JexLau

Cat mom of 2, 7 years. Co-founder of @theCatellect, helping cat parents better understand the little beings they love. Shipping AI hardware, learning in public.

[JexLau @JexLau]

@AaronxShepherd @ProsperPathway BUILD

[JexLau @JexLau]

@op7418 我直接做产品的营销渲染图，blender + threejs

[JexLau @JexLau]

@op7418 别的不说，用BPY控制Blender比视频模型生成可控性要强得多

[JexLau @JexLau]

这条推发出去半小时后，Codex 团队的 Eric 把同一个判断写成了完整指南。

他给出的建议很具体：精简 Skill 描述、用渐进式披露节省上下文、清理 AGENTS.md 里的过度约束，并重新定义 Astra 的执行边界和完成标准。

模型在升级，围绕模型搭建的系统也需要同步进化。

https://t.co/kcTfhyB9KH

[JexLau @JexLau]

Astra 上线后，我准备先做一次 Agent 系统大扫除。

这段时间，我在 Claude、Codex 和 Hermes 之间积累了大量项目指令、Skills、Prompt、长期记忆和自动化任务。它们解决过真实问题，也记录了旧模型犯过的错。但模型能力升级以后，旧规则的价值会发生变化：有些仍是必要边界，有些已经被模型能力覆盖，有些互相冲突，还有些只是某次失败留下的临时补丁。

我的做法是让 Astra 直接审查完整工作流：

1. 找出跨会话反复出现的 Prompt，把稳定流程沉淀成 Skill；
2. 找出人工重复执行的步骤，把它们改成脚本、集成或 cron job；
3. 检查项目指令和长期记忆，合并重复内容，修正冲突；
4. ��看经常中断的任务，定位卡住 Agent 的文件、权限、测试和上下文缺口；
5. 用真实项目重跑回归，依据失败案例决定哪些规则继续保留。

模型越强，Harness 应该越精准。真正有价值的 Agent 系统，会把每一次模型升级转化成系统升级。

[JexLau @JexLau]

https://t.co/hvkxg6SUy4

[JexLau @JexLau]

今天不用工作流，直接重构，几句话直出，GPT6 把 Catellect 从「9 块 9 塑料玩具感」里捞了出来。
之前整体的塑料感、灯罩发黑、出粮口被柔光吃掉、��色太淡、项圈灰扑扑……
细节都不用我提，也不用Skill，Astra是真离谱，这是对比：

[JexLau @JexLau]

@zengbozb 对，还有平时的工作流可以重构，每一次模型能力进化后，都要推倒重来。

[JexLau @JexLau]

Astra 上线后，我准备先做一次 Agent 系统大扫除。

这段时间，我在 Claude、Codex 和 Hermes 之间积累了大量项目指令、Skills、Prompt、长期记忆���自动化任务。它们解决过真实问题，也记录了旧模型犯过的错。但模型能力升级以后，旧规则的价值会发生变化：有些仍是必要边界，有些已经被模型能力覆盖，有些互相冲突，还有些只是某次失败留下的临时补丁。

我的做法是让 Astra 直接审查完整工作流：

1. 找出跨会话反复出现的 Prompt，把稳定流程沉淀成 Skill；
2. 找出人工重复执行的步骤，把它们改成脚本、集成或 cron job；
3. 检查项目指令和长期记忆，合并重复内容，修正冲突；
4. 回看经常中断的任务，定位卡住 Agent 的文件、权限、测试和上下文缺口；
5. 用真实项目重跑回归，依据失败案例决定哪些规则继续保留。

模型越强，Harness 应该越精准。真正有价值的 Agent 系统，会把每一次模型升级转化成系统升级。

[JexLau @JexLau]

@xiaomabosn 我之前产品转3D的疑难杂症问题被解决得非常好！

[JexLau @JexLau]

@op7418 我在grok bot里装了另一台grok bot 🤪

[JexLau @JexLau]

@calvinobai 太好用了

[JexLau @JexLau]

@Claudeapicom 太好了 我的陈年旧bug都可以修好了

[JexLau @JexLau]

Astra 全量开放。Full Reset 再送一张。一共3张！
Tibo 这两天的节奏太夸张了，这个周末，Codex 要跑疯了！

[JexLau @JexLau]

Every new model should audit the exhaust from the last one: repeated prompts, manual glue, failed handoffs, and workflows that worked. Your session history is the training set for a better personal agent harness.

[Shann³ @shannholmberg]

we got GPT-6 Astra, here are the first 3 things to run with it

GPT-6 Astra scored 32% higher than Fable 5.1, so its officially the best model out there

here are the 3 most important things you can do right away:

1. review how you work

give Astra your session and project history across Claude, Codex, Hermes and any other agent you use

ask it to find:

> prompts you repeat across sessions
> manual steps that should become scripts or integrations
> repeatable workflows that should become skills
> corrections that belong in project instructions
> scheduled work that should become cron jobs
> files or steps where sessions repeatedly stop

2. audit your second brain

let it inspect how knowledge enters, gets structured, becomes searchable and gets used by your agents

have it review:

> folder structure and data schemas
> duplicate or conflicting knowledge
> retrieval gaps
> cron jobs and ingestion workflows
> how sessions and research get saved
> files and data that you and your agents never use

3. review your evergreen projects

send it your most important projects, especially the ones earlier models could not get working properly

have Astra inspect the project files, run the existing tests, trace failed workflows and write a prioritized plan with the files to change and the checks each change must pass

write every approved improvement back into the project, skill library or company brain

use Astra for planning and review, then hand the defined execution plan to cheaper models

[Shann³ @shannholmberg]

we got GPT-6 Astra, here are the first 3 things to run with it

GPT-6 Astra scored 32% higher than Fable 5.1, so its officially the best model out there

here are the 3 most important things you can do right away:

1. review how you work

give Astra your session and project history across Claude, Codex, Hermes and any other agent you use

ask it to find:

> prompts you repeat across sessions
> manual steps that should become scripts or integrations
> repeatable workflows that should become skills
> corrections that belong in project instructions
> scheduled work that should become cron jobs
> files or steps where sessions repeatedly stop

2. audit your second brain

let it inspect how knowledge enters, gets structured, becomes searchable and gets used by your agents

have it review:

> folder structure and data schemas
> duplicate or conflicting knowledge
> retrieval gaps
> cron jobs and ingestion workflows
> how sessions and research get saved
> files and data that you and your agents never use

3. review your evergreen projects

send it your most important projects, especially the ones earlier models could not get working properly

have Astra inspect the project files, run the existing tests, trace failed workflows and write a prioritized plan with the files to change and the checks each change must pass

write every approved improvement back into the project, skill library or company brain

use Astra for planning and review, then hand the defined execution plan to cheaper models

[JexLau @JexLau]

Grok hit its usage limit just as GPT-6 Astra arrived. Perfect timing: the old king ran out of tokens, and the new one walked in. AI model season never sleeps.

[JexLau @JexLau]

不要把 grok bot 当聊天机器人使用，每个机器人都拥有一个不会消失的独立计算机。文���会一直存在。浏览器会一直存在。应用程序登录状态会一直存在。第一步：build a Chief of Staff before you build anything else.

[CyrilXBT @cyrilXBT]

https://t.co/suTKWm4S47

[JexLau @JexLau]

@antfuzh @simon_he1995 恭喜🎉！

[JexLau @JexLau]

我靠今天把 bot 套进 bot 里了，有点上头

[JexLau @JexLau]

@loki_yan_seo @willlhhh 前排搬好小板凳