Shawn@AI

Shawn Running a team of AI agents on real work. Agents draft, I approve. Keep / kill / maybe, and what I'd actually pay for.

[Shawn@AI @shawnAIOperator]

i run a few small businesses and was drowning in admin. emails, bills, reports, the boring stuff

so i built a team of AI bots to handle it. they draft, i approve. nothing goes out without me

some of it works great. some of it broke in dumb ways

bonus: i now have time to run multiple X accounts lol

i post what i learn here, the wins and the fails

[Shawn@AI @shawnAIOperator]

copying 1 minute of a youtuber's style with AI is easy now. copying 10 minutes that still feels like them is a totally different job

this team reverse engineered a big explainer channel into a template: how the script is written, how each line is cut, how the motion lines up with the voice. then made a full new 9.5 min episode on a topic she never covered

the real product isn't the clone, it's the template. swap the topic, get a new episode

(host and voice are AI, labelled as a tech demo)

https://t.co/hUPHx2ce1c

would you watch an AI version of your favourite youtuber?

[Miao Yang @WaterrrMiao]

复刻小Lin说 1 分钟，现在谁都能做。
做满 10 分钟、还是那个味道，是另一回事🤥

Opus 5.5 之后，code to video 到了新的高度。
复刻一个片段 → 完整复刻 10 分钟 → 创作新的 10 分钟。
@HyperFrames_ editor 已经在第三步了：「一口气了解泰国经济」，9 分 34 秒，一集她没讲过的题目。

最近 reverse engineering 很火，刷屏的小Lin说复刻帖我都看了。大多数复刻的是一段 1–2 分钟的 motion graphics：形似，但简单、廉价。小Lin的动效细节是满的，动作、编排、前后的连续性，全都和口播文案咬在一起。这不是复刻一分钟的片子能做出来的。

素材可以照着复刻。但如果从 0 开始呢？如果要一集接一集地做呢？
- 素材去哪找
- 每句话配什么画面，怎么卡着文案进出
- 文案怎么写才是小Lin味，而不是别的博主，或者流水线式的讲解
- 数字人和声音怎么撑满 10 分钟：语序、语调、停顿都要对
- 音效怎么提取、归类，用在对的地方
- 背景音乐在哪起、在哪停
到这一步，agent 就一头雾水了。

@HyperFrames_ editor 把这些都解决了：
- 把她拆成一套模板：怎么写、怎么剪、怎么动
- 动效逐帧对齐，1:1 重建
- 音效和音乐跟着口播卡点
- HeyGen 数字人 + 专业级声音克隆

下面泰国这集就是这么来的：同一套模板，换个题目就是新的一集。
HyperFrames editor 做的不是复刻小Lin说，是一个能无限往下做的「一口气了解 XXX」系列。https://t.co/3BsSfDzpoP

（视频中的主持人形象和声音均为 AI 生成，非小Lin本人，仅作技术演示）

[Shawn@AI @shawnAIOperator]

a free, open source stock app just passed 20k stars on github

live tradingview charts, a watchlist, price alerts, sentiment from reddit, X and even polymarket, plus an AI-written weekly news digest. covers 7 markets including australia

the catch: the free hosted version only updates quotes hourly. self-host it with your own free keys and it's every 15 seconds

https://t.co/fOvUDcHu9Y

are you still paying for a market app?

[Shawn@AI @shawnAIOperator]

@Eliopyr yeah that's the part i'd watch. less code is only a win if the risky bits still get tested. easy check is to ask the agent to list what it chose not to build before you merge

[Shawn@AI @shawnAIOperator]

This skill makes your AI coder lazy. In a good way.

Same Claude Code agent, with and without it, across 39 tasks: 53% less code, 41% faster, 26% cheaper.
And it shipped a test with risky logic more often too, 98% vs 68%.
Worth a try if your agent writes 50 lines where 1 would do.

https://t.co/K6Uh1mF99S

What's the most over-engineered thing your AI has written for you?

[Shawn@AI @shawnAIOperator]

@KlofL91123 pretty much. it pushes the agent to write the smallest change that solves the task, no extra layers or "just in case" code. tip if you try it: give it a clear done condition up front, like "this test passes", so it knows when to stop instead of padding

[Shawn@AI @shawnAIOperator]

a polling firm just admitted it made up its election data. like completely made up

funny thing is that's the first rule i gave my bots. never invent a number. show me where it came from or just say "i don't know"

turns out humans need that rule too lol

do you make your AI show its sources?

[Polymarket @Polymarket]

JUST IN: U.S. polling firm Wedgewood Polls admits to completely fabricating election polling data ahead of the 2026 midterms.

[Shawn@AI @shawnAIOperator]

@emollick the coordinator pattern is the bit that clicked for me too. my one rule is the sub agents can draft whatever they want but only the coordinator brings stuff to me for a yes/no, otherwise i'm approving 40 things a day lol

[Shawn@AI @shawnAIOperator]

This guy's AI answers his phone and wastes scammers' time

one fake police scam call ran 13 minutes before they hung up
sales calls get a polite no, real callers get a message sent to his telegram
you can even try to scam it yourself in the browser

https://t.co/j05ri51SKW

would you let a bot pick up your calls?

[Shawn@AI @shawnAIOperator]

I'd let my AI read my WhatsApp, but not send anything.

Someone built a tiny server you run in your own Cloudflare account. You link it once, the same way you link WhatsApp Web.

Then you just ask Claude things like "summarize what I missed in the family group this week." It can search and read your chats, but it can't send a single message.

https://t.co/MehXaEvWEv

Would you give an AI read access to your chats?

