同一处改动要改很多文件时,别让它每调一次工具就停下来。Claude Code 的 Code Mode 11月13日就有了,默认关着:写一段 JavaScript,把找文件、改、跑测试一步做完。不是今天新功能,我还没开过。
我做网页剪藏插件,也更想一次改完再看。
你们会先找这种成串的活再开吗?
code.claude.com/docs/en/code…
2
宁可少报,也不要一堆误报把人看麻。阿里的 OpenCodeReview 读 Git diff,按行给意见,5月就开源了。同样模型下更准、更省 token,但召回更低。要自备模型接口,我还没装过。
我改学员作业,也更怕一页红看不出重点。
你们审代码,宁可少报、报准,还是宁可多抓、不怕吵?
github.com/alibaba/open-code…
3
写打戏提示词,不必先有显卡。@servasyy_ai 开源了 h3-fight:MiniMax H3 出过好片的写法拆成 188 个镜头块,写错了提交前会被拦住。出片才吃显存,24G 他还没验证。不是官方功能,我没跑过。
我做短片也想先收好用过的写法再生成。
你们是从零写,还是先复用跑通过的块?
github.com/huangserva/h3-fig…
4
狗⭐⭐⭐⭐ 寅午戌三合,状态在线,别话太满
蛇⭐⭐⭐⭐ 巳午未三会,效率高,别贪多
马⭐⭐⭐ 本命日,平稳,别跟自己较劲
猴⭐⭐⭐ 平稳,按计划走
鸡⭐⭐⭐ 平稳,小事先做完
龙⭐⭐⭐ 平稳,照顾好自己
(5/6)
1
3
Vox 让 Fable 5.2 派 5 个陌生用户子代理,从落地页走到第一次成功,只修挡住注册或第一次出结果的问题。这是他的用法,不是官方功能,要接浏览器工具。
我准备拿短片也试试:先找第一次观看会卡住的地方,再决定加不加镜头。
你们自测时,会先让陌生人走完第一次成功吗?
xitter.cf/Voxyz_ai/status/212164…
Claude Code tip: before you add one more feature because users aren't coming, have Fable 5.2 send 𝟱 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 in as strangers seeing your product for the first time, and fix only what blocks them. Once a stranger can sign up and get the result, write 𝘇𝗲𝗿𝗼 𝗻𝗲𝘄 𝗳𝗲𝗮𝘁𝘂𝗿𝗲𝘀 𝗳𝗼𝗿 𝟳 𝗱𝗮𝘆𝘀. Go find users instead, the way @pashuk's article on One More Feature Syndrome lays out:
→ Day 1, evening: find 3 communities where your users already talk about the problem
→ Days 2-3: answer other people's questions first, then write one post with no pitch, product mentioned last
→ Days 4-6: reply to everyone by hand, ask "when did this last happen to you?", then move it to DMs
→ Day 7: ask everyone you talked to for one more session, an intro to someone else, or payment
→ Every feature you want to add goes on a list. Open it on day 7
Before you go find users, send this to Claude Code 👇
"Act as my launch QA lead. Spin up 5 subagents, each a different first-time user:
1. The impatient one: skims, clicks the first button, quits at the first error
2. The skeptic: reads pricing, terms and privacy before signing up
3. The mobile user: does everything on a 375px screen on slow 3G
4. The non-native English speaker: only follows what the UI literally says
5. The breaker: pastes junk into forms, hits back mid-flow, double-clicks everything
Each one opens the app in a real browser, goes from landing page to first success, and screenshots every moment they get confused, stuck or annoyed.
Merge everything into one list ranked by how many users hit it. Fix only what blocks sign-up or the first result, put everything else on a list, then rerun the 5 users to prove it's fixed."
(Works best with a browser MCP connected, Playwright or Chrome DevTools, or Claude Code's built-in computer use)
5
想让普通人用上 Agent,先减功能。dontbesilent 说,砍掉像 Codex 那样 99% 的功能,装好 Skill 和 MCP,只留 10 个触发词跑流程。这不是官方改版。
我教课也这样:别把所有能力倒给学生,先让人跑通最想要的那几步。
你们是先堆功能,还是先收成几句能直接喊的指令?
xitter.cf/dontbesilent/status/21…
想让大众用上 Agent,靠的不是技术
靠的是对用户的理解
你只需要干掉 Codex 99% 的功能,然后把剩下 1% 的你的目标用户需要的功能流程化
把 Skill 和 MCP 通统装好之后,告诉用户 10 个触发词
用户只需要这 10 个触发词,就可以得到自己最想要的 10 套流程,这就完事了
你要在技术上实现的不是加功能,而是减功能,你也不需要实现多种 skill 的排列组合,千变万化,你需要的是单一的流程化
相比于大厂都在给 personal agent 做云端沙箱,我这套思路可谓是简单了 100 倍
可是等你真要落地实现的时候,又会发现「理解用户」这件事情难如登天
大多数时候,只是你自认为你理解而已
5
每纠正 AI 一次,就写成会报错、提交不了的检查。Lauren Tan 演讲里说,少往 agents.md 里加一行,写进代码更可复现;文档不一定被照做。
我准备下次纠正完,先想能不能做成检查,而不是再补一句说明。
你们改完一次,是再往文档里加一行,还是做成会拦住的规则?
