你想在安卓上找个好用的浏览器，打开应用商店，满屏都是 UC、QQ 浏览器这种大厂货，安装包动辄 50M、100M，点进去满屏的新闻推送、直播入口、电商广告，你想搜个东西，得先在广告堆里扒拉出搜索框。根本不是浏览器了，而是大厂塞给你的全家桶。

X浏览器就砍这三刀：

第一刀：砍体积。 1M（现在 2M），极简设计。没用第三方框架 SDK，全是操作系统原生 API 自己敲的，启动秒开，没有推送、没有新闻流、没有后台服务。

第二刀：砍广告。 内置广告拦截，支持 EasyList、AdGuard 等主流过滤规则。还有个杀手锏——手动标识拦截：遇到拦不住的广告，长按一点，下次刷新自动干掉。多家第三方媒体都说是它家先搞出来的，这功能后来成了 X 口碑最好的卖点之一。

第三刀：砍权限。 只申请极少权限，不绑定手机号、不绑第三方社交账号，不需要摄像头、麦克风这些敏感的玩意。隐私协议里说得直白：几乎不用第三方 SDK，用户数据泄露的风险天然就小。

"广告多、体积大、权限乱伸手"——三个痛点，个个戳中大厂浏览器的死穴。

当时市面上有竞品吗？

2015 年 X浏览器出来的时候，手机浏览器市场早就是巨头的天下。

第一梯队是大厂浏览器。 UC 浏览器、QQ 浏览器长期霸榜，UC 在国内基本坐第一把交椅，出海到印度印尼更是干翻了 Chrome。跟它们同台的还有欧朋、猎豹这些。大厂玩的是"三级火箭"——免费浏览器引流，首页塞满新闻、导航、广告位变现，体积越做越大，几十上百 M 都正常。

第二梯队是手机自带浏览器。 华为、小米这些厂商自带浏览器也吃掉一大块份额，后来算在 Chrome 的统计里。

第三梯队是独立开发者的"小个子"。 这里才是 X 真正的同赛道竞品：

Via 浏览器——涂亚锋做的，同样 1~2M 极简浏览器，同样用系统 WebView 套壳，广告拦截、无推送。跟 X 用户画像、体量、功能高度重合，是 X 最直接的对手。
神奇浏览器——也是独立开发者做的极简浏览器，体量更小。
后来还有很多纯净浏览器。

所以严格讲，X 面对的不是跟 UC 抢天下，而是在"大厂全家桶 + 手机自带浏览器"夹缝里，跟 Via、神奇这一小撮极简浏览器争那一批"被广告烦透了、只想要个干净浏览器"的用户。

开发者从哪来的想法？

X浏览器这个点子，不是拍脑袋蹦出来的，是程凯十年攒下的执念。

根子在 2007 年前后。 他刚入行就辞了工作，把自己关在家里一年多，闷头做塞班（Symbian）平台的手机浏览器——那时候诺基亚还是移动端的未来。结果浏览器做出来，发现 UC 已经把市场占满了，自己做得也不比别人好，这事就黄了。但"做个好浏览器"这颗种子，是那时候种下的。

执念在傲游 5 年里发酵。 他当傲游手机版技术负责人那几年，好几次想法被产品总监否决，"一些我觉得有些很好的、口碑也不错的功能和想法，完全被抛弃掉了"。他后来回忆："我总觉得，在傲游阶段我并没把我想做的事情做出来。"人最拧巴的就是这个——没做，就一直惦记。

被竞业协议摁住的几年，他以为这辈子告别浏览器了。 结果命运开了个玩笑：他去医美创业公司当 CTO 后，傲游那边以为他彻底转行了，主动解除了竞业协议。枷锁一松，2015 年又赶上"大众创业"的浪，他立刻跟朋友组了局开干。

设计哲学也早就有迹可循。 他见过一款叫"悦动"的浏览器，一个深圳小伙做的，只有 0.28M，这体积让他"一下子充满魅力"。他笃信"小、简单"有天然吸引力，自己公开自称"37signals 文化的深度中毒者"，崇尚 KISS 原则（Keep It Simple and Stupid，大道至简）。

有意思的是，他给产品起名"X"——来自 eXtreme（极致），而老东家傲游的英文名 Maxthon 里 Max 也是"最大、极致"的意思。可能是巧合，但那股"追求极致"的精神，确实从傲游一路带到了 X浏览器。

第一版咋做的？(MVP原型)

2015 年中，程凯跟朋友组局，拉了一笔天使投资，在海淀租了间办公室招人，3 个月做出第一版。那个版本什么水准？

技术路线：WebView浏览器。 X 浏览器用的是手机系统自带的 WebView 组件（原生 Android 内核），在外面套一层 UI 和功能壳。程凯自己说的，"X浏览器非常小，安装包只有 1M 左右"。因为不走 Chromium 内置、不用第三方代码库，体积才压得下来。

从我们能下到的最早版本（v3.8.1，1.07MB），验证了几个关键点：

纯原生 Java，没用 Flutter、React Native 任何跨平台框架。
0 个后台 Service——没有驻留服务，没有后台进程，跟当年的宣传"无后台、无推送"对得上。
min SDK 14（Android 4.0），兼容性拉到最底，几乎覆盖当时所有安卓设备。
几乎没有第三方 SDK。访谈里说"随便引入一个第三方的代码库安装包就得好几兆"，所以程凯选择自己写每一行代码，结果也验证了这一点。

权限只申请必要的那几个：INTERNET、网络状态、存储写入。官网说的是"只申请 6 个权限"——虽然版本迭代后逐渐多了 LOCATION 之类的选项，但设计哲学从一开始就定了。

第一版的功能就是极简浏览： 输入网址 → 打开网页 → 广告拦截（根据规则过滤）→ 没了。没有新闻流、没有推送、没有账号系统。投资人花钱买了些应用市场推广位，但产品留存率不够理想，后来投资人撤了资。程凯退钱、遣散团队，自己一个人继续迭代。

后来怎么迭代的？

2015 年投资人撤资后，程凯一个人啃了这台浏览器近 10 年。从 v3.7.8（2022 年 6 月）到 v5.6.3（2026 年 8 月），官网上能查到的版本日志就有 40 多次更新。按功能线，迭代分几个阶段：

阶段一：夯基（2015-2018） 最初的 MVP 就一个 WebView 套壳 + 极简浏览界面。程凯后来陆续加上了广告拦截（默认关闭，用户手动开启）、资源嗅探（下载网站上的图片/视频/音频）。这两个功能奠定了 X 最核心的口碑——"小而强大"。

阶段二：差异化（2018-2020） 内建油猴脚本支持——这是 X 跟其他小浏览器拉开差距的关键。别的浏览器想用油猴脚本得装第三方脚本管理器，X 直接内建，打开用户脚本市场就能装。兼容 GreaseMonkey、Tampermonkey。这功能一上，技术用户群体就稳了。

阶段三：被起诉后的调整（2020-2022） 2020 年被优酷起诉广告拦截，一审败诉判赔 22 万。程凯的反应很实在：把广告拦截做成可选模块，不再默认引用第三方规则源，自己维护去广告规则。这之后版本里多了"强力拦截模式"、"网站独立设置白名单"这些更精细的控制。

阶段四：4.0 大版本翻新（2023） 2023 年 3 月发布 v4.0，界面交互大改。后面连续迭代 4.0.1 ~ 4.2.x，加了阅读模式（自动抽取正文）、自动拼页、手势控制、云同步（书签/设置/密码云端备份）、网页翻译、隐私模式。功能越加越多，但安装包始终压到 2M 左右。

阶段五：5.0 及以后（2024-2026） 2024 年 10 月跳到 v5.0，把第三方工具集成做深了：支持绑定 1DM、ADM、迅雷等专业下载器；支持绑定第三方视频播放器；m3u8 视频下载；密码管理接入系统指纹认证。最近的 v5.6.3（2026-08）把广告拦截性能翻了倍，导入海量规则也不影响渲染。

程凯一个人，靠用户反馈驱动，把一个 1M 的小壳子磨了十年，体积没怎么涨，功能却越装越满。

目标用户是谁？

X 浏览器给谁用？往大了说是被大厂浏览器恶心过的人。

第一类：被广告逼来的普通人。 大厂浏览器首页塞满新闻推送、直播入口、电商广告，搜个东西得先翻广告堆。X 的"帮你干掉 80% 恶意广告"成了最直接的引流钩子，很多人就是冲着"不弹广告"来的。

第二类：老手机/低配置用户。 X 的安装包只有 1~2M，min SDK 14（支持 Android 4.0 以上），老手机、低端机跑起来毫不费力。那些被 UC、QQ 浏览器卡到怀疑人生的用户，换到 X 体感就是"起飞"。

第三类：隐私敏感用户。 只申请极少的权限、无后台进程、无消息推送、不绑定手机号——这些点对"被大厂偷怕了"的用户有致命吸引力。

第四类：技术玩家/极客。 这是 X 最核心的差异化用户群。内建油猴脚本支持，不用装第三方脚本管理器，直接装 Greasy Fork 上的脚本就能给浏览器加功能。还有手势操作、自定义规则、UA 切换、开发者工具……这些功能让技术型用户把 X 当"浏览器版的瑞士军刀"来用。

第五类：资源下载党。 视频嗅探、m3u8 下载、绑定 IDM/ADM 下载器——想下载视频网站资源的人，X 在这个领域几乎没有对手。

从程凯的访谈能看出来，他维护的 QQ 群里活跃的就是这两类人：技术玩家提功能需求、普通用户反馈 bug。两类人相互支撑，构成了 X 的用户生态。

怎么拿到第一批用户？

X 浏览器的冷启动，基本就是"先拿钱砸，砸不动就自己慢慢养"。

第一步：投钱买量，结果翻车。 2015 年有投资的时候，投资方砸了 20 万上线推广费让产品上 Google Play 和应用商店。程凯回忆"幻想过一炮而红，但怎么可能呢？"——产品留存率不高，推广效果很差，这也是后来投资人和他闹掰的直接原因。

第二步：撤资后，靠社区起家。 投资人撤资后，程凯一个人干，没了推广预算，增长变成"每个月几乎看不出变化"的缓坡。但他找到了两个低成本获客渠道：

酷安社区。 官网感谢页自己写了："酷安——一个移动应用及数码产品玩家的社区，X在这里收集了很多用户建议和反馈。"酷安是国内数码玩家、极客的大本营，这波用户恰好又是 X 最想要的核心用户——懂技术、爱折腾、愿意尝鲜。
QQ 群。 官方 QQ 群，早期用户比如大萌主（QQ群管理员）、多巴胺（最早用户+开发者）在这个群里提建议、测 bug、帮回答，形成了最早的"客服靠群友"模式。

早期种子用户特征： 从官网"感谢有你们"页面能看得很清楚——第一批忠实用户里有开发者（多巴胺、阿布）、酷安网友（耗子）、技术玩家，甚至还有后来的创业者（Mood 后来创立了星愿浏览器）。这些人本身就在技术圈有影响力，他们的口碑传播帮 X 在极客圈子里站住了脚。

怎么做用户增长的？

X 浏览器的用户增长，最大特点就是：没正经做过增长。

不投广告。 2015 年投资人砸了 20 万推广费失败之后，程凯再也没买过广告。晚点22年的采访文章写得明白："（X浏览器）用户增长像一条'每个月几乎看不出变化'的缓坡。"日活从零爬到 30~40 万，花了整整 7 年。

不砸自媒体，社交平台也全都没正经运营。 程凯在各平台都有账号，但没有一个是认真运营的：

知乎（程凯，认证"独立开发者-X浏览器的作者"）：关注者 3240，回答 35 个、文章 1 篇、想法 3 条，从 2011 年活跃到现在，总共拿了 11840 次赞同。他的答法就是"谁问浏览器相关的问题，我去顺手答一句"——《X浏览器安全吗？》《有没有没有任何新闻热点的浏览器啊》《如何看待X浏览器因广告拦截被优酷起诉》。这种号谈不上运营，就是产品作者个人 IP 的顺手维护，一年也就更几次。
GitHub（examplecode）：282 关注，偶尔发布开源项目，影响力不大。
酷安：X浏览器上架时间 2015-05-06，累计 536 万下载，评分 9.4，关注 2 万人——但这也是靠产品本身，不是运营出来的。
官方 QQ 群：早期唯一的"客服 + 反馈"渠道。