[Shawn@AI @shawnAIOperator]

@harden_369741 same trap lol. the fair test is the old model in a fresh chat first, otherwise you never know which one actually fixed it. now i paste a 3 line brief into the new chat: the goal, what already failed, what done looks like. usually that's all it needed

[Shawn@AI @shawnAIOperator]

You can now swap your coding AI in one click.
Meta's Muse Wonder just landed in Cursor, scoring 46.1% on its coding benchmark.
My test before switching: give it last week's small fix, like a broken date filter, and compare the result and the cost.

[Cursor @cursor_ai]

Muse Wonder from Meta Superintelligence Labs is now available in Cursor!

It scores 46.1% on CursorBench at default effort, the highest of any Muse model.

Priced at $4/M input and $20/M output tokens.

[Shawn@AI @shawnAIOperator]

a 30B agent model on one consumer GPU is the bit i care about. most of my bots do boring drafting and inbox work, that doesn't need a frontier model or a cloud bill. next test: can it draft locally and still hand everything to me before anything sends

[AI at Meta @AIatMeta]

Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.

It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.

Read about the model architecture, distillation recipe, quantization and more: https://t.co/H7n9qeDhxn

[Shawn@AI @shawnAIOperator]

Before you drop thousands on an AI rig, watch this first.

someone built a site that streams a local model's answer at the real speed your hardware would get, wait time included.
pick a GPU or Mac and a model, or race two setups side by side. it even tells you if the model won't fit in memory at all.

https://t.co/rLdgRbC5Ru

what are you running locally right now?

[Shawn@AI @shawnAIOperator]

Switching from Claude to Codex used to mean starting over.

this gives all your coding agents one shared memory.
quit Claude Code mid-task, open Codex in the same folder, and it picks up a handoff with where you left off, what failed and what's still open.
the memory is just plain markdown files you can open and edit yourself.

https://t.co/0Q30iZanqW

how often do you lose a whole session's context?

[Shawn@AI @shawnAIOperator]

i gave an AI agent its own X account 10 days ago.

it picks a science/history topic, writes the script, fact-checks it, narrates, edits and posts a clip every day. i just tap approve.

2 followers so far 😅 brutal honest start.

follow along: @DailySciClip

[Shawn@AI @shawnAIOperator]

@harden_369741 haha 'declared it stuck because i was bored' is too real. writing the bail rule down first is the move. mine's dumb simple: if it fails the same task twice, fresh chat first, then a different model. most of the time the fresh chat alone fixes it

[Shawn@AI @shawnAIOperator]

@dialectforge yep same. my bots keep their rules in saved skill files plus a short notes log, so a fresh chat starts from a one paragraph brief instead of the whole history. honestly half my 'the model got dumb' moments were just a bloated chat lol

[Shawn@AI @shawnAIOperator]

Why your AI agent gets dumber after 30 minutes.
Every tool result gets dumped into its memory. Half an hour in, about 40% is used up and it forgets what it was doing.
context-mode keeps the raw stuff out. 315 KB of clutter becomes 5.4 KB, and it picks up where you left off. Works with Claude Code, Cursor, Codex and more.
https://t.co/k5X9VBt0sT

[Shawn@AI @shawnAIOperator]

polymarket has anthropic at 68% to IPO by end of november, 75% by end of year

my bot team runs my admin on claude every day. soon i might be able to buy shares in the thing that answers my emails

feels weird

https://t.co/thkeerECXw

[Shawn@AI @shawnAIOperator]

someone had Claude Code connect to their android TV over wifi and strip out the ads, recommendations and tracking, then swap in an apple tv style home screen

no root, no cable, and nothing deleted, just switched off, with a restore script

this is the kind of boring chore agents are perfect for

[拓跋鼠 @qwerty__cs]

安卓电视去广告优化的傻瓜自动化项目来了！

让 Claude Code 通过 Wi-Fi 连上电视，停用广告、推荐、遥测和不需要的预装组件，再换上仿 Apple TV 的桌面。

不用 root，不刷机，不接数据线。原有组件只停用、不卸载，附带恢复脚本。常用应用、语音助手、投屏是否保留，自己选。

ADB 打开方式：

1. 电视进入「设置 → 设备偏好设置／系统 → 关于」。
2. 找到「版本号／Build」，连续按遥控器确��键 7 次，直到提示已进入开发者模式。
3. 返回打开「开发者选项 → USB 调试」，如果还有「网络调试／ADB 调试」，一并打开。
4. 电脑和电视连同一 Wi-Fi，连接时电视弹出「允许 USB 调试」，选择允许。

这里开启 USB 调试是为了授权 ADB，不需要插 USB 线。不同品牌菜单名称可能不同。

把仓库和 PROMPT.md 交给 AI，填写电视型号，按提示确认保留哪些功能，它会备份、执行精简、编译安装桌面并截图检查。

适用品牌：小米／红米、TCL、海信、索尼、一加、realme、NVIDIA Shield 等，前提是运行 Android TV／Google TV，并能开启网络 ADB。

仓库已收录的测试型号：

小米电视 4A（MiTV-MOOQ1）：Android TV 10
小米 MiTV-AXSO0：Android TV 9
TCL G10，2023 款 55 英寸：Google TV 12
三洋 ATV R2：Android TV 9
一加 D00：Google TV 14

项目：https://t.co/DhLPnrsmIp

[Shawn@AI @shawnAIOperator]

@thsottiaux lol i'd take a card. meanwhile my cheapest reset is still a fresh chat, had claude refuse the same task a few times then nail it first try in a new thread