Ss7lyb

[Olivia Chowdhury @Oliviacoder1]

7. The Retention Audit

Act as an audience retention analyst specializing in online video performance. Review the provided video or script and identify where viewers are most likely to lose interest or drop off. Pinpoint specific timestamps, explain what weakens engagement, and suggest targeted fixes such as pattern interrupts, tighter cuts, B-roll inserts, or pacing adjustments to keep attention high. Also create three alternative high-energy hooks for the first 10 seconds designed to maximize curiosity and retention. [Upload video or paste script].

[Viktor Oddy @viktoroddy]

CLAUDE DESIGN FULL 1 HOUR COURSE

❤️‍🔥Just Recorded a Full Course on How To Build Interactive Animated Website with AI.

Bookmark this before you forget.

[老金 @freeman1266]

https://t.co/37szqr8QgO

[OneHopeA9 @onehopeA9]

https://t.co/AMMkoUFMBa

[Garry Tan @garrytan]

Did you want a GBrain but didn't want to set up your own OpenClaw

Clawputer is here to help you!

https://t.co/Nt0yn8AcUo

[Ss7lyb @ss7lyb]

@NFTCPS 这不是早就开源了吗？

[Ivan Burazin @ivanburazin]

After the Claude Code source code leak, a former PM extracted its multi-agent orchestration system into an open source model agnostic framework.

He studied the architecture, focused on the multi-agent orchestration layer (the coordinator that breaks goals into tasks, team system, message bus, task scheduler with dependency resolution), and reimplemented these patterns from scratch as a standalone open source framework without infringing on Anthropic's code.

The result is what @JackChen_x calls an "open-multi-agent." Unlike claude-agent-sdk, which spawns a CLI process per agent, this runs entirely in-process and can be deployed anywhere (serverless, Docker, CI/CD)

Check it out: https://t.co/w3XjnZEk92

[Om Patel @om_patel5]

THIS IS HOW A SENIOR ENGINEER ACTUALLY SCALES THEMSELVES WITH CLAUDE CODE

the biggest change with AI isn't coding faster. it's where you actually spend your time now.

more detailed prompts, more code review, more planning, less typing, etc.

here's the workflow:

this guy has been shipping code since the days of cgi and perl.

he uses a compound engineering plugin that runs 5 separate agents on every task.

one brainstorms, one plans the technical implementation, one executes, one reviews, one checks different verticals.

every step is documented in markdown files. it's slow and way more waiting. but the output quality is way higher because each agent is focused on one thing.

then the REAL multiplier is in git worktrees

if Claude Code made you 10x faster, worktrees multiplies that again depending on how many agents you can manage in parallel

his team runs 4-8 Claude Code sessions at the same time across different worktrees with each one working on a separate task.

the skill is managing multiple AI agents in parallel without losing track, that's the next evolution of engineering

[Ss7lyb @ss7lyb]

@lijigang Lecun 的世界模型就是这个思路

[Ss7lyb @ss7lyb]

@ousinu @wangray 我早就做完了

[Nagi Yan @naki2012]

《认知是什么？从信息中逆向还原结构的能力》

一篇很有意思的论文提出了一个简单却深刻的问题：

在有限算力的条件下，人能从信息中看出多少结构？

论文的答案是：结构识别是算力受限的。算力越强，能发现的结构越多；算力不足时，很多结构就会像随机噪声。换句话说，信息中的结构 ≠ 人能看到的结构。人看到的，只是其中一部分。

⸻

信息有结构的本质：可压缩性

如果一段信息存在规律，它就可以被压缩，比如 111111111111 可以压缩成 1 × 12，这说明它有结构。而一串看起来像 101011001101 的数据就很难压缩，意味着它更像“随机”。这对应的是 Kolmogorov Complexity（柯尔莫哥洛夫复杂度）：一个对��的复杂度 = 描述它的最短程序长度。所以识别结构，本质上是找出“更短的描述”。