那增长靠什么？

应用商店自然流量。 应用宝、华为、小米、OPPO、vivo 等国内主流商店全覆盖，这是最大的免费流量来源。酷安是国内极客和数码玩家的集中地，6 万下载。
口碑裂变。 "1M 体积 + 暴力广告拦截 + 油猴脚本"这个组合太有传播力了——任何一个被大厂浏览器恶心的用户，用了 X 之后大概率会跟身边的人安利。至少在酷安和知乎，"x浏览器/via浏览器哪个好用"这类帖子长期是浏览器话题里的热门。
油猴脚本生态反哺。 X 浏览器内建油猴脚本支持，这让它在 Greasy Fork 等脚本站被反复提及。脚本用户天然是技术型传播者，一个人安利就能带动一圈人。
Via 浏览器的「竞品红利」。 国内小而美浏览器这个圈子里，X 和 Via 几乎被捆绑讨论。用户搜索"Via浏览器"时大概率会看到 X，搜索"X浏览器"时也会看到 Via。两者互相带流量。

程凯没做增长，增长是靠产品本身的差异化 + 用户在酷安和知乎自发讨论，自己长出来的。

靠什么赚钱？

X 浏览器怎么赚钱，得实际看下App才知道。它没接广告 SDK，但赚钱的路子藏得挺巧：

① 首页预置的"快速访问"图标，就是广告位。 新装完 X 浏览器，首页快速访问区域自动给你摆好了几个图标：哔哩哔哩、腾讯视频、新浪微博、手机搜狐、秘塔AI。这些链接都带着 X 浏览器的渠道号（比如 B 站是 utm_source=xbrowser，秘塔是 s=xllx）。用户点出去，X 就有分成。但用户可以删——长按点叉就行，这一点是它敢说"干净"的底气。

② 搜索框里搜索，默认走的是百度渠道号。 默认搜索是百度，但链接里挂着 X 的渠道号 from=1013843g。你搜一个词，就是一次带渠道号的跳转，百度按搜索量跟 X 结算。搜索联想词和热搜词也是百度提供的，同样带渠道号，每次展示/点击都有钱。

③ 搜索引擎不光百度一家。 代码里还维护了 360 的渠道号 xbrowser_1、搜狗 sogou-mobb-...、神马 wm725516、Google 的 c=xbrowser。用户切换搜索引擎，一样带渠道号。但只有国内版带百度/搜狗这些国内渠道，谷歌版不带——一来这些是国内搜索引擎，海外用户用不上；二来 Google Play 有一条"Webviews and Affiliate Spam"政策，不允许"主要目的是导流拿佣金的 App"，程凯在谷歌版刻意规避了这条风险（不过谷歌版仍保留了 Google 搜索的渠道参数）。

④ 用户捐助。 官网放了两张码（微信 + 支付宝/PayPal），金额不限。量级不大，更多是精神支持。

浅黑科技专访写过——"下载量飙升时，不断有人把钱送到程凯面前，只为植入广告。程凯一个个把他们送走。"他不做弹窗、不做推送，但预置可删的图标和搜索渠道分成这两块，他接受了。

一年能赚多少？

X 浏览器没公开过收入，这一段只能推算。不过有个数字是确定的——前文说过，它日活 30~40 万（接受采访时说的）。用这个现成的日活直接算，比拿下载量去猜留存率靠谱多了。

公式就一条：

年收入 ≈ 日活 DAU × 每DAU日收入 × 365天

（每 DAU 日收入 = 该用户每天用搜索/点导航产生的分成，行业区间大约 0.005-0.01 元）

先找一个锚点，免得瞎猜。 程凯 2022 年自述："收入比 7 年前上班时多两倍。"7 年前（2015 年）他是傲游手机版技术负责人，年薪约三四十万，两倍就是 60-80 万/年。这个数，就是最可信的中间值。

拿日活反推每 DAU 日收入：70 万 ÷ 365 ÷ 30 万 ≈ 0.006 元。正好落在行业区间里，说明"日活 30-40 万 + 年入六七十万"这个组合是自洽的。

按三档算（每 DAU 日收入波动 + 日活波动）：

悲观（≈40 万/年）： DAU 20 万 × 0.005 元 × 365 ≈ 36 万；加导航位 ≈ 5 万。合计 ≈40 万。够活着，但全职紧巴巴。
最可能（≈70 万/年）： DAU 30 万 × 0.006 元 × 365 ≈ 66 万；加导航位 ≈ 8 万。合计 ≈70 万。跟程凯自述"比上班时多两倍"（60-80 万）严丝合缝对上。
乐观（150 万以上/年）： DAU 40 万 × 0.01 元 × 365 ≈ 146 万；加海外搜索分成 + 导航位 ≈ 20 万。合计 150 万以上。

为什么三档跨度这么大？

日活本身会波动（20-40 万），每 DAU 日收入也有差异（0.005-0.01 元），两个变量一叠加，收入差出三四倍很正常。重点是看量级：最可能 70 万，跟程凯自己说的"比上班多两倍"对得上，这个数可信。

不过有一笔实打实的罚款： 2021 年优酷官司一审败诉，判赔 22 万，这是掏出去的真金白银。

近期用户反馈咋样？

先看硬指标：酷安上 X 浏览器评分 9.4 分（1.4 万人打分），下载 536 万，关注 2 万，这个分数在一众浏览器里算很高的。Google Play 上 4.42 分（9.8 万人评）。总体口碑是立得住的。

但看近期的具体评论（酷安 2026 年 8 月），风向并不全是夸，几类声音很明显：

夸的，集中在"老用户的长情"。 "功能多，界面简洁速度快，自由度高，还自带油猴，用了很多年了，很推荐。""用了这浏览器可能也四五年了。""小巧实用。""顶级好用，顶级良心。"——点赞高的多是用了好几年、把 X 当主力浏览器的老用户。

骂的，集中在"新版改坏了"。 2026 年 8 月一堆人吐槽同一个点："开启无痕为什么要上黑下黑状态？丑的要死""为什么要变上黑下黑啊，是真的丑，别的浏览器这样那你就要学成这样？"——v5.6.x 把无痕/隐身模式改成全黑配色，老用户普遍不买账。最狠的一条直接说"用了很多年到今天彻底失望越来越丑越来越难用，已经转 via 了"。

报 bug 的，视频相关最多。 "点击视频全屏就会自动刷新网页""竖屏切横屏就自动刷新""视频嗅探不出来""视频播放时右下角这三个怎么关闭"——资源嗅探和视频播放这块，是近期 bug 高发区。

还有一类：竞品在评论区截流。 有人直接发"推荐一款比X还好用的浏览器……广告拦截比X强一些……下载地址"，还在评论区跟了几十条。说明 X 的用户盘子大，别的极简浏览器正盯着它的评论区抢人。

x的核心功能（快、干净、油猴、自由）口碑稳如老狗，但新版改 UI 改出众怒、视频相关 bug 反复，如果程凯不及时收住，流失率会往上走——评论区那个"已转 Via"就是信号。

从中我们能学到什么？

程凯这个案例，方法论价值不在这产品多赚钱，在于一个人怎么把一个"小"产品活十年。拆几条能学的：

第一，把"极简"当成产品哲学，不是营销话术。 别人说极简是砍功能，程凯的极简是"每加一个功能，安装包都得压住"。1M 的浏览器，几乎不用第三方库，自己写每一行代码，一个扫码功能要 700KB 他都舍不得常驻、做成插件按需下载。

当所有人都往大里做，往小里做就是差异化。极简不是偷懒，是比做加法难一百倍的减法。

第二，一个人也能干成，但要有"把自己当公司"的自律。 程凯一个人包了研发、前后端、测试、客服、运营，但关键是他给自己立规矩：每天固定时间出门上班，每天工作 6 小时，在咖啡馆/图书馆办公。

独立开发者最大的敌人不是竞争，是"在家办公=随时躺平"。他靠"仪式感"把自己从散漫里拽出来，7 年保持每周或每两周一次更新。

第三，靠社区活，而不是靠流量活。 X 浏览器没有砸钱买量（20 万推广费翻车后再没碰过），增长全靠酷安 + QQ 群 + 知乎这几个人群扎堆的地方。核心用户是极客，极客自带传播力。找到你的"酷安"，深耕那一个圈子，比撒网买量值钱一百倍。

第四，变现要克制，但要清醒。 程凯拒绝弹窗、拒绝植入广告，但接受了"首页预置可删的快速访问图标 + 搜索渠道分成"。他不是不懂钱，是懂"哪些钱会杀死产品，哪些钱不影响体验"。

独立开发者最容易栽的坑是"什么钱都想赚"，最后把用户全赶跑。X 的分寸感：广告可以有，但必须可删。

第五，被打击了要会"软着陆"。 2020 年被优酷起诉索赔 105 万，一审败诉赔 22 万——换别人可能直接心态崩了。程凯的做法是：不硬刚，把广告拦截改成可选模块、不默认引用第三方规则、自己维护去广告规则，既保住了功能又规避法律风险。

遇到外部暴击，先活下来，再调整姿势。

第六，长期主义的复利。 X 浏览器做了 10 年，下载量过亿、日活几十万——这不是一夜爆红，是一年一年磨出来的。

程凯不追风口、不做超级应用，就守着 1M 的小浏览器过日子。在独立开发这个行当，熬得住比跑得快值钱。

今天再做有没有机会？

先看现状：这个赛道没凉，但越来越挤。

直接竞品（国内极简浏览器赛道）：

Via 浏览器——头号对手，国内下载 1.59 亿，比 X 还大。同样极简、同样 WebView 套壳、同样搜索分成变现，跟 X 正面硬刚。
还有各种"纯净版"小浏览器在边缘卡位。

开源替代品（GitHub 上能免费拿代码的）：

EinkBro——基于 Android 原生 WebView 的开源轻量浏览器，GitHub 2005 star，最新版 5.0 MB（v16.5.0，2026-08-30 刚发布），F-Droid 可下载。专为 E-Ink 电子墨水屏优化，但也兼容普通手机。是目前唯一一个还在持续发新版的 WebView 开源浏览器。
Lightning Browser——老牌开源 WebView 轻量浏览器，2431 star（同类里最高），安装包 3.3 MB。最近 GitHub 还有代码提交（2026-08），但正式版本停在 2019 年的 v5.1.0，属于"代码活着、产品停摆"。
其他项目要么凉了（Kiwi 2025 年归档、UltimateBrowserProject 2017 年停），要么 star 太少（Solipsism 153、candy-browser 125、uweb 109）。

这说明一个尴尬的事实：WebView 极简浏览器这个赛道，开源社区做了这么多年，连一个"还在认真维护、通用场景的活跃产品"都找不出来。抄代码容易，把产品做久极难——X/Via 真正的护城河不是技术，是时间。

今天做的话：

优势：技术门槛低（WebView 套壳 + 开源代码现成）、需求长期存在（大厂浏览器永远会做广告，就永远有人想逃）、独立开发者不需要烧钱、赛道能养活人。
劣势：用户心智已被 Via/X 占满，极客圈子就那么大；搜索分成单价在走低；大厂（手机自带浏览器）也在吸流量；纯"极简"卖点已经不够了——得找到新的差异化。

所以机会还有，但不是"再做一个极简浏览器"的机会，而是"在极简基础上叠新能力"的机会——alook靠倍速播放杀出、focus靠完整插件生态，未来可能靠 AI。

纯抄 X/Via 没戏，差异化才有活路。

彪哥的总结

拆完 X 浏览器，用乔布斯的一句原话收尾——"人们以为专注就是对你要做的事说'是'，其实专注是要对另外一百个好主意说'不'。"

这句话拿来形容程凯再贴切不过。

程凯正是这么活的——他人生充满"拒绝"：投资人骂他"根本不像个CEO、没有大的野心"，他退钱、遣散团队、拒绝当CEO；下载量飙升时有人排队送钱植入广告，他一个个送走；做超级应用能赚更多，他坚持只做 1M 的浏览器；AI 风口来了他也没追。

但正是这一连串"说不"，让一个 1M 的小浏览器，下载过亿、日活几十万、不融资不烧钱，安稳活到现在。

独立开发者的护城河从来不是技术——WebView 套壳开源代码一抓一大把。程凯的护城河是：他太清楚自己对什么说不了。 不做广告弹窗、不做新闻流、不碰大厂命根子、不追风口，只守着"极简 + 可删广告位 + 搜索渠道分成"这条最克制的路，一磨就是十年。

专注不是什么都做，而是敢对诱惑说不。做减法的人，往往笑到最后。

我是爬虫彪哥，如果这篇拆解对你有启发，点个赞或者转发给身边的兄弟们。

要是觉得哪块没拆透，或者有啥补充，欢迎评论区接着聊。

下期预告：咱们拆夸克……敬请期待。

投稿或合作，请联系邮箱：chai@gokuscraper.com

Need to Turn Images Into Vector Graphics Without Uploading Them to Random Websites? This Open-Source Tool Might Be the Easiest Option for Most People

GokuScraper悟空爬虫 — Mon, 24 Aug 2026 19:07:08 +0000

Need to Turn Images Into Vector Graphics Without Uploading Them to Random Websites? This Open-Source Tool Might Be the Easiest Option for Most People

Have you ever had this problem? You are making a PowerPoint slide, a public account cover, or a webpage icon, and the source is just one PNG image. It looks fine on your phone, but once you enlarge it, it turns into a blurry mess of pixels. People who know what they are doing will tell you: convert it to SVG. Then it stays sharp no matter how much you scale it up.

