Pinned Tweet
我是 gptkit.eth,全栈开发,做 iOS 逆向和移动安全研究(防御视角),这两年主要在折腾 AI Agent。这里主要写三类内容:① Agent 实战:自己做、自己测、自己翻车 ② AI × 安全:沙箱、护栏、提示词注入实测 ③ 工程观点:怎么让 Agent 的产出可核对。代表作👇
1
11
374
I designed my live GTA 6 video site with taste-skill, then spent $7 trying a new look using Uizze’s recommended workflow.
Dark on the left, light on the right. Which do you prefer?
Try the live version: gta6.terryso.dev/
2
1
26
SimplexAI 的 Linger 开源了,本地跑个人 Agent,16GB Mac 就能装。
看了一眼,16GB 那档默认只跑 4B 模型,24GB 才到 27B;邮件、日历、语音、后台整理记忆还是要联网,Apache 2.0 还加了条商业托管限制。
16GB 的别指望它多聪明。
github.com/simplexai-labs/li…
1
32
I've been learning Cantonese, and the textbooks weren't cutting it — so I built a small site: real Cantonese and Mandarin audio, sentence by sentence, from actual conversations. Vlogs, cha chaan teng dialogues, the stuff people actually say.
sheng.terryso.dev/
3
3
69
Danshipper 在 Every 测了几周 Astra 6.2 / Sol 6.2:不如 Fable 5.2。
写作训得太短,难读懂;代码实战常提前停;设计全是“家风格”。
别急着换,等适配。
nitter.cf/danshipper/status/2122…
BREAKING:
OpenAI just launched Astra 6.2 and Sol 6.2!!
We've been testing them for the last few weeks @every and, unfortunately, these models aren't great. Overall a disappointing release compared to Fable 5.2.
- Writing. We found Astra 6.2 to be a step back from Astra 6 on writing. 6.2 appears to have been trained to very extremely concise and actionable—so much so that its writing is hard to understand, and it gives short answers when long answers would be appropriate.
- Coding. We found performance to be similar to Astra 6 on most of our benchmarks. However, in real coding tasks Astra 6.2 seemed to stop too early frequently, which made it quite frustating to use on end-to-end builds.
- Design. Astra 6.2 has good default design instincts. But it struggles to vary its designs, resulting in output that has a noticeable "house style".
Full vibe check coming soon on @every!
1
37
微软出了个给 coding agent 省钱的插件 TokenEconomizer:读你仓库里跑过的会话,剪掉重复的发现和从来不用的工具。
277 个 SWE-Bench 任务实测,成本降 18.4–30.9%,成功率基本不动。第 6 个任务起,每个仓库都开始比默认便宜。
microsoft.com/en-us/research…
1
41
Muse 终于能搜 X 了。
装一个技能 muse-grok-x-search + Grok Build CLI,设备码登录一次就行。实测 21 秒返回 5 条,抽查原帖是真的。
只读搜索,重要判断回原帖核对。
github.com/kgmkm/muse-grok-x…
2
71
Claude Code 的 Projects 现在有个会主动汇报的总管:并行线程跑着,你不用一个个去查,它在项目聊天里直接回答问题、推送各线程进度。
等于把“盯会话”这活也外包出去了。Pro/Max 公开 beta,另加只读分享链接、提醒、/btw 和 /fork。
code.claude.com/docs/en/clau…
1
47
一台 1994 年的 Jeep,没有 OBD2 口,也没有能用的诊断仪。车主花 3 美元接了块 Arduino Nano,让 Claude Code 和 Codex 直接跟发动机电脑对话:一周读完 32KB 固件,找到加州版 ECU 在尾气检测工况下偷减油的原因,最后过了检测。
编程 agent 的尽头是修车。
youtube.com/watch?v=WPj3aQ-x…
1
93
Claude 现在直接坐进 Google Meet 开会了:边聊边写文档、画白板、记待办,会还没开完活已经干了一半。
以前 AI 是会后整理纪要的,现在是会中同事。差别不在记笔记,在它能当场查你接好的工具、把决定变成可执行的下一步。
但只在 Team/Enterprise beta、只支持 Google Meet。Zoom/Teams 的团队先别急。
claude.com/resources/article…
1
1
81
同一个 Agent 连上几台电脑,方便的背面是:发一句「帮我处理」,它到底去哪台执行?