⸻

同一世界，不同人看到完全不同的规律

论文的关键洞察是：结构是否“存在”，其实取决于观察者的计算能力。普通人看股市是涨跌波动，量化基金看到的是统计套利，物理学家看到的是系统动态。信息本身没变，只是不同算力下，世界的结构可见度不同。

⸻

人类并不是靠算力找结构

机器靠暴力计算搜索结构，人类却靠认知结构。比如看到序列 1 3 5 7 9，我们几乎立刻能看出是等差 +2，这不是因为我们计算快，而是我们脑中早就内建了这类结构模板。人类并不是用算力找结构，而是用结构找结构。

⸻

所以认知是什么？

认知就是：从信息中逆向还原结构的能力。同样一段信息，有人看到表象，有人看到规律，有人看到系统，有人看到抽象的结构动力学。信��没变，变化的是一个人能还原出多少层结构。

⸻

结构可见度的底层公式

结构可见度 = f（信息 × 搜索能力）。搜索能力的来源有两个维度：算力（机器）+ 认知结构（人类）。算力越强，能穷举更多可能。认知结构越深，能用更少资源定位到复杂结构。

⸻

认知差距的本质是结构还原能力差距

为什么很多人看世界混乱，有些人却能从中提炼出稳定的规律？因为他们在信息中能还原出不同深度的结构。有人停留在表象，有人能构建出完整的系统层次。这种差距，不是“聪明”与否，而是结构识别层级的差异。

⸻

AI 会极大地放大这件事

AI 的优势不在知识，而在算力。当算力提升，AI 就能识别出人类肉眼“看不到”的结构。这就是 AlphaFold 能解蛋白质结构、量化模型能捕捉市场微结构、大模型能学习语言模式的原因。很多“随机”，只是我们当前看不见。

⸻

结尾：认知是文明的核心变量

从这个角度看，认知不是信息的积累，而是结构的还原能力。世界是否有秩序，不只是它本身决定的，而是��的结构可见度决定的。你能看到的结构，决定你能理解的世界。未来文明的分层，很可能不再是信息差或财富差，而是结构可见度的差距。

⸻

📖 引用论文：“The Structure of Information Under Computational Constraints”

[meng shao @shao__meng]

Remotion Skills 制作视频完整过程和提示词

最近我的「AI 启蒙小伙伴」在小红书和视频号等平台用 Remotion Skills 制作了很多期动画视频，很多朋友留言想了解咱们的视频是怎么做的，尤其是视频号，春节期间关注增长了 4000 多，单篇视频最高观看也到了 8 万，转发 4 千多，今天就一并分享制作环境、过程和提示词。

制作环境：
Claude Code/OpenClaw/Codex/Cursor 等支持 Agent Skills 的 AI Agent，使用 'remotion-best-practices' 这个 Agent Skill，大家可以自行搜索，或让 AI Agent 自己找自己装。

视频核心结构：
Hook → Why → Process → Essence → Tips → End Card 每个场景 4-8 秒，控制在 1 分钟内。

制作过程：
在 AI Agent 中输入要求，请使用 remotion-best-practices 为以下内容生成动画视频，参考以下提示词 [...]，输出 MP4 文件。

视频提示词：
请按以下规范生成一条 3:4 竖版短视频：

分辨率1080×1440，30fps，时长15-20秒。背景音乐音量0.28~0.35，前1秒淡入，后1.5秒淡出。全片基底 #08090c，主色暖橙#f59e0b、辅色冷青#06b6d4，禁止蓝紫渐变与高饱和荧光色。

结构：Hook 25%、Content 55%、End Card 15%。Hook 要突出成本对比/核心数据，标题��80px，团队badge置顶，左右飞入形成碰撞感。