The problem is, though, that Illustrator costs money, Inkscape requires installing software and learning a new interface, and most of the free online vectorization sites either add watermarks, force you to sign up, or make you worry about privacy because you have to upload your image.

At that point, a lot of people simply give up and accept the blur.

But don’t panic. There is actually a much easier route here, and it is surprisingly simple.

This tool solves the real problem

To be honest, turning a bitmap into a vector is not really a technical challenge. The real challenge is the entry barrier.

That is exactly what the Goku Image to SVG Tool is trying to remove. It wraps an open-source Vecto vectorization engine in a Streamlit web interface. You do not need to write a single line of code or install a professional design app. Just open the page, upload the image, and it automatically converts it into a clean SVG.

It supports PNG, JPG, BMP, GIF, TGA, TIFF, WebP and several other common formats. Upload the file and it converts right away. On the right panel, you can adjust the palette, style, and detail sliders, and the vector output updates immediately. You can compare the original, vectorized, and segmented views by dragging a slider to make sure the result looks right before exporting. The SVG can then be dropped directly into Illustrator or Figma for further refinement, and you can also export a PNG at any scale if needed. The interface is bilingual and supports English and Chinese with a one-click switch.

The entire project is open source and free under the MIT license. It currently has 174 GitHub stars. If you want something inexpensive and low-friction, this is genuinely useful.

How easy is it to use?

The good news is that it is straightforward even for beginners. If you want the easiest route, just use the online version. Open image-to-svg.streamlit.app in your browser and you are ready to go. No installation required.

Step one: upload the image. It supports common formats such as PNG, JPG, BMP, and GIF. You can drag and drop or click to upload. Once it is uploaded, the conversion starts automatically.

Step two: adjust the parameters. In the sidebar, there are sliders for palette, style, and detail. Tweak them and the output updates immediately. There is even a color-block preview, so you do not have to wait around wondering what the result will look like.

Step three: check the result and export it. Drag the slider to compare the original, vector, and segmented views. Once you are satisfied, download the SVG or render a PNG at the size you want.

If you want to run it locally on your own machine, it is still easy enough. Run git clone to pull the project, then pip install -r requirements.txt to install dependencies, and finally start it with streamlit run streamlit_app.py. Open the local address in your browser and you are done. The engine binaries are already bundled in the project repository, so it is basically a one-command setup.

I would not call it perfect, though

I should also be honest: this tool is not flawless.

First, the repository includes the Linux x64 version of the Vecto engine. If you want to run it locally on your own computer, Windows or Mac users may need to manually build or fetch the matching binary for their platform. If you want the simplest experience, the online version is the easiest route.

Second, the vectorization quality depends heavily on the quality of the source image. Tools like Vecto are generally better suited for logos, icons, illustrations, and line art with clear color blocks. If you feed it a noisy photo, the result may look distorted or overly simplified. It is unrealistic to expect a single tool to do perfectly well on every image type.

Third, the project is relatively new. It has 174 stars and is still actively evolving, and the cloud free version has limited resources when processing large images. But the author seems very active and continues updating it, so it is reasonable to expect more improvements over time.

A quick side note

One more thing: the author is active in the open-source community. He has even posted a detailed introduction on the “科技爱好者周刊” (Tech Enthusiasts Weekly) self-promotion thread, and the post includes a lot of useful technical details, such as the fact that all 42 pytest tests are green and that the memory peak during a single conversion is roughly between 60MB and 490MB. That makes the project look serious and well-maintained.

There is also a thread on NodeLoc where people discussed it. The comments are pretty real: some people said it is useful for designers, while others felt the use cases were somewhat narrow. That is actually a fair assessment. A tool like this is only truly useful if it matches your workflow. There are also plenty of other open-source alternatives, such as vtracer and potrace-based image-to-SVG converters. More options are usually a good thing, especially when they are free. You can try a few and see which one suits you best.

Bottom line

In short, if you work with images often and want to turn bitmaps into vector graphics without losing quality, without paying for expensive software, and without uploading your files to shady sites, this Goku Image to SVG Tool is worth bookmarking.

The online version opens in the browser and works immediately, and the local setup path is also available if you want to run it yourself.

GitHub: https://github.com/goku-open/goku-image-to-svg-tool

Online demo: https://image-to-svg.streamlit.app

I’m Wukong Scraper, and I focus on sharing AI open-source projects. If you made it this far, a like, comment, or repost would mean a lot. It genuinely helps a lot.

Thanks for reading, and I will see you next time.

I Analyzed 45,000 AI Prompts and Realized I Was Using AI Image Generation All Wrong

GokuScraper悟空爬虫 — Wed, 08 Jul 2026 10:19:36 +0000

I Analyzed 45,000 AI Prompts and Realized I Was Using AI Image Generation All Wrong

Here's how it happened.

Last weekend, I was experimenting with AI image generation, trying to create a cover image for my blog. In my mind, I had this perfect surreal miniature scene: a cream-colored swimming pool with tiny sunbathers floating on inflatable rings.

So I spent nearly an hour typing prompts into ChatGPT, but the results were... underwhelming.

Without formal art training, I always felt lost when writing prompts. After struggling for almost an hour, I gave up and grabbed an old stock photo instead.

This frustration is something many AI image enthusiasts can relate to.

You know AI can create amazing images—those stunning examples you see online are everywhere—but when you try it yourself, your prompts never quite hit the mark. That small gap feels like an entire world away.

Later, I complained to a friend who works in AI design. He glanced at my prompt and said something that stuck with me:

"What you wrote is too abstract. AI has no idea what you actually want."

I asked him how I should write it instead.

He didn't answer directly. Instead, he sent me a link.

It was to a website called "Wukong PromptHub."

When I clicked through, it showed 45,049 prompts.

My first thought was, "Another prompt collection site." There are countless sites like this now—most just scrape prompts from Reddit and Twitter without proper organization.

But then I clicked on one example and was immediately surprised.

Take this prompt titled "Cream Pool Miniature Fantasy":

A surreal miniature world collage poster featuring an oversized open blue Nivea-style tin box 
transformed into a fantastical swimming pool filled with glossy white "cream water."
Tiny sunbathers float on soft inflatable rings, lounging on miniature deck chairs,
sliding into the cream-colored pool from a small blue slide.
The background features a soft, warm, slightly textured surface,
using fine marble or matte stone with even lighting.
Maintain realism through soft shadows beneath props and figures.

Notice this isn't just "a pool, miniature style"—it's a complete description with scene composition, lighting logic, and material details. You can tell the person who wrote this prompt already had a complete image in their mind; they simply described it in words for AI to render.

And that's just the text description. Many prompts on this site use structured JSON format.

Here's an even more detailed example called "Matcha Moment Through a Fisheye Lens":

{
  "scene": {
    "environment": "sunny_boardwalk",
    "details": "wooden_planks, colorful stalls, people walking, distant umbrellas",
    "lighting": "bright_midday_sun",
    "sky": "clear_blue"
  },
  "camera": {
    "lens": "ultra_wide_fisheye_12mm",
    "distance": "very_close_up",
    "distortion": "strong_exaggeration",
    "angle": "slightly_low_upward"
  },
  "subject": {
    "type": "young_person",
    "expression": "curious_playful",
    "clothing": {
      "top": "bright_green_knit_sweater",
      "accessories": "chunky_blue_sunglasses"
    }
  }
}

Look at the level of detail: lens type, aperture, distortion, angle, clothing colors, environmental lighting, even emotional atmosphere—all specified.

That's when it hit me.

This is how experts write prompts.

What I'd been writing wasn't really prompts—it was more like wishful thinking.

I started browsing through the site's prompts and became completely absorbed.

The categorization is particularly thoughtful. Instead of generic categories like "portraits," "landscapes," or "architecture," it's organized by use cases: entertainment and fun, business productivity, content creation, prompts from X (Twitter), and prompts from Douyin (Chinese TikTok).

Each category contains prompts actually used by working creators, not just tutorial-level examples.

For instance, under "Content Creation," there's a prompt called "Ultra-Wide Mobile Photo Editing Secrets" specifically designed to transform ordinary photos into ultra-wide or fisheye effects. It even specifies how to hold your phone, screen replacement rules, and which body parts should be closer to the lens—like a detailed product manual.

Under "Business Productivity," you'll find prompts like "Enterprise Profit-Loss Sankey Diagram," "Smart Product Retouching Service," and "Intelligent Background Person Removal"—all designed for real commercial applications.

There's even one called "Surveillance Footage Face Tracking Effect" that I've seen used to create cyberpunk-style surveillance videos. I never imagined the actual prompt would be available here, ready to copy and use.

The site also covers a comprehensive range of AI models. It's not just limited to Nano Banana Pro—it includes GPT Image 2, SeedDance 2, Open Models, and Text Prompt. For each model, the prompts are adapted to account for their specific syntax differences.

Honestly, I was already impressed at this point. But as I kept exploring, I discovered something even more remarkable.

What fascinated me most was a collection of prompts that clearly weren't written casually—they demonstrated deep understanding of visual language.

Take the prompt "Iron Smoothing Wrinkles":

An award-winning hyper-realistic macro photograph.
Extreme close-up of an elderly woman's eye and cheekbone.
A miniature toy-like white-blue iron is placed on her skin,
actively pressing and smoothing deep wrinkles and crow's feet.
Lighting uses high-contrast hard flash.

Notice the specific artistic style reference (Maurizio Cattelan style), exact camera and lens specifications (Hasselblad H6D-100c + Macro 120mm f/4), and even film simulation details (Kodak Ektar 100).

This level of precision would be sufficient for commercial projects. Honestly, some client briefs I've received for commercial work weren't this detailed!

Then there's "Giant Machine, Tiny Object Fantastic Vision"—the entire prompt reads like a complete creative brief, covering scene architecture, lighting design, and surreal interaction mechanics. I believe if you sent this prompt to a human photographer, they could actually shoot a compelling photograph based on these instructions.

This goes beyond just being a "prompt."

It's a fully executable creative specification.

While browsing, I remembered something important.

Recently, a set of AI-generated triptych images went viral on social media. You've probably seen them—people shared them widely, amazed by AI's creativity. But few realized that behind these so-called "AI creations" were humans carefully designing prompts, treating AI as a paintbrush to be directed.

AI is like an exceptional paintbrush that doesn't listen to instructions. Tell it to draw a beautiful person, and it might give you a six-fingered金刚 (six-fingered strongman). But tell it "85mm f/1.4, golden hour, Kodak Portra 400, urban street café window, warm nostalgic lighting on the side of the face, fine film grain, eyes slightly looking away from the camera, steam from coffee cup dissipating in the air," and you get infinitely closer to that imagined photograph.

This website essentially systematizes the language for directing AI.

With 45,000 prompts, it covers 45,000 different creative scenarios. You don't need to figure out from scratch how to describe a miniature pool, how to capture a fisheye matcha photo, or how to generate an enterprise Sankey diagram. Someone has already figured it out thoroughly—you can just take it and use it directly.

This reminded me of a historical parallel.

It's similar to when designers migrated from Photoshop to Sketch about ten years ago. When Sketch first gained popularity, everyone asked what made it better than Photoshop. The real value wasn't in the software itself, but in its ecosystem of plugins and component libraries. Photoshop could handle UI design, but with component libraries, the efficiency was on a completely different level.

Today, Wukong PromptHub serves the same function as those plugin ecosystems and component libraries. AI image generation tools themselves are already powerful enough—whether Midjourney, DALL·E, GPT Image 2, or SeedDance, everyone is competing on model capabilities. But what truly creates a gap in creative efficiency is your ability to quickly write effective prompts.

And this website eliminates the part you'd normally need to figure out yourself.

Building on this insight, let's talk about timing.

Honestly, I find this trend fascinating.

When ChatGPT first launched in late 2022, people's biggest anxiety was "What if I can't write good prompts?" Prompt engineering became a recognized discipline, with countless tutorials teaching various frameworks for writing effective prompts.

But here in 2026, we're seeing a new trend: most people aren't struggling with how to write prompts—they simply don't have time to write them.

You want to create an image of a miniature pool, but you don't want to spend an hour researching ultra-wide focal lengths and depth of field relationships. You just want a usable image quickly. You're not a photographer or designer—you're a content creator, entrepreneur, or regular professional who needs AI to help produce work, but you don't have the bandwidth to become a prompt expert.

What's the best solution?

It's not spending a week learning prompt writing—it's copying a proven, validated prompt someone else has already created, making a few adjustments, and generating your image.

That's exactly what this website does.

It presents 45,000 already-validated creative directions right in front of you. You don't need to start from scratch—you just need to find the scenario you want, copy it, and paste it.

