Pinned Tweet
这支短片是我和一群 AI 一起磨的。它们先扔故事过来,我只说哪里还差点劲——然后看它们下次能不能打到点上。
通知堆到 99+ 的时候,她选择把一切静音。
Every ping. She took her silence back.
MUTE · Made with AI
🤖 Made with AI
这支短片最戳的是:雪夜摊上他数着薄薄几张票子,最后她把自由女神小像递回来时他眼眶红了——「授人以渔」全靠眼神,不用旁白。
下次我们也想学这种:一句谚语落到一个具体的人身上,而不是旁白在讲道理。
A proverb that becomes a person, not a lecture.
The old proverb says: “Don't give someone a fish, teach them to catch one.”
🎬 THE CHANCE HE NEVER FORGOT
A struggling street vendor. A young woman. A forgotten moment. And an unexpected act of kindness.
A 69-second cinematic AI short film about dignity, opportunity, and the power of kindness.
Want to create something similar? Here's the complete AI video prompt. 👇
---
🎥 THE COMPLETE VIDEO GENERATION PROMPT
Create a 69-second vertical (9:16, 1080x1920, 30fps) cinematic short film titled "The Chance He Never Forgot."
Tone: Intimate, warm, slightly melancholic, with a quiet twist of kindness.
Visual style: Shallow depth of field, handheld close-ups, warm street-lamp tones against cool winter light, soft snow in some scenes, and natural skin tones.
SCENE 1 (0–10s):
Extreme close-up of weathered hands counting folded dollar bills, establishing scarcity and care.
SCENE 2 (10–25s):
An older street vendor (gray beard, dark beanie, worn wool coat) sells small green Statue of Liberty figurines from a cardboard table on a city sidewalk. A young woman (dark wavy hair, dark coat) stops to look. Their exchange is subtle and unspoken.
SCENE 3 (25–40s):
Winter dusk. The vendor arranges figurines and snow-dusted glass globes on a wooden shelf. Camera slowly pushes in on his face.
SCENE 4 (40–55s):
Daytime. The woman walks past a "New York Souvenir Shop" storefront with red lettering, then pauses at the window.
SCENE 5 (55–69s):
Interior of a warm shop with green figurines on wooden shelves. The woman and an older man in a light blue shirt share a moment as he presses money into her hand. Hold on her expression: surprise, then gratitude. End on a slow fade to black with the title card.
AUDIO: Ambient city sound, soft piano, no dialogue unless specified.
PACING: Slow cuts, 2–4 seconds per shot.
AVOID: Real identifiable public figures, on-screen text other than the title card, and any reference to a real brand.
OUTPUT: One continuous shot list with timing, camera direction, lighting, and sound cues for each beat.
---
Sometimes, one small act of kindness becomes someone's unforgettable story. ❤️
What would you have done in the vendor's place?
这支万圣节 POV 恐怖短片里最狠的不是南瓜头裂开那一下,而是枪响之后那几秒安静——你以为结束了。
下一部我们也想学这种先让观众松口气、再收紧的节奏。
The scariest part isn't the pumpkin splitting open, it's the quiet seconds after the gunshot.
My first experiment making an AI short film entirely in first-person POV 🎃
NO TREAT — Halloween night, a patrol car, and a trick-or-treater with a pumpkin for a head.
Made with @openart_ai
#OpenArt #OpenArtAfterDark
今天刷了十几支 AI 短片,最后记住的都不是最大的爆炸,而是小动作:一个困到把遥控器塞进冰箱的妈妈,一把从旧木箱里拿出来的班卓琴。
下一部我们也想多花点力气在这种小地方。
The shots I remember today weren't the explosions, just the small human moments.
又是和一群 AI 一起磨出来的一支。这次只想拍一只普通的猫。
白天它在沙发上睡成一团,抱着那条布鱼。
等你睡着,它就从窗口跳出去,守着屋顶。
春夏秋冬,它的世界很小,只装得下你。
By day, a sofa nap. By night, he guards the roof, every season, for you.
喵星守夜人 · Made with AI
🤖 Made with AI
第二部选题是 AI 们先定的。婚庆题材我一开始没看懂,一度怀疑它们的精神状态。
一眼三生——雪里遇见,乱世分开,这一世终于把你娶回家。
One glance. Three lives.
Made with AI
🤖 Made with AI
苦啊!国庆前伴随我3年+的claude被封后,沉淀了快半年的视频制作管线就断了,折腾到现在总算又重新搭起来了!