公司文件留在公司电脑,个人账号留在个人电脑。同一个任务,换台机器,碰到的文件、登录态和权限就全变了。
多机 Agent 最值得做成默认设置的,不是「自动挑一台在线的」,而是先记清楚:这件活允许在哪台跑。在线是状态,不是授权。
1
2
68
把上次让 AI vibe coding 的 TrailerTube 展示版代码公开了。
它也是一个 Agent 友好的静态站:不只是给人浏览,也把 Agent 当作使用者来设计。
仓库自带 80 部电影样本,不用申请 API key,本地就能跑。浏览、搜索、收藏、稍后观看都能试。
github.com/terryso/trailertu… · trailer.terryso.dev/
nitter.cf/Suchuanyi/status/21172…
Karpathy weekend-hacked HackerTube.
I followed the idea and built TrailerTube — ~a decade of movie trailers, one page.
trailer.terryso.dev/
1
2
114
你是不是也在 X 上存了很多书签,却很少回看?
我做了个「X 书签消化助手」:先把帖子原文存好,再定时挑出值得关注的内容,告诉你接下来能做什么。每次最多 3 条,没用的就不打扰。
想清理书签也可以,但必须先得到你的同意;没读全、拿不准的都留着。
x.ai/bot/su2T1C_TnKNjwu1ICft…
3
80
苹果刚开放 Siri AI 开发者预览:
Intent 管单步,Skills 管把几步串起来。
官方说明:developer.apple.com/news/?id…
2
147
给 Agent 用的 X 发帖 skill:ego-skills / x-post。
复用本机 Ego Lite 已登录会话,page.fetch 直发,不点 UI。短帖 CreateTweet,长文 CreateNoteTweet。可带图。
用法见附图 · github.com/terryso/ego-skill…
开源了一组基于 @ego_agent 的 Agent Skills:在 Ego Lite 里用纯脚本跑(session / page.fetch),吃本机登录态,不靠点 UI,比 Playwright 点页面快一截。
仓库:github.com/terryso/ego-skill…
已上 SkillHub / ClawHub(ego-*):
· Pinterest / Instagram Saved 下载
· 微信视频号发草稿、快手 / YouTube / 抖音发视频
· @teambition 测试用例管理(官方skill不支持用例管理)
装好 ego-browser 后可用 Claude / Cursor / Codex 等多端安装。欢迎试用、提 PR。
Karpathy weekend-hacked HackerTube.
I followed the idea and built TrailerTube — ~a decade of movie trailers, one page.
trailer.terryso.dev/
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
4
224
很多 Agent 浏览器还在一轮一轮点工具。
@ego_agent 反过来:脚本直接在本机 Ego Lite 里跑——点、填、抓页面连着做,继承你已登录的 cookie,任务在独立的 Agent Space,不抢你标签页,人和 Agent 互不影响
我开源了几个技能挂在这条链上:github.com/terryso/ego-skill…
Pinterest / IG Saved / 微信视频号发布,有兴趣可以拿来玩玩,成功率和效率比用普通 Chrome 浏览器高多了。
我之前发微信视频号一直搞不定的封面上传问题,用 ego lite 一次就能搞定。
4
93
马斯克说 serving Grok Bot 相对收入撑不住。这边立刻对上:Opus 路由变少、额度往上跳。不是模型变笨,是补贴先断。
以前聊 Personal Agent 付钱,还在想怎么让它帮你下单;现在是——钱没到位,很多能力你根本进不去。
2
38