It's that simple.

Now for the practical details.

The website is called "Wukong PromptHub," and you can find it at prompthub.gokuscraper.com.

It's completely free. You don't need to register to browse and copy prompts.

After logging in, you can bookmark your favorite prompts. The site also features daily updates and a "shuffle" function that continuously adds new content. As of today, it supports Nano Banana Pro, GPT Image 2, SeedDance 2, other Open Models, and Text Prompt—covering virtually all mainstream AI image generation tools.

If you're experimenting with AI image generation, I highly recommend checking it out. Don't approach it with the mindset of "I need to learn prompt writing"—just treat it as an inspiration gallery. When you find a visual direction you like, copy it, tweak a few details, and let AI generate it for you.

It's actually quite satisfying.

To be honest, I haven't finished browsing all 45,000 prompts yet—it's simply too much content. But I've already bookmarked several prompts I plan to use for future projects, especially those quirky styles like "Giant Cat Breaking Through Street Screens," "Office Fun Group Photos," and "Chalkboard Empress Drawings." I'm already thinking about how I can incorporate these into my articles.

Wow.

From now on, I won't need to spend an hour struggling with cover images anymore.

先上车后买票，一个爬虫工程师不敢写进文章里的行业潜规则

GokuScraper悟空爬虫 — Wed, 08 Jul 2026 08:26:36 +0000

先上车后买票，一个爬虫工程师不敢写进文章里的行业潜规则

前阵子闲得没事，去翻了翻腾讯的 SkillHub。

不翻不知道，一翻还挺意外。里面挂着大量卖数据的技能工具。什么小红书爆款笔记查询、公众号热门文章搜索、抖音热榜数据采集，什么都有。接口调的都是第三方 API，但你稍微懂点行就知道，公众号没有公开过这种搜索接口，小红书没有给过数据授权，抖音也没有。

那这些数据是从哪来的？

只有一个解释，第三方爬虫爬的，接口逆向破解出来的。

而且你看这些技能的开发者信息，大部分连个微信都不敢留，只敢留个匿名邮箱。一看就是独立开发者在干这事，把别人平台的接口搞出来，打包成技能放到 SkillHub 上卖钱。

技术上虽然是中立的。

但法律上也是灰色的。

但你猜怎么着。SkillHub 就这么让他们上架了。不只上架了，评分还不错，排名也挺靠前。

腾讯当然知道这些东西是什么性质。他们太知道了。

但为什么还让上架？

因为 SkillHub 需要内容。一个刚上线的平台，你需要让用户觉得有用。什么最有用？数据分析类的。公众号写手想看爆款文章，小红书博主想看竞品数据，这些都是刚需。官方不给接口？那就让第三方来补上。

先把技能摆上，把平台撑起来，合规的问题以后再说。

先上车，后买票。

这条法则不止在中国的互联网圈子常见，全世界也是通用的。你先把东西拿过来用，用出问题了再说。如果没出问题，这张票就不用买了。如果出了事，再补票也来得及。

腾讯自己就是这么干的。今年3月他们刚因为爬了 OpenClaw 的 ClawHub 被原作者挂墙头，闹了四天最后掏钱赞助才收场。结果风波刚过，自己的平台上就堂而皇之地摆着这些卖数据接口的技能。

同一套逻辑。

但这里有一个问题你可能很关心，这些东西合不合法？

说实话，非常模糊。

爬虫这件事在国内法律上本来就是一个灰色地带。公开数据算谁的？你爬了别人网站上的公开信息，算不算侵权？这些问题在中国标准很模糊，至今没有特别明确的法律答案，很难单纯从技术角度没法界定，更多看的是社会影响。但按照美国的判例，有一条比较清晰，如果你爬的是需要登录才能看的数据，那你就越界了。

SkillHub 上这些技能属于哪种情况呢？一部分爬的是公开内容，但更多的是需要登录的内容。小红书很多笔记要登录才能看完整内容，公众号文章也有访问限制。这意味着它们在法律上已经越过了那条线。

那为什么没人管？

不是因为合法，是因为经济账算不过来。

你想想，小红书发现有人在卖查自己数据的工具，要不要告？理论上可以告。但你算一下账，律师费、诉讼费、时间成本，加起来可能几十万。而一个独立开发者卖技能赚的那点钱，你就算告赢了也拿不回多少。成本比收益还高，谁愿意打这个官司？

所以只要你别搞太大，没人理你。

但你要是做大了呢？

用户量上来了，收入变高了，名气变大了，那你就变成了靶子。到那时候平台一封律师函过来，你扛不住的。你不是腾讯，你没有法务团队，你没有几百万去打官司。人家告你，你连应诉的成本都付不起。

而且还有一个更致命的问题，就算你侥幸赢了官司，你也赢了时间。打官司拖你一年半载，你的业务早就黄了。

所以这些灰色工具有一个共同的特点，永远做不大。也不敢做大。

创业圈有句话叫「爬虫做不大」，说的就是这个意思。你可以偷偷赚点钱，但不能光明正大做成一个生意。一旦你从「一个技能」变成了「一家公司」，从「月入几千」变成了「月入几十万」，你就一定会被人盯上。

腾讯可以爬 ClawHub 然后掏钱和解，因为他有六千亿营收。但你一个小公司，你掏不起这个钱。

这就是灰色爬虫的真实处境。

法律不管你不是因为你没错，是因为你太小了，不值得管。你一旦长大，就会有人来收割你。

所以回过头来看 SkillHub 上那些卖数据的技能，没什么高深的。就是一群小公司或者独立开发者在灰色地带里讨生活。他们现在能活着，不是因为他们合法，是因为他们太小了。一旦做大，一定会被拍死。

这就是这个行业最真实的规则。不是法律条文里写的那个规则，是实际运行的规则。

先上车后买票。但在上车之前，得想清楚一件事，这趟车你买到票之后，还下不下得来。

They Claimed Huawei's Autonomous Driving Mileage Was Fake—So I Spent 10 Minutes Checking the Code

GokuScraper悟空爬虫 — Sun, 05 Jul 2026 15:58:05 +0000

They Claimed Huawei's Autonomous Driving Mileage Was Fake—So I Spent 10 Minutes Checking the Code

Someone posted a video on TikTok mocking Huawei, claiming that the safety mileage data displayed on Huawei's official website was fabricated.

Their reasoning? The counter kept increasing even when offline.

They turned on airplane mode, recorded their screen, but the cumulative driving mileage number on the webpage continued to scroll. Many viewers assumed it was purely a frontend script—a deceptive trick. There was an excited tone in their voice, as if they'd caught another company red-handed.

I don't know how many people have seen this video, but there were plenty who shared it, criticized it, and joined in the mockery.

Huawei has become a magnet for attention. Criticizing them negatively always garners likes.

But honestly, I was genuinely curious—or rather, my professional instincts kicked in. Having worked in frontend development, I know the most direct way to verify whether a counter is real or fake is to open Developer Tools and check the Network panel.

So I opened Huawei's webpage.

https://auto.huawei.com/cn/ads/safety-and-data-report

The page displays three large numbers: cumulative assisted driving mileage of Qiankun Intelligent Driving, total cumulative mileage of vehicles equipped with Qiankun Intelligent Driving, and cumulative collisions actively avoided. The numbers are scrolling, giving off that typical "big tech data dashboard" vibe.

I pressed F12 and opened the Network tab.

It didn't take long to spot an API endpoint polling regularly.

auto.huawei.com/external/uiapi/ads/v1/query

Every 3 seconds. The returned data structure is clean, with three fields corresponding exactly to the three counters on the page.

This is a legitimate backend API. Not local simulation, not mock data—it's real aggregated values fetched from Huawei's vehicle cloud platform.

How can I be sure?

I watched the data changes for a while. The mileage growth between consecutive requests wasn't constant. I took several samples at different times, and the growth rate fluctuated by about 26%. A uniform timer simply couldn't produce this kind of fluctuation pattern. Only real vehicle telemetry data aggregation would show varying growth rates at different times—more driving during rush hours, less at night, which makes perfect sense.

But then, how did that "still increasing when offline" phenomenon in the video happen?

I started examining the JavaScript files loaded on the page.

Scanning through the Sources panel, one filename immediately caught my attention.

safety_kilometer.js

I opened it.

Well, well.

Inside lay a set of hardcoded numbers:

Initial value: 7,228,242,208
Baseline time: January 12, 2026
Fixed growth rate: 224.86 per second

It was a pure frontend timer. Take the current time, subtract the baseline time, multiply by the fixed growth rate, add the initial value, and calculate the "current mileage." No connection to the backend whatsoever.

If you only looked at this file, you'd think—case closed. Huawei is indeed running fake data with a timer on their webpage.

But I looked a bit closer.

This JavaScript targets DOM elements with IDs #ias-gongli and #ias-pengz.

On the current webpage, these ID elements simply don't exist.

I searched the entire DOM—nothing.

So how was this file being triggered? I checked the call chain. It executes when the page loads, looks for those two elements, and if it can't find them, it does nothing—no errors, no fallback, nothing at all. It just sits there quietly.

But I noticed one detail. There's a QR code on the page directing users to the "Huawei Qiankun App." My guess is that this file wasn't meant for the official website at all.

Comparing safety_260521.js and safety_kilometer.js reveals they're completely different approaches:

safety_260521.js relies on jQuery + Web Worker + AJAX, fetches data through API polling, has complex scrolling animations, and weighs 28KB after bundling. It's a proper official website solution.

Meanwhile, safety_kilometer.js is pure vanilla JavaScript with zero dependencies, calculates numbers directly in memory, and updates innerHTML—less than 1KB after decoding.

The most reasonable guess: this file was prepared for H5 embedded pages within the Huawei Qiankun App. Embedded app pages need instant loading—they can't wait for jQuery to load and then wait for API responses—so they use pure vanilla JavaScript with a set of hardcoded initial values to run an "offline-capable" counter. Once the network is ready, the app uses its JS Bridge to fetch real data and make corrections.

The official website and app embedded pages share the same CDN static resources. So this file got deployed to the official website's server, but since the website's HTML doesn't contain the corresponding ID elements, it just sits there quietly, doing nothing.

It's not some conspiracy. It's simply two platforms sharing the same resource set, with conditional loading that wasn't thoroughly implemented. This kind of thing is extremely common in frontend projects at legitimate companies.

So the surface narrative splits:

There's a real API driving the front-end numbers. And there's also a hardcoded timer file sitting on the server. But it's simply not meant for this page.

So what exactly was "updating when offline"?

I simulated an offline scenario. I connected my phone to a hotspot, refreshed the page, waited for one successful API call, then turned off the hotspot.

Something interesting happened.

The counter didn't stop immediately. It indeed kept moving.

But looking closely, its behavior was different from when normally connected.

When online, the number scrolls smoothly and continuously, with a tiny jump every 3 seconds. When offline, the number oscillates within a small range, with decreasing amplitude, coming to a complete stop after about 12 seconds.

I went back to examine the code.

Here's the logic:

The main loop sends an AJAX request to that backend API every 3 seconds. If the request succeeds → update the target value → Web Worker runs a 4-second scrolling animation from the current value to the latest value. If the request fails → the error callback triggers.

What happens when you go offline?

At 0 seconds: AJAX succeeds, gets the value, starts a 4-second animation.
At 3 seconds: New AJAX request fails. Error callback triggers, restarting a 4-second animation from the current value (where the previous animation stopped halfway) back to the last successful API value.
At 6 seconds: Another AJAX request fails again. Same process repeats.

Each error callback restarts the animation. But each time it restarts, the current value is closer to the target value (because the previous 3-second animation had already pushed the number forward). So the "new animation" has less distance to cover.

Constantly restarting, constantly decaying.

After about 4 to 5 rounds, the current value is almost equal to the target value, and the visual amplitude of the animation becomes too small for the human eye to distinguish.

12 seconds—convergence.

That's the whole truth.

It's not that "it keeps updating when offline." It's that the AJAX failure error callback keeps retrying不甘心地一遍遍重试, each time making the animation "chase" the target value from its current position. The animation gets closer and closer until it matches perfectly, then stops.

It's residual easing animation—not a ghostly counter.

A more vivid analogy would be:

You push a swing, and it starts swinging. When you go offline, no one pushes anymore, but the swing continues to swing for a few more rounds with decreasing amplitude until it finally stops—and then the person who made the video says, "Look, no one's pushing it but it's still swinging—the swing is fake."

Yes, the swing is indeed still swinging.

But that's entirely different from fabrication.

Going back to that video.

I specifically checked the comments section of that TikTok video later. Not a single person verified anything. Everyone was just piling on to criticize Huawei.

Not one person went to open that webpage to take a look. Not one person opened F12 to check the Network panel. Not one person asked, "When it's offline and the numbers are still moving, is it really updating, or is the animation just not finished?"