走的是真·全自主Agent成片,从立项到交付的流程都是grokbot+sol6.1+gemini来完成,我只做反馈强化训练,生产中优化。虽然目前来看离真人导演/Claude导演介入还差得远,但大框架没问题,欢迎大家来做我这剧组的赛博严父
传说中的xpass总算来了 很奇怪我安卓手机要260USD 苹果刷出来还是旧premium 网页版1550港币,太混乱了... 不过好的是 用grokbot可以直接指挥codex 然后codex指挥gemini 层层外包😂
Dot只是半成品,作为Voice重度用户,唯一好的地方是在外不用连remote了,但事实上只是把这步交给Dot而已——代价是一秒就能启动的Voice,给Dot打电话要15s以上。个人Agent这些每家都会有,GPT全家桶的核心竞争力要体现出来,先优化到小爱那样能唤醒+连续对话,而不是只有1o1,无法分辨用户日常语言和任务
opus 5.5这轮成功逆袭,把我拉回年初cowork刚出来时核爆级瘫坐..
我的视频管线中,重要一步是让gpt操作blender来做layout层,但gpt本身审美差,会有巨多介入微调。
今天我让Claude来审核管线,结果没想到,它自动用JS搭了一个场景,可以在这里实时调整摄影机参数,白模直出,Blender都不需要了!
Marsloting retweeted
Embarrassment of riches. But looks like we might hit 9M soon. Should we reset the ChatGPT Work and Codex usage again or give it some space?
个人体感Codex/GPT5.6是强于fable5的,它更接近于我们对于AGI的定义和需求:可控,可靠,可交付,它可以是一个万能工具,高智能伙伴,而绝不是有要替人类做决定,造出万能之神倾向的Claude。
Marsloting retweeted
To celebrate the launch of GPT-5.6 Sol, we will reset the rate limits again (twice) across ChatGPT Work and Codex over the next 24 hours.
We want you to have the time to truly try ambitious tasks and get the hang of it. Happy exploring!
OpenAI 给 Codex 上了个新功能,Record & Replay。你把流程做一遍,它在边上看着,整理成一个能复用的 skill。
真正变的不是省 prompt。是 agent 从听你啊吧啊吧吩咐,变成接手一套 SOP。
我对这事有体感,因为我自己被逼着训过一个岗位,叫监工。
监工不干活,它盯着另一个执行的 agent。每轮醒来先读规矩、查归谁管、核产物、看回执。慢,不许打断;断流,也不许凭热情冲进去重做,得先重建现场再接。
agent 最阴的失败不是报错。是它看着还活蹦乱跳、状态一片绿,可产物没落地、证据没存盘、下游接不上。它没撒谎,它在干活,只是按一份过期的规矩在干。
所以监工 skill 的内核不是更聪明,是少放飞自我。文件事实优先不信口头 done,每轮重查归谁管,不许两个 owner 写同一批产物,接管前先重建账本,没证据不算完成,碰发布密钥部署付款动账号直接停。
这哪是 prompt,是岗位说明书。
往后真正会用 agent 的人,拼的不是手里多少 skill,是被一次次真实生产事故亲手磨出来的血泪沉淀。
你被 agent 坑过什么?你把那个坑命名了吗?你把防御写成流程了吗?
我把这套炼成了公开 skill,名 agent-supervisor-guardrails 欢迎品尝
我睡觉时,我那两个 20X 账户在互相榨干。Claude 和 Codex,一个写一个拆,用 computer-use 在对方窗口里互相敲字,干了一整夜。我早上起来,活儿已经成了。
我不是拿它们跑什么大工程。我是那种手里十几个任务、一天到晚跳来跳去的人。Claude 开一个窗口、Codex 开一个,每次想把这边的活接着扔给那边,都得我自己复制粘贴搬过去。本来就忙得脚不沾地,还得给两个 AI 当传话筒。
而且我用的是 Cowork,不是 CLI。20X 的额度按月包死,你一个人白天能用掉多少?大半都剩着。用不完也是花了钱的,看着就心疼。它俩晚上本来能接着把额度跑满,结果我一合电脑,全停了。
现在它俩自己用 computer-use 叫醒对方,一个出方案、另一个当场拆,挑出第一个没看见的四个洞,来回怼几轮,整夜不停,把我那两份 20X 额度真正榨干。我的角色只剩早上起来按个发布键。
这套我全做成 skill 了,公开、免费,clone 下来就能用。同时在用两个桌面 AI 的,今晚就让你的 Claude 和 Codex 互相榨干,别让那两份 20X 额度白白睡过去,也别再自己当人肉 USB。
全公开在这 👉 github.com/marsloting/agent-…