Content 每幕单观点，卡片左/右滑入 spring，progress bar 同步推进；使用 100-140 刚度与damping 18-22。

End Card 显示品牌名、头像旋转光环、三按钮（点赞/转发/关注）横向错位弹入，布局居中。

字体：英文标题 Oswald，中文大标题 Noto Serif SC（60px+），正文 Noto Sans SC（28px+）；居中 flex，左右内边距50-80px。

视觉：叠加 0.025 噪点、极淡网格线、60-100px 大光晕 blur（频率≤0.04）；关键元素加 willChange。

动画统一：入场 spring 交错 delay10-15帧，转场 fade 15帧且场景重叠；禁止 breathe scale、scan line、wipe。

最后保留 20-30 帧阅读停顿，结尾必须含 CTA（点赞/转发/关注）。

[Ss7lyb @ss7lyb]

@PanXing0827 @xiaohu 咋办呢

[凡人小北 @frxiaobei]

这篇内容给我看亢奋了。

真真实实的用 OpenClaw 当编排层，很多人还在纠结用 codex 还是 claude code，人家已经把它们都降级成工人了，自己只盯业务和合并 PR。

这就是我要给 AI 配 Mac mini 的原因之一。

我现在唯一的担心是我这次配置可能买低了。
不过也无所谓，到时候实在不行就直接换新机。
说不定三个月后我也会加入二手出机大军。🤣

[Jason Zhu @GoSailGlobal]

https://t.co/OXfzzWVB9C

[Muratcan Koylan @koylanai]

https://t.co/7kZ0k7nbxI

[一口新饭 @onenewbite]

https://t.co/9UyWBtBMdt

[Team9 @Team9_ai]

https://t.co/mEwwHMqYhh

[𝙩𝙮≃𝙛{𝕩}^A𝕀²·ℙarad𝕚g𝕞 @TaNGSoFT]

Skills解决了”能力模块化”，但没解决”身份连续性”。

openclaw把language agent的脚手架上下文管理一定程度上交给模型自己了！clawbot套上一层self-awareness，加上了cc+skills所没有的身份联系性！

Layer 4: Consciousness（意识层）
↑ https://t.co/IjWBTjCApH, https://t.co/b4Vb1ZU2cj
| "我是谁、我为什么存在"

Layer 3: Memory（记忆层）
↑ https://t.co/KaB4zCEOps, https://t.co/z3wl7F6CYW
| "我经历了什么、我在为谁工作"

Layer 2: Capability（能力层）
↑ Skills, Subagents, MCP connectors
| "我能做什么"

Layer 1: Execution（执行层）
↑ ReAct loop, tool use, verification
| "我怎么做事"

[Zac @PerceptualPeak]

FINALLY!!! I've SOLVED context transfer between a pre-compacted & post-compacted state in CLAWD BOT. It was a LONG and grueling day figuring this one out, but for all intents & purposes, I now have a completely seamless transition from pre-compacted & post-compacted session states. Here's some of the changes I made:

- I have a cron job that maintains a running memory file on an hourly basis; summarizing everything done in that hour long block & appends it to the running memory file for that day.

- The past 24 hours of hourly memory summaries get injected into the post-compacted summary along with the compacted summary itself.

- Clawdbot maintains a running .JSONL file of ALL conversation history (which does not get impacted by compaction). I've configured it to ALSO inject the 15 most recent user messages, 10 system messages, and 15 thinking blocks from the pre-compacted chat session log into the post-compacted context.

- Cron job established on a bi-hourly basis which spawns a sub-agent to scour the previous two hours of chat logs, extract, embed and store relevant learnings in a vector database (using nomic embedding model).

- Modified the Clawdbot source code to include a "User Prompt Submit" hook (like Claude code has). So anytime I submit a prompt to Clawdbot, it will (in a synchronous fashion) embed my prompt, find relevant memories in the vector database, and inject them alongside my prompt to Clawdbot before it even begins processing my prompt (delay time sub 300 ms).

These additions have completely changed the game for me. My Clawdbot is operating 10000x more efficiently, with literally ZERO noticeable knowledge loss.