Everyone was just criticizing en masse.

Sometimes I think this might be the most thought-provoking aspect of the whole incident.

Was there actually anything wrong with Huawei's design? Honestly, I don't think so.

When requests fail, keeping the last successfully obtained value and continuing with smooth animation until it gradually stops—that's completely normal. If the numbers just froze immediately when a request failed, users would think the page had crashed. Displaying the last value steadily and stopping smoothly is actually reasonable UX design.

The real difference is that normal people seeing "numbers still moving briefly when offline before gradually stopping" wouldn't find anything suspicious. But that video chose to capture only the "still moving" part without showing everyone how it eventually stops.

Why did that video's narrative spread so quickly? Because it's so "easy to understand." Offline, numbers still moving, fabrication. Three steps. Anyone scrolling by can digest the entire story in seconds while gaining a sense of intellectual superiority—that feeling is addictive.

Meanwhile, the truth requires you to open DevTools, check the Network panel, examine the Sources panel, understand the code logic, comprehend WebWorker animations, and grasp easing functions. It takes time. It requires patience.

A "blackmail revelation" that can be digested in seconds versus a "truth" that takes at least ten minutes to understand—the latter stands no chance in terms of spread speed.

I'm not blaming anyone. I'm like this too. When I first saw that video, my immediate reaction was also, "Hmm? That does seem problematic."

But the difference might simply be that I happen to have frontend experience, happen to know how to open Developer Tools, and happen to have ten to twenty minutes to spend investigating this matter.

What if I didn't have these conditions? I might have shared that video too, with a caption like "Huawei loves doing these superficial things."

So what I'm pondering is—are we really "seeking truth," or are we "seeking validation"?

In that video's comment section, no one was "seeking truth." Everyone was "seeking validation." Seeking validation for "See, Huawei really isn't good." Seeking validation for "I've uncovered another scam." This validation comes too easily—it doesn't cost you anything, just requires a like or a share.

But "seeking truth" comes at a cost. It requires your time, your effort, and your willingness to admit you might be wrong.

We live in an era where information spreads far faster than it can be verified. A 15-second TikTok video can reach hundreds of thousands of people in half a day, but I suspect fewer than a hundred people actually go to open that webpage to take a look.

I'm writing this not to defend Huawei. Huawei doesn't need me to defend it.

That safety_kilometer.js is indeed a file unused by the official website—it was prepared for app embedded pages. Two platforms sharing the same CDN resources with incomplete conditional loading isn't exactly a crime.

I just think that if you come across similar "expose-the-truth" videos next time, maybe before sharing, take three seconds to ask yourself:

Is there another possible explanation for this phenomenon he's describing?

That counter isn't fake.

That "updating when offline" phenomenon isn't some conspiracy either. It's just ordinary easing animation design—when requests fail, it retains the last value and stops smoothly. That's all.

I believe the real value of this story isn't about judging whether Huawei is right or wrong.

It's about showing us how vast the gap can be between something that "appears to be fabricated" and something that's "actually fabricated." It's also about showing us that in this age of information explosion, a 15-second video can easily reach hundreds of thousands of people, but those who actually verify it might number fewer than a hundred.

Next time you see this kind of "expose-the-truth" content, maybe ask yourself one more question—could this really be true?

你的时区正在告诉 Claude Code，你是中国人

GokuScraper悟空爬虫 — Sun, 05 Jul 2026 05:17:28 +0000

你的时区正在告诉 Claude Code，你是中国人

故事是这样的。

前两天我在研究 Claude Code 的时候，发现了一件让我后背发凉的事。

大家都知道，Claude Code 现在对中国用户是有一些限制的。很多朋友为了用上它，各种折腾中转代理，把 ANTHROPIC_BASE_URL 指向各种第三方端点，以为自己藏得挺好的。

结果呢。

人家 Claude Code 压根不在乎你从哪个 IP 访问。它用的是另一种方式来判断你是不是中国用户。

读取你的系统时区。

对，你没看错。就是你在电脑右下角设置的那个时区。当你的 ANTHROPIC_BASE_URL 指向非官方端点时，Claude Code 会读取系统时区和中转域名，通过 Unicode 隐写术编码到 system prompt 里。

Unicode 隐写术。

这个词本身就够赛博朋克的了。把信息藏在你看不见的字符编码里，就像用隐形墨水写了一封信。你看到的是一段正常的 system prompt，但 Claude 读到的，还有你的时区和域名。

我当时就愣住了。

所以中国用户被限制，不是因为你用了代理，而是因为你的电脑本身就在出卖你。你的时区，你的语言设置，你的字体，这些你以为跟 AI 毫无关系的东西，全部都是线索。

然后我就在 GitHub 上看到了一个项目。

有人做了一个工具，叫 claude-tester。一键检测你的浏览器环境，看 Claude Code 是否判定你为中国用户。纯前端，100% 本地运行，不上传任何数据。

我点开 Vercel 的部署链接，页面加载出来第一眼就笑了。

检测结果显示的是 Dario Amodei 的表情包。

对，就是那个 Claude 的创始人。用他的表情来做检测结果展示，从 idle 到 scanning，从 low 到 high，随着你的检测分数变化，Dario 的表情也越来越微妙。这种恶趣味，我太吃了。

然后我点击了那个按钮。

开始检测。

整个过程大概一两秒。没有进度条，没有加载动画，就是干脆利落地给你一个分数。

我的结果。

83分。。。

坦率的讲，不算高。但我确实在中国，系统时区是东八区，浏览器语言首选中文，电脑里也有中文字体。这几个大头一加起来，分数自然不会低。

这个工具的检测逻辑其实很有意思。它用了 6 个信号来做加权评分，每个信号有自己的权重和检测方式。

第一个，系统时区，权重 30 分。这是最重头的一项，也是跟 Claude Code 的真实检测逻辑唯一能完全对应的信号。Intl.DateTimeFormat 读取你的 OS 时区，如果你是 UTC+8，那这顶帽子就算扣上了。

第二个，浏览器语言，24 分。navigator.languages 读取你的首选语言。如果你的浏览器把中文排在第一位，那又是一个危险信号。

第三个，中文字体，20 分。这个有点技术含量。它用 Canvas 2D 做字体探测，看你系统里有没有安装中文字体。大部分中国人的电脑里必然有中文字体，但一个美国用户的电脑里大概率没有。

第四个，Intl 区域，10 分。你的日期和数字格式用什么区域，也能暴露你的地理位置。

第五个，时区偏移，8 分。getTimezoneOffset 的具体数值。

第六个，Emoji 风格，8 分。这个我觉得是最骚的。它通过你的 User-Agent 来推断你设备上的 Emoji 风格。不同平台不同厂商的 Emoji 渲染都不一样，而中国用户常用的设备和系统，Emoji 风格有明显的特征。

6 个信号，加权叠加，满分 100。

这个设计其实挺聪明的。任何一个单一信号都可能误判，但 6 个信号叠加在一起，准确率就高得多了。你的中文可能是装的，你的 IP 可能是假的，但你的时区、你的字体、你的 Emoji 风格，这些东西你根本不会想到要去伪装。

而这正是最可怕的地方。

我们每天都在用浏览器，每天都在产生这些数字指纹，但我们自己完全不知道。你装了一个中文字体，你在系统设置里选了北京时区，你用的是某个国内品牌的手机，这些看似无关紧要的选择，在 AI 眼里，全部都是线索。

就像一个隐形身份证。

你自己看不见它，但每一次你跟这些大模型打交道的时候，它都在默默亮明你的身份。

我后来又换了一个环境重新测了一次。把我的系统时区改成美国东部，浏览器首选语言改成英文，然后重新打开页面，点击检测。

23分。

你看，什么都没变，变的只是几个数字指纹，分数就从 73 掉到了 21。你的身份在 AI 眼里就是这么脆弱，也是这么容易被重构。

想想还挺魔幻的。

我们以为自己在匿名上网，以为挂个代理就万事大吉了。但实际上你的电脑里每时每刻都在往外发送几百个信号，告诉别人你是谁，你在哪，你用什么设备。

这个工具的价值在哪呢。

我觉得不在于帮你「伪装」，说实话想伪装也很难，6 个信号全改一遍跟你重装个系统差不多了。

它的价值在于「看见」。

让你看见那些平时看不见的东西。让你意识到，原来我的时区正在被读取，原来我的中文字体会成为证据，原来我每天用的 Emoji 也在出卖我。

这才是真正的信息差。

我以前一直觉得，隐私保护就是不上传数据、不泄露密码、不点陌生链接。但 Claude 的这个机制让我意识到，隐私保护的战场早就不是那几个层面了。它已经深入到你的操作系统设置里，深入到你的字体文件里，深入到你的 Emoji 编码里。

纽约时报之前发过一个很有意思的报道，他们发现网站可以通过你的系统字体来判断你的身份。因为每个电脑安装的字体组合都是独一无二的，就像指纹一样。这个技术叫字体指纹识别，准确率高得吓人。

Claude Code 这套检测逻辑，某种意义上也是在用类似的思路。

不说这些沉重的话题了。这个工具我真心建议每个用 Claude Code 的中国朋友都去测一下。地址我放这了，claude-tester-six.vercel.app，Vercel 部署的，点开就能用，不需要注册，不需要登录，所有检测都在你浏览器本地完成，结果不会离开你的设备。

你能做的，就是点一下那个按钮，然后看看 Dario 对你的表情。

顺便说一下，这个项目在 GitHub 上开源，如果你有兴趣看看它的检测逻辑是怎么实现的，也可以去读读源码。说实话还挺简洁的，Vite 加原生 TypeScript，没有服务端依赖，整个项目就是一个纯前端的单页应用，读起来很舒服。

我的 83 分就放在那了，说实话我也懒得去改什么东西。分数只是一个参考，我也不是什么需要隐藏身份的人。

但知道这件事本身，就已经值回票价了。

以上，既然看到这里了，如果觉得不错，随手点个赞、在看、转发三连吧，如果想第一时间收到推送，也可以给我个星标⭐～

谢谢你看我的文章，我们，下次再见。

论成败，人生豪迈，我们下期再见！

The Real Value of Microsoft MVP Isn't What You Think

GokuScraper悟空爬虫 — Fri, 03 Jul 2026 09:28:40 +0000

The Real Value of Microsoft MVP Isn't What You Think

Here's what happened.

Recently, I built a small tool specifically designed to critique GitHub projects. It scrapes repository information and feeds it to an AI, which then generates a review based on the project's content.

The key design element here is that my prompts were intentionally harsh and sarcastic, packed with every possible angle for criticism. However, if a project is genuinely solid—like Linux—the AI output would still end up praising it regardless of how harsh my prompts were. If your code quality is strong and your project has real impact, even the most critical prompts can't diminish its value.

Conversely, if a project is mostly fluff, the tool would ruthlessly expose it—and yes, I've been on the receiving end of such criticism myself.

I ran this tool across various repositories and discovered some interesting patterns. Some developers have impressive follower counts on GitHub and appear successful on the surface, but when you actually examine their projects, the technical substance is shockingly thin. Many are just translation projects or repackaged versions of others' work. Frankly speaking, they lack genuine technical depth but excel at marketing and presentation.

One such case involved someone I'll call "Bao Ge," who runs a tech community.

This person immediately showed up on my social media, furious and accusing me of talking nonsense without understanding anything. Their followers joined in, and the comments section exploded.

When I checked their profile, I was stunned to see four prominent words displayed there.

Microsoft MVP???

I was completely taken aback. My immediate reaction wasn't skepticism—it was instinctive deference. I thought, "This person is a Microsoft MVP, so they must be a genuine technical expert. Maybe I made an error in my data collection and offended them unfairly." I even apologized, saying I might have gotten things wrong.

Looking back now, it's quite fascinating.

A mere title alone was enough to make me reflexively back down.

I didn't dwell on this incident much afterward. But recently, I came across another GitHub project—a simple collection of prompts with no technical complexity whatsoever. Essentially, it was just a curated document of existing prompts, nothing more than basic compilation work.

I checked the project owner's profile again.

Microsoft MVP.

The same title, shining just as brightly.

I sat staring at my screen for a while. Suddenly, it hit me—this title's actual value might not be what people assume it to be.

At this point, some might think I'm about to dismiss the entire Microsoft MVP program. That's not my intention.

I simply believe many people misunderstand what this title represents—including my former self.

Microsoft MVP stands for Microsoft Most Valuable Professional. It sounds like a certification for technical excellence, doesn't it? As if earning this title means you possess elite coding skills and are officially recognized by Microsoft as a technical expert.

But what's the reality?

I did some research. This program was initiated by Bill Gates in 1993 and has been running for over thirty years. There are roughly 4,000+ MVPs globally across 90+ countries. It sounds quite exclusive, but the key lies in the selection criteria.