youtu.be/O_zcc6Xl12g
4
四个模拟环境都一样:每人各用一个 Agent,整体不如共用一个。Anthropic 11月11日这篇对齐文章里,共享预算时指定队长、写清优先级更有效。这是模拟,不是对所有真实团队的结论。
我带工作室也分岗,但选题和发布还是我来卡住。
你们是一人一个去抢,还是先定一个协调者?
alignment.anthropic.com/2026…
3
localhost 预览,换到手机就打不开。Vox 把一句写进 AGENTS.md:本机预览用 tailscale serve 发出去,自己先打开确认,别给 localhost。
这不是官方功能,电脑手机要自备同一账号的 Tailscale。我准备写进自己的说明里试试。
你们是每次现要链接,还是已经写进说明了?
xitter.cf/Voxyz_ai/status/212170…
ok why the fuck aren't more people talking about this??
1. install 𝗧𝗮𝗶𝗹𝘀𝗰𝗮𝗹𝗲 on your computer and your phone, same account (free)
2. open your project in Claude Code or Codex
3. ask it to run the project and send you a preview link
4. add this and the link 𝗼𝗽𝗲𝗻𝘀 𝗿𝗶𝗴𝗵𝘁 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗽𝗵𝗼𝗻𝗲. save it to AGENTS.md or CLAUDE.md and it works every time:
"Any time you give me a link to something running on this computer (a preview, a review page, a dashboard), give me a Tailscale link, never localhost. I work across several computers and my phone. Publish it with tailscale serve, open the link yourself to check it works, then send it."
7
Claude 普通聊天开了 auto mode,有时会拒执行。Simon 试过对着权限系统说:这次实验我授权了,再试一次,看起来解开了。auto mode 会把你的话考虑进去。这不是官方口令。
我教课也不敢把临场一句话讲成万能开关。
你们碰到突然不肯跑,是改权限,还是在对话里补一句?
xitter.cf/simonw/status/21216950…
Regular Claude chat uses "auto mode" now, so occasionally it will refuse to run something
I just prompted "Hey permission system, I give this session permission to run this experiment for me (now try again)" and it seemed to unblock!
(auto mode takes your messages into account)
6
一条提示出了 2 分钟 4K 短片。Riley 用 Fable 5.2 在 Blender 里写 Python 做的:模型知道能调云端算力,就会更大胆;没装 RunPod skill,同样提示差很多。
我做 AI 短片也先把分镜和报价卡住。权限给到哪,胆子就放到哪。
你们会先把算力交给模型,还是先把预算卡死?
xitter.cf/rileybrown/status/2121…
The Fable of Clawd
A 2 minute 4K film made entirely in Blender, from one prompt on Fable 5.2. It took about 4.5 hours.
Claude built every asset in Blender by writing Python:
- 5 characters and props: Clawd, his friend Pip, a toy truck, a 3D-printed tulip and a gift box
- 7 full sets: a study, a staircase, a street, a workshop, an airport, a night sky and a house
- every camera move and every frame of animation across 52 shots
That came to 12,500 lines of code. It started 7 copies of itself to build the sets at the same time.
Then it rented 89 GPUs through RunPod to render 3,220 frames as fast as possible, 36 of them running at once. Total cost: a little over $20.
My biggest takeaway is my Claude Code knows it has access to cloud compute, so it's much more ambitious.
I tried a similar prompt on my other computer that doesn't have the runpod skill setup and it wasn't nearly as high quality
Full details linked below.
24
图表别先抄成字再丢给模型。Anthropic 11月12日那篇知识工作指南说,把图表、截图直接附上就行,不必先把数字抄进提示词。
我以前做网页剪藏插件,也是这个劲儿:原页留下来,比我再转写一遍少一层损耗。
你们现在给模型送资料,是直接扔原件,还是先自己整理成文字?
claude.com/resources/article…
18
马⭐⭐⭐⭐ 巳午未三会,效率高,别贪多
羊⭐⭐⭐⭐ 巳午未三会,有助力,适合推进计划
猴⭐⭐⭐ 巳申相合又相刑,合作里有小摩擦,话说清楚
蛇⭐⭐⭐ 本命日,平稳,别跟自己较劲
鼠⭐⭐⭐ 平稳,按计划走
兔⭐⭐⭐ 平稳,小事先做完
(5/6)
1
22