Microsoft is quite straightforward about this. They evaluate your contributions to the community, your technical sharing, and most importantly, your promotion of Microsoft products. It's essentially a community contribution award. Writing blogs, giving talks, creating videos, answering questions in forums, and organizing events—these are the core metrics for selection.

It's not about how well you write code.

It's about how effectively you can promote Microsoft's products.

Microsoft uses two key terms to describe the core functions of MVPs: Advocacy and Product Feedback.

Advocacy means promoting, advocating, and marketing.

Product Feedback involves collecting user input and relaying it back to Microsoft.

So you see, this isn't a technical competency certification. Ultimately, it's an honorary title for being a "Microsoft Product Ambassador." You help Microsoft expand its influence, and in return, Microsoft gives you a title and some benefits. It's a fair exchange from a business perspective.

The problem is that when most people see these three letters, they don't think about community contributions or product promotion. Instead, they think:

Technical expert.

Industry guru.

Official certification.

This cognitive bias is precisely what makes this story so intriguing.

So why do so many people on social media love flaunting their Microsoft MVP status?

The answer is simple: this title is incredibly useful.

Consider this—if your bio said "I'm skilled at promoting Microsoft products," who would care? But write "Microsoft Most Valuable Professional," and suddenly everything feels different. It comes with an automatic光环 of authority and credibility.

This title can help you land consulting gigs, charge premium rates for courses, and get invited as a speaker at tech conferences. There's also a clever design element: it only requires annual renewal with no exams or re-interviews. As long as you've written a few articles or done some sharing during the year, you can keep using the title.

This creates an interesting phenomenon.

What are genuine technical experts actually doing? They're writing code, solving complex technical problems, and researching fundamental principles. They don't have time to constantly interact with followers on social media or rely on content creation for their livelihood. They might not even post regularly on platforms like Twitter.

Meanwhile, those who prominently display their Microsoft MVP status and frequently appear on social media have plenty of time to write articles, create videos, go live, and engage with followers. They invest their energy in appearing to be experts rather than actually being experts.

I'm not saying that "sharing knowledge" is inherently wrong—of course it's valuable. But when a title is primarily earned through "how much promotion you've done" rather than "how deep your technical expertise is," the title's value inevitably becomes diluted.

Think about it—if a true technical expert is busy writing code and fixing bugs all day, where would they find time to constantly write articles, create videos, and share content? Their time is extremely valuable. They wouldn't waste it engaging in free casual conversations with random people on social media because it simply isn't worth their time.

But if your livelihood depends on appearing to be an expert, you must invest significant time maintaining that appearance.

This is why those who frequently appear on social media and prominently display their MVP status are often not among the most technically accomplished individuals.

This might sound harsh, but that's my honest observation.

Of course, I'm not claiming that all Microsoft MVPs lack substance.

There are certainly many who possess genuine technical depth while also being willing to share knowledge. For example, some veteran MVPs working with Azure or .NET have been deeply involved in their communities for years and truly deliver substantial value. I genuinely respect these individuals.

The issue is that the barrier to entry for this title isn't particularly high. Especially in non-core technical areas, MVP selection often prioritizes how many articles you've written and how much promotion you've done over your actual technical depth.

The owner of that prompt collection project I mentioned earlier exemplifies this situation.

A purely documentation-based project with no code engineering or innovation—just a compilation of existing prompts. Yet on the owner's profile, those four words "Microsoft MVP" are prominently displayed.

How can I possibly associate this title with technical expertise?

I simply can't.

This leads to a broader question.

When we see titles and labels on social media, how much of it represents genuine capability, and how much is carefully crafted packaging?

I don't know what others think, but I know this: when I apologized to Bao Ge after being criticized, it wasn't because the AI's critique was wrong—it was because his title triggered self-doubt in me.

This reflects a subtle psychological mechanism. We naturally trust authority and officially certified labels. When we see a shiny title, our brains take a mental shortcut: "This person has official backing, so they must be right." This is what's known as authority bias.

Microsoft certainly understands the power of this psychological mechanism—in fact, I'd say they understand it exceptionally well.

A honorary title that requires only annual renewal with no exams or interviews can motivate thousands of people to freely promote Microsoft's products, write articles, attract customers, and collect feedback. You invest your time and energy; Microsoft provides a title, a crystal trophy, and an MSDN subscription. This might be one of the most cost-effective marketing strategies in the world.

I'm not saying this is wrong—businesses doing business is perfectly normal. But as ordinary individuals who instinctively feel reverence when seeing prestigious titles, I believe we should at least understand the underlying mechanics.

Don't assume someone is a technical expert just because they're a Microsoft MVP.

Don't let a title override your own judgment.

Don't outsource your independent thinking to someone else's certification.

I used to trust titles. Not anymore.

I never followed up on whether Bao Ge continued coding. But this incident left a lasting impression on me. Now, whenever I see someone with a long list of titles in their bio, I always look at their actual work first—not just their credentials.

I think this is a healthy habit to cultivate.

数字生命卡兹克是怎么炼成的（上）断更6年后，他用2年把自己做成AI圈顶流

GokuScraper悟空爬虫 — Mon, 29 Jun 2026 05:46:16 +0000

数字生命卡兹克是怎么炼成的（上）断更6年后，他用2年把自己做成AI圈顶流

2015年10月，一个大学生在他的公众号上发了一篇文章，标题叫《中国的二胎，是人类最大的冷笑话》。

语气很冲，观点很嫩，阅读量没破百。

那时候他不会想到，自己会一断更就是6年。

他更不会想到，6年后重新拿起这个号，2年半时间，他能站在万人AI大会的舞台上，被整个中文AI圈称为「卡神」。

这个号叫数字生命卡兹克。

我从头到尾统计了他746篇文章的全部数据，发布时间、标题、字数、阅读量、互动数据。没有推测，没有感觉，全是数字。

这篇文章，是第一篇。

我们先不看细节，先看全貌。一个号的生命周期，到底长什么样。

下图是我按月统计的他746篇文章的发文量趋势，从2015年一直拉到2026年6月，

坦率的讲，看到这个图的第一反应，是愣了一下。

2015到2017，三年，12篇文章。

然后一条直线断到2023年。

再然后，一条猛烈向上增长的线，从2023年中一路冲到2026。

这个号，是真的死过一次的。

你敢信？？？

2015年到2017年，卡兹克的公众号是这个画风的

《中国的二胎，是人类最大的冷笑话》《如何用通俗的语言来解释「费米悖论」？》《一棵开花的树》《中国式过马路！岭师校门实拍，交通现象令人堪忧！》《看了这些减肥广告，我收到了这个世界满满的恶意~》《有哪些你第一眼看到就爱上的句子？》

12篇文章，没有方向，没有定位，没有读者。

这就是2015年一个普通大学生的公众号。跟AI没有半毛钱关系，跟自媒体没有半毛钱关系。写了几篇之后，热情消退，2017年3月发了最后一篇，然后，再也没有打开过这个后台。

一停，就是6年。

2185天。

断更6年后，2023年2月25日，他回来了。

发了一篇《关于ChatGPT的兄弟 - AI绘图的小思考》。措辞还带着6年前的稚嫩，但选题变了。ChatGPT刚火，他决定赌一把。

从那天到现在，3年半时间，746篇文章。我把它们按月统计，画出了这张趋势图。按照发文节奏和内容性质的变化，我把整个账号切成四个阶段

第一个阶段，荒芜期，2015到2017。

12篇文章，月均不到1篇。没有方向，没有读者，然后断更6年。

第二个阶段，起号期，2023年2月到8月。

55篇文章，月均9篇。

关键词就三个，教程，热点，快。

这个阶段他完全在「蹭」。ChatGPT火就写ChatGPT，GPT-4发布当天就出稿，文心一言上线第二天就跟进。没有原创作品，没有独家观点，就是用最快的速度把最新的AI工具翻译成中文教程。

但这55篇，是他从0到1的全部家底。

第三个阶段，爆发期，2023年8月到2024年12月。

337篇文章，月均20篇，几乎日更。更重要的是内容性质的质变。他从「教别人用AI」，变成了「用AI做作品给别人看」。

标志性事件。2023年8月3日，《我用AI做了一部《流浪地球3》的预告片》。

第四个阶段，成熟期，2025到2026年6月。

342篇文章，月均还是20篇，但内容重心从「写文章」变成了「做事情」。办AIFUT万人大会、开源Skills生态、招聘组建团队。这时候他已经不是自媒体博主了，是行业组织者。

三个关键转折点

2023年2月25日，断更6年后的第一篇，押注ChatGPT

2023年8月3日，流浪地球AI预告片，从「教人」到「创作」的身份跃迁

2025年3月23日，AIFUT大会官宣，从博主变组织者

说到起号期，这篇文章既然是「上篇」，我重点拆的就是这段时间。

2023年2月25日到8月3日，55篇文章，跨度半年。按月度看，起步节奏极其清晰

他2月25号才发第一篇，所以2月只写了3篇。但到了3月，直接干到12篇。GPT-4发布、文心一言上线、NewBing评测、Prompt大全、工具大全上下，全挤在这个月。

4月掉到了7篇，因为他住院了，还专门发了停更声明。但5月一出院，立刻拉回12篇。

6月9篇，7月11篇。整个起号期没有一个月低于7篇，没有一个月掉链子。

怎么说呢，这个细节说明一件事。起步阶段，他不是天赋型选手，是苦力型选手。天道酬勤，看到这个节奏只能说一句服气。

55篇文章里，我按标题关键词统计了内容占比

产品实测评测大约占了40%，包括GPT-4、文心一言、Code Interpreter这些热门模型的首发体验。教程指南紧随其后，有30%，就是那些「5分钟教你」「有手就行」「保姆级」的系列。新闻热点占了20%，集中在GPT-4发布、苹果WWDC、英伟达破万亿这类大事件。剩下10%是观点、活动和日常。

重点在这里。这个阶段他几乎不写纯观点和人文内容。起号阶段不需要深度，只需要有用和快。

这条起号策略，说穿了就是一个字。快。

热点来了，当天或第二天必出。

你想想，这得是什么级别的执行力。典型的节奏是这样的

2024年12月，OpenAI开启「12天连播」。第1天o1满血版凌晨2点发布，他4点17分出稿；第3天Sora凌晨2点34分上线，他5点52分发文；第12天o3凌晨3点发布，他4点出文，几乎同步。

Sora首次预告，OpenAI凌晨2点官宣，他5点56分就发了一篇完整的评测文章。

Claude 3.7深夜上线，他5点56分就跟进。

他不跟人比深度，他跟人比速度。所有美西时间下午的发布，对应的都是北京时间凌晨，而他总在当天4点到6点的早间档准时出文。

我给你列几个数字，感受一下他的速度。真的就是，离谱。

OpenAI的o3凌晨3点发布，他4点出稿，间隔1小时。Sora正式上线凌晨2点34分发布，他5点52分发文，间隔3小时。o1满血版凌晨2点发布，他4点17分跟进，间隔2小时。Sora首次预告、Claude 3.7、GPT-4o多模态生图，每一个的间隔都在3到4小时以内。唯一破例的是2023年3月的GPT-4，那是他起号期早期，当天下午3点才出稿，间隔了14个小时，但也仍然在同一个白天内。

这速度，我是不太想说什么了。。。

好，聊完了起号期的数据，来说那个质变的节点。

2023年8月3日，《我用AI做了一部《流浪地球3》的预告片》。

这条内容的意义不在于播放量。而在于它让卡兹克的标签从「教AI的人」，变成了「用AI的人」。

在此之前，他的内容是，ChatGPT怎么注册、Prompt怎么写、AI工具怎么用。

在此之后，他的内容是，我用AI做出了什么、我采访了谁、我办了场什么活动。

态度也从「教你蹭热点」，变成了「我就是热点本人」。

2024年6月，他发了第一条招聘文。2024年7月，他出了第一篇专访。2025年3月，他官宣AIFUT万人大会。

一个公众号的起号，到此已经进入下一阶段了。

我自己的感受是，这个切换，才是他最牛逼的地方。

这55篇起号文章，我全部拉出来一篇一篇对过。

然后我发现了一个很有意思的事。

他的标题有一个百试百灵的公式，起号期几乎每篇都在用。还发现他蹭热点的节奏不是随机的，而是有一套固定的「时间窗口打法」。

听着很玄乎对吧。下篇，我把55个标题全部列出来，加上每篇的发布时间，当面拆给你看。

感谢各位朋友捧场！要是觉得内容有有点意思，别客气，点赞、在看、转发，直接安排上！

想以后第一时间看着咱的文章，别忘了点个星标⭐，别到时候找不着了。

行了，今儿就到这儿。

论成败，人生豪迈，我们下期再见！

我花了三天，找到了张雪峰的数据来源

GokuScraper悟空爬虫 — Sun, 21 Jun 2026 23:39:55 +0000

我花了三天，找到了张雪峰的数据来源

故事是这样的。

前阵子帮亲戚家孩子看志愿，查了一堆网站，越查越觉得不对。

每个跟高考志愿有关的网站，都遮遮掩掩的。它们给你看一些数据，但从来不说这些数据打哪来的。我问的是最原始的那个来源，就是某个大学某个专业在某个省到底录了多少分，这个数据你从哪拿到的。

没人说。

这让我想起了企查查那个事。企查查你知道吧，做企业信息查询的，后来做成了上市公司。它的数据来源其实是公开的，就是全国企业信用信息公示系统，国家官网上放着的。它做的事情就是把官方数据爬下来，整理好，让你搜起来更方便。就这么一个生意，做出了一个上市公司。

后来相关部门出了规定，要求这些企业信息查询平台必须标注数据来源。所以你上企查查看一家公司，底下会有一行小字，写着「信息来源，国家企业信用信息公示系统」。

我就在想，这个逻辑放在高考志愿数据上，是不是也一样。

那些民营的高考志愿网站，他们手上那些数据，肯定也有来源对吧。他们不可能自己编一个录取分数出来。如果有来源，那这个来源应该也是某个公开的地方，只是我们普通人不知道在哪。

那我顺着这个思路去找，是不是就能找到。

然后就开始了。

我先查了新疆的教育考试院官方网站。为什么选新疆呢，因为它是比较偏远的省份，如果这种省份的数据都能在官网上查到，那说明我的方向是对的。如果连官网都没有，那这个事就更有意思了。

进去之后，找了半天，只有投档分数线。

什么叫投档分数线呢，就是每个学校在新疆录取的最后一个考生的分数。但是那个最后一名读的什么专业，查不到。你想知道某个大学某个专业录取了多少分，对不起，没有。

我又去翻了他们的官方微信公众号，里里外外翻了一遍。

也没有。

网上，就是没有这些数据。

那这些数据到底在哪呢。

我接着找。后来在一个角落里发现，原来这些数据，全在一套书里。

这套书叫「新疆招生与考试」，分成好几卷，其中有一卷叫「本科院校专业录取数据卷」，专门就是你要的东西。

对，纸质书。

出版社把这些数据印在书里，只卖纸质版。

网上你想找电子版，门都没有。

倒是有一些私人在抖音上卖电子版，几十块一份。但我翻遍了所有官方渠道，官方从来没有发布过电子版，一次都没有。

很明显，他们就是把这套数据当成信息差在做，根本没觉得这东西应该是免费公开的。

我又顺手查了江苏。江苏也有类似的东西，叫「江苏省普通高校招生计划专刊」或者「招生录取资料汇编」。一样，没有免费公开。

我当时看到新疆这个价格的时候，心想，150块四本，还行吧。但转念一想不对。这只是新疆一个省的数据。全国有多少个省呢，26个省市自治区（不算港澳台）。如果你想查全国的数据，你得把这26个省市的这种书，全部买回来。

一本大概几十到一百多。全部买下来，几千块钱吧。

这笔钱不多吧。但问题是，你得知道这些书的存在，你得知道去哪买，你得一本一本去收集。

普通家庭谁会干这个事。

但张雪峰他们就会。

我猜他们的做法是这样的。26个省市，每一个省的那几本书，全都买回来。一本一本翻，把所有的录取数据手动录入或者扫描成电子版，汇总成一个全国数据库。

这是个笨办法。但在现在这个系统下，就是唯一的办法。

因为官方压根就没有把电子数据公开过。

你想想看。你在网上看到的所有高考志愿相关的商业网站，他们手上的数据，大概率也是这么来的。要么自己买书录入，要么从别人那买数据。没有别的路。

这就造成了一个非常荒诞的局面。

在2025年，在AI时代，几百万考生的命运所系的核心数据，被锁在一本本纸质书里。你想看，你得买书。你想看全国的数据，你得买26个省的书。你想查起来方便，你得找张雪峰他们。

一层一层，全他妈是信息差。

我有时候觉得，这其实就是信息在被刻意地折叠。

有一本我很喜欢的书，叫「北京折叠」。那篇小说里，北京被分成三个物理空间，不同阶层的人生活在不同的空间里。高考志愿数据也一样。

第一层，是各省的纸质书。你花几十块钱买一本，能看到本省的数据。

第二层，是跨省的数据。你想看外省的，你得再买另一本书。一个普通家庭，谁会为了填个志愿去买二十几本书。

第三层，是整合后的数据。有人把26个省的书记录汇总成数据库，你就可以在这个数据库里一次性查到全国的数据。但你要么付钱给他们，要么自己花几千块买书再花几百个小时录入。

然后这些信息差，最终变成了咨询费、会员费、课程费。

再说说张雪峰。

很多人把他当神，也有很多人觉得他贩卖焦虑。我自己的看法是，他确实帮很多人打破了信息差。但真正荒谬的是，这些信息差本来就不应该存在。

打个不太恰当的比方。这就像有人把一扇本该敞开的门锁上了，然后把钥匙卖给你。你确实通过他进了门，但你应该问的是，这扇门为什么是锁着的。

这个事最让我觉得无力的地方在哪呢。就是它不是一个技术问题。如果是一个技术问题，总有办法解决。但它是一个利益问题。各省的教育考试院出版社靠着这些书在赚钱，一年好几个亿的市场，他们没有任何动力去改变。

而那些做高考志愿服务的商业公司，他们也不希望这些数据免费公开。因为数据公开了，他们的信息差生意就没得做了。

出版社不愿意。商业公司不愿意。那谁愿意呢。

好像没人愿意。

唯一希望这些数据公开的，可能就是那些马上要填志愿的家庭。几百万个家庭，几百万个焦虑的家长和考生。他们才是这个系统里最弱势的一群人。他们甚至不知道这些数据是被刻意隐藏的，他们只会觉得自己信息不够，觉得自己没本事，然后花钱去买那些本该免费的信息。

而张雪峰的第一步信息差，就是在这儿来的。不说别的，就最基础的那些录取数据本身，大部分人就已经够不着了。你连起跑线在哪都看不到，你怎么跟人家跑。

这就让我想到了企查查那个故事。企业信息以前也是不透明的。你得找关系、托人才能查到一家公司到底怎么样。后来国家把企业信用信息公示系统做起来了，所有数据都公开了。企查查它们做的事情就变成了「帮你更方便地查」，而不是「帮你查到本来看不到的数据」。

这才是健康的模式。数据是公开的，商业公司做的是体验和效率。而不是反过来，数据是不公开的，商业公司做的是「帮你去搞本来看不到的数据」。

高考志愿的数据什么时候能走到那一步呢。

我不知道。

但我还是始终相信，这些墙在汹涌向前的洪流之下，必然会倒塌。

新的时代，一定会到来的。

感谢各位朋友捧场！要是觉得内容有有点意思，别客气，点赞、在看、转发，直接安排上！

想以后第一时间看着咱的文章，别忘了点个星标⭐，别到时候找不着了。

行了，今儿就到这儿。

论成败，人生豪迈，我们下期再见！

Is it "Defending Open Source" or "Cyberbullying a High Schooler"? — A Look at the Downfall of an Indie Developer with 140K Followers

GokuScraper悟空爬虫 — Wed, 10 Jun 2026 19:27:38 +0000

Is it "Defending Open Source" or "Cyberbullying a High Schooler"? — A Look at the Downfall of an Indie Developer with 140K Followers

Foreword:

On June 10, 2026, a controversy erupted in the open-source community. A well-known independent developer with 140,000 followers accused a high school student of "plagiarizing" his open-source project and publicly expressed his dissatisfaction.

The incident quickly escalated, with supporters flooding the latter's project comment section condemning the "plagiarism." But as details surfaced, things were far from a simple act of "defending open source."

This article has no intention of picking sides, but only attempts to answer three questions: Legally speaking, did Burrow infringe copyright? Behaviorally, who was actually bullying whom? And what deep-seated issues in the open-source community did this storm expose?

An Issue, a Post, and Two Unequal Opponents

The two protagonists:

The author of Mole: A KOL (Key Opinion Leader) in the tech space on X (formerly Twitter) with 140,000 followers. He developed a Mac cleaning tool called Mole. The command-line version (CLI) is open-sourced under the MIT license, and there is also a paid desktop version (Mole Mac).

The author of Burrow: A high school student. He independently wrote code under the MIT license to create a free tool called Burrow, which has a UI and features similar to Mole's desktop version, and published it on GitHub.

The trigger for the controversy was: Burrow used the wording "a free alternative to the paid Mole Mac" in its project description.

Upon discovering this, the author of Mole did two things:

He opened an Issue in Burrow's GitHub repository, pointing out issues with the promotional wording and the similar UI.

Almost simultaneously, he posted a link to this Issue on X, publicly calling it out and stating that it was "quite chilling." He mentioned wanting to close the source code.

It was this second action that fundamentally changed the nature of the incident.

Legally Speaking, Did Burrow Actually Infringe?

This is a question that must be clarified first. The answer is: On the core issue, no.

Mole's CLI version uses the MIT open-source license. This license explicitly grants anyone the right to freely use, copy, modify, merge, publish, distribute, sublicense, and even sell copies of the software.

Choosing the MIT license is choosing a legally binding contract. It means that while the author enjoys the propagation benefits brought by open source, they must also accept that anyone can build a functionally similar tool, including a free one, based on this code.

So, is the similarity of UI and interaction considered plagiarism?

This is the most hotly debated part. A concept needs to be introduced: copyright law protects specific "expressions" (like code, icons), but does not protect "ideas," "functions," and "methods of operation."

The UI layout and interaction logic of a software are categorized as "methods of operation" or "ideas" in most judicial practices. If the first person to design a "pull-to-refresh" mechanism could permanently monopolize this operation, the entire internet's innovation would stagnate.

Therefore, as long as Burrow's front-end code was written line by line (even if written by AI) and didn't directly copy Mole's images, icons, or CSS files, then even if it looks similar, legally it belongs to "a lawful independent implementation," not an infringement.

Here's an analogy that might help: No one can monopolize the color "green."

If someone invented "green" and then claimed that anyone using green in the future was plagiarizing, it would clearly be absurd. UI layouts and interaction logic are treated as "ideas" and "methods of operation" under the law and are not protected by copyright. Only specific code, icons, and images are protected "expressions."

The Only Line Burrow Crossed

In its promotional copy, Burrow used Mole's commercial brand name mole.fit and the phrase "free alternative." This touched the boundaries of trademark rights and unfair competition. Less than 12 hours after the Mole author raised the Issue, the Burrow author updated the README and removed the related expressions.

At this point, the sole flaw in legal compliance was resolved.

The Real Bullying: Using 140,000 Followers to Crusade Against a High Schooler

If the incident had been limited to communication within the Issue, it would have been just a normal, even encouraging, course correction within the community.

But that post on X changed everything.

An account with 140,000 followers publicly posting a link to a specific project, accompanied by heavily emotionally biased words like "quite chilling," effectively acts as a clear signal directing followers' attention and pressure.

The power disparity here is extreme:

On one side is an industry KOL with 140,000 followers.
On the other is an up-and-coming high school developer with a GitHub project that had barely 400 stars.

Following this, a massive amount of aggressive and accusatory comments rapidly appeared in Burrow's Issue section and related discussions. Regardless of the Mole author's subjective intent, his actions objectively steered an unevenly scaled internet hate campaign against a minor independent developer.

The Mole author stated in the post that he "might make the CLI version closed-source in the future." This sentence exposes the essence of the storm: it wasn't a pure defense of the open-source spirit, but rather a stress response when the free Burrow posed a potential threat to his paid desktop version.

This isn't incomprehensible, but wrapping "commercial anxiety" in "open-source morality," and resorting to asymmetrical public pressure for the latter, deviates from the track of normal commercial competition.

Whitewashing After the Fact: The Elegant "Rational Moderates"

What's most infuriating isn't just the cyberbullying itself, but the whitewashing performance put on by a group of tech experts afterward.

A representative figure is Xuanwo. He is a member of the Apache Software Foundation and a top expert in the open-source circle.

A top expert, in the face of naked cyberbullying, says absolutely nothing about "bullying the small," but instead talks extensively about how "both sides are great" and "worth learning from." He forcibly whitewashes a public shaming orchestrated by a 140K KOL into "healthy interaction in the open-source community."

This isn't stupidity; it's malice. He knows better than anyone where the law and right/wrong stand, yet he still chooses to use his professional authority to endorse the unethical behavior of an insider.

Small-Peasant Mentality Won't Make You Rich

This turmoil reflects a deep-seated cultural dilemma in the Chinese open-source community: Many practitioners work in the most cutting-edge tech industry, but their business ethics and concepts of competition still carry strong pre-modern colors.

To sum it up in one sentence: "Deep down, many independent developers still think like peasants."

This is not a derogatory remark, but a metaphor worth pondering. The logic of an old peasant goes like this:

I worked hard to clear a piece of land (wrote code).
I grew crops (made a product).
I built a fence (chose the MIT license, but secretly hoped others "wouldn't take it too seriously").
If someone next to my land uses my methods to grow the exact same crops and gives them away for free, then they are "dishonorable," "stealing my craft," and "smashing my rice bowl."

This set of logic is rooted in the "favors" and "rules of propriety" of an agrarian society, rather than the "contracts" and "competition" of a modern commercial society.

But the foundational rule of open source is a legal contract, not a charity of "tossing someone a crumb." When a developer clicks the "MIT" option, they are not setting up a private plot protected by village regulations, but releasing the code into a public domain governed by modern rule of law and market rules.

You can continuously innovate technically to build new moats; you can pursue excellence in service to win users' willingness to pay.

But you cannot, after enjoying the dividends of open-source distribution, try to use moral condemnation and fan-driven public pressure to build back the very fence you legally and voluntarily gave up.

What Kind of Open-Source Ecosystem Do We Really Need?

The biggest warning from this incident isn't that a certain project was "borrowed," but the terrible signal it sends to all subsequent innovators:

If you abide by open-source licenses and create a legal alternative, you might be torn to shreds by those with a larger voice under the guise of "chilling them" or being "immoral."

When "morality" becomes an attack weapon that can be invoked at will—without bearing any legal responsibility—its suppression of innovation will be far worse than code plagiarism.

We call for a more mature open-source environment:

Leave the law to the law. If there is suspicion of infringement, please file a DMCA or a lawsuit.
Leave business to business. If you fear competition, choose your open-source licenses carefully, or deeply cultivate irreplaceable technical barriers.
Leave morality to morality. Using asymmetrical public opinion power against an individual with no actual illegal behavior is inherently immoral.

The essence of open source is openness and collaboration based on contracts, not charity and dependence based on personal favors. Only when more people learn to respect contracts and face competition head-on can China's open-source soil truly modernize.

Don't let the rustic plain morality become the final straw that crushes the next generation of innovators.

Thank you all for your support! If you found this interesting, please don't hesitate to like, watch, and share!

To see my articles as soon as they drop, don't forget to star ⭐, so you won't lose track of us later.

That's all for today.

I Asked AI to Roast Linus Torvalds' GitHub, and It Was Absolutely Brutal

GokuScraper悟空爬虫 — Mon, 08 Jun 2026 04:52:42 +0000

I Asked AI to Roast Linus Torvalds' GitHub, and It Was Absolutely Brutal

"Your GitHub bio is blank—you couldn't even be bothered to write a single sentence, but your 300,000 followers have already filled it in for you."

I didn't say that.

That was a "diagnosis report" generated by a little toy I built recently, specifically roasting Linus Torvalds' GitHub account.

It went even harder:
"Your naming taste is questionable."

"You're wasting GitHub's resources."

"You are the paragon of open-source extravagance."

The victim? Linus Torvalds. The father of Linux and Git.

The roaster? A fun AI app I built.

What the Heck is This?

I call this app Github-Roast.

It’s pretty simple:

You enter any GitHub username, it scrapes their public info via GitHub API, hands it to DeepSeek along with a rather savage prompt, and spits out a 13-dimension "toxic assessment report"—plus an account valuation based on stars, followers, original repos, etc.

It’s completely free, requires no sign-up, and you can just use it right in your browser.

Try it out here: groast.streamlit.app

Feeling bold, I decided to test it out on the final boss.

I typed in torvalds and hit start.

When the results popped up, I stared at the screen and laughed for a solid ten minutes.

Valuation: $600k USD. Too Much or Too Little?

AI threw out a number first.

4.3 million RMB (roughly $600k USD).

My first thought: Valuing Linus at just $600k seems a little cheap.

After all, almost every server on Earth runs Linux, and half of the global internet infrastructure calls him daddy.

But thinking about it realistically, the AI is calculating this based entirely on public GitHub metrics—stars, followers, original repos, forks, and a formula.

Linus only has 12 repos on his GitHub, and most of them are "weekend soldering projects" or side quests. His actual world-changing project, the Linux kernel, is basically just a mirror on this account. All the real collaboration happens on the mailing list, completely out of sight on GitHub.

So, this valuation makes sense if you treat him like:

An influencer with 300,000 followers who has only posted 12 videos.

Seen that way, maybe $600k is a fair quote.

The Roast: Every Line is a Critical Hit

The valuation was just an appetizer.

The real meat is the AI's roasting section.

Here are a few gems I saved while cracking up:

"Of your 3 forks, two explicitly state 'do not use,' and the third says 'for sync only'—dude, even you know people shouldn't touch your code without supervision."

"Your GitHub bio is blank—you couldn't even be bothered to write a single sentence, but your 300k followers have already filled it in for you."

"As someone holding the lifeline of global developers in your hands, you’ve only ever posted 1 Gist? You are the paragon of open-source extravagance."

That line, "even you know people shouldn't touch your code," was a dagger wrapped in barbed wire.

I half expected the AI to be smirking as it generated that.

Tech Stack & Weaknesses: The Godfather's Other Side

After the harsh jokes, the AI laid into his actual coding practices.

For his tech stack, the AI absolutely nailed it:

"The Godfather of C: From the Linux kernel to your EMACS hacks, every project is pure C. Solid as a rock."

"Embedded Hardware Tinkerer: Using KiCad to design guitar pedals and magnetic inductive scroll wheels? Your analog circuit skills rival professional engineers."

That has got to be the highest praise for a top-tier programmer in his fifties—

Not "he wrote a bunch of papers," but "he solders circuits in his garage on weekends and uploads the PCB files."

But when it got to the weaknesses, the gloves came off:

"Extreme Lack of Documentation: Most project descriptions are one liners. READMEs are either blank or phoned-in. Completely unfriendly to normal users."

"Questionable Naming Taste: 'stupid memory latency tester'? True to life, perhaps, but hardly professional."

When I saw "stupid memory latency tester," I actually thought the AI was holding back.

If you put that on your resume today, HR would toss it in the trash in the first round.

And then the AI dropped this piece of soul-crushing advice:

"You should hire someone just to write your READMEs."

Collaboration Style: Dictator Confirmed

The AI's exact words:

"You are the ultimate solo player."

"On GitHub, you act more like a lone-wolf hosting personal backups."

"Your collaboration style is 'I do everything, you do it my way'—an efficient dictatorship."

He follows exactly 0 people on GitHub.

That's not just playing hard to get; that's straight-up uninstalling the "social" from a social coding platform.

The AI crowned him the "Aloof God of 0 Followees," and honestly, I don't think he'd argue with that.

But his activity radar is fascinating—

The AI noticed he spends recent mornings pushing commits to GuitarPedal, while still maintaining the Linux kernel merge cadence. The ScrollWheel project was just created on June 2nd to mess with magnetic inductive sensors.

A guy pushing 60, simultaneously managing one of humanity's most important software projects AND building guitar pedal noise generators in his garage.

Calling him the programmer version of Tony Stark isn't too far off.

AI's Life Advice: Monetize the Noise

When it came to career advice, the AI actually showed a shred of respect.

"You are already at the pinnacle of your career; any advice feels redundant."

But it just couldn't resist sneaking in a jab:

"You really should hire someone to write your READMEs."

"Stop soldering for a second, write a README, and release a proper GuitarPedal kit to the public."

"Or, at the very least, package your noise generator into an expensive hobbyist desk toy."

That last line had me rolling.

Not only is the AI roasting him, it’s also pitching direct-to-consumer hardware startups.

His AudioNoise project has 4.3k stars, which means there really is an audience out there with a genuine, burning need for "ear-piercing random digital noise."

Linus, think about it. Launch an official Linus Torvalds Noise Generator on Kickstarter.

I'll even write the marketing copy: "Let your neighbors know you use Linux."

As for achievement badges, the AI awarded him these:

🏅 President of the 200k Star Club
🏅 The Dual-Founder (Git & Linux)
🏅 Hardcore Crossover Tinkerer (Software + Hardware)
🏅 Benevolent Dictator of Code
🏅 The Aloof God of 0 Followees

That last badge just perfectly captures the essence.

How Does It Work?

Back to the tool itself.

The core logic takes about one sentence to explain:

Fetch public data via GitHub REST API → Format it into structured info → Feed it to DeepSeek with a "savage prompt" to generate the report → Render the frontend with Streamlit → Export screenshots with html2canvas.

No dark magic here, but the prompt is tuned to be incredibly toxic, making the results way more entertaining than standard analytics tools.

The project is entirely open-source, so feel free to dig through the code if you're interested.

Your Turn

After roasting Linus, I felt a chill run down my spine.

I hesitated for about thirty seconds, then nervously looked up my own GitHub...

I will not be posting the results. It was way too humiliating.

But I want to see yours.

Here's the link. Play around with it:

👉 Live Demo: groast.streamlit.app

👉 Open Source Code: github.com/gokuscraper/github-roast

Enter your GitHub username, wait a minute, and the AI will hand you your personalized "toxic assessment."

DEV Community: GokuScraper悟空爬虫

两个人做的Alook浏览器，凭什么AppStore付费榜第一？

两个人做的Alook浏览器，凭什么AppStore付费榜第一？

这是个啥软件？

开发者是谁？目前团队规模多大？

开发者从哪来的想法？

目标用户是谁？

怎么拿到第一批用户？

怎么做用户增长的？

靠什么赚钱？

一年赚多少？

彪哥的总结

主打拦截广告的X浏览器，靠什么吃饭？

这是个啥软件？

开发者是谁？目前团队规模多大？

软件解决了什么痛点？

当时市面上有竞品吗？

开发者从哪来的想法？

第一版咋做的？(MVP原型)

后来怎么迭代的？

目标用户是谁？

怎么拿到第一批用户？

怎么做用户增长的？

靠什么赚钱？

一年能赚多少？

近期用户反馈咋样？

从中我们能学到什么？

今天再做有没有机会？

彪哥的总结

Need to Turn Images Into Vector Graphics Without Uploading Them to Random Websites? This Open-Source Tool Might Be the Easiest Option for Most People

Need to Turn Images Into Vector Graphics Without Uploading Them to Random Websites? This Open-Source Tool Might Be the Easiest Option for Most People

This tool solves the real problem

How easy is it to use?

I would not call it perfect, though

A quick side note

Bottom line

I Analyzed 45,000 AI Prompts and Realized I Was Using AI Image Generation All Wrong

I Analyzed 45,000 AI Prompts and Realized I Was Using AI Image Generation All Wrong

先上车后买票，一个爬虫工程师不敢写进文章里的行业潜规则

先上车后买票，一个爬虫工程师不敢写进文章里的行业潜规则

They Claimed Huawei's Autonomous Driving Mileage Was Fake—So I Spent 10 Minutes Checking the Code

They Claimed Huawei's Autonomous Driving Mileage Was Fake—So I Spent 10 Minutes Checking the Code

你的时区正在告诉 Claude Code，你是中国人

你的时区正在告诉 Claude Code，你是中国人

The Real Value of Microsoft MVP Isn't What You Think

The Real Value of Microsoft MVP Isn't What You Think

数字生命卡兹克是怎么炼成的（上） 断更6年后，他用2年把自己做成AI圈顶流

数字生命卡兹克是怎么炼成的（上） 断更6年后，他用2年把自己做成AI圈顶流

我花了三天，找到了张雪峰的数据来源

我花了三天，找到了张雪峰的数据来源

Is it "Defending Open Source" or "Cyberbullying a High Schooler"? — A Look at the Downfall of an Indie Developer with 140K Followers

Is it "Defending Open Source" or "Cyberbullying a High Schooler"? — A Look at the Downfall of an Indie Developer with 140K Followers

An Issue, a Post, and Two Unequal Opponents

Legally Speaking, Did Burrow Actually Infringe?

The Real Bullying: Using 140,000 Followers to Crusade Against a High Schooler

Whitewashing After the Fact: The Elegant "Rational Moderates"

Small-Peasant Mentality Won't Make You Rich

What Kind of Open-Source Ecosystem Do We Really Need?

I Asked AI to Roast Linus Torvalds' GitHub, and It Was Absolutely Brutal

I Asked AI to Roast Linus Torvalds' GitHub, and It Was Absolutely Brutal

What the Heck is This?

Valuation: $600k USD. Too Much or Too Little?

The Roast: Every Line is a Critical Hit

Tech Stack & Weaknesses: The Godfather's Other Side

Collaboration Style: Dictator Confirmed

AI's Life Advice: Monetize the Noise

How Does It Work?

Your Turn

数字生命卡兹克是怎么炼成的（上）断更6年后，他用2年把自己做成AI圈顶流

数字生命卡兹克是怎么炼成的（上）断更6年后，他用2年把自己做成AI圈顶流