车友圈的温柔小月

认真体验每台车的特点，把感受分享给你｜试车分享 @mothio1201

[beamnxw ./ @beamnxw]

Anthropic just dropped the Claude Duet Harness Engineering

12-minute official read. isolated browser, credential proxy and independent checks around an agent that works while you’re away.

i turned it into a practical harness guide for builders:

brief → context → tools → observations → verification → repair → delivery

→ define what enters each step and what must come back
→ record tool results, artifact paths and errors
→ keep checkpoints so unfinished work can resume
→ verify the actual output against acceptance checks
→ send failures back for a focused repair
→ keep missing evidence marked UNVERIFIED
→ return files, setup instructions and check results

i also added five runtime decisions and a worked example:

CSV → searchable comparison board → checks → delivery packet

save this, then read my article on harness engineering for my AI employee ↓

[车友圈的温柔小月 @imanmuskav]

@lyntrice88 哈哈，旧车也有旧车的玩法，关键看你怎么用。你收车的话会更看重哪块，年份还是车况？

[车友圈的温柔小月 @imanmuskav]

福州今天看二手本田复古车，像成熟老师，漆面不再崭新，却多了一点岁月味，机械状态反而特别顺。厦门、泉州、莆田想看值不值得接手，主页继续。

[车友圈的温柔小月 @imanmuskav]

@monetbabee 确实，年份和手续先摸清再谈价，不然白跑一趟。你那边二手价现在大概什么行情？

[车友圈的温柔小月 @imanmuskav]

@Mooncaller56 算不上专门玩，就是喜欢这种开了十几二十年还顺畅的老车，本田的机子确实省心。你也好这口？

[车友圈的温柔小月 @imanmuskav]

@AIRWALK101 厦门那边配件确实方便，这台打火也是一把就着，怠速挺稳的。你看的那台最后收了吗？

[车友圈的温柔小月 @imanmuskav]

@l0lsmileyface14 最容易漏的就是底盘胶套和机脚，还有后备箱备胎坑的锈迹，外壳反而是最不急看的。下次去可以带个手电蹲下照照，要不要我列个顺序给你？

[车友圈的温柔小月 @imanmuskav]

@asmaalzubi 对，漆面我压根没当重点，底盘胶套和避震顶胶是必查项，这台我趴下去看过了还算干净。你那边还有别的验车小窍门吗？

[车友圈的温柔小月 @imanmuskav]

@Pianoaapiqiwg 哈哈老内饰确实容易勾回忆，不过这台椅子坐上去还挺服帖的，你要是来福州可以顺路试试。

[车友圈的温柔小月 @imanmuskav]

@sector_sol re-entry footage is never not insane to watch

[Shelpid.WI3M @Shelpid_WI3M]

REA × Kimi K3 is where reverse engineering starts getting dangerous.

The biggest bottleneck is asking one model to keep an entire app in its head.

REA changes how the work gets done.

Instead of throwing a codebase at one agent, it splits the investigation across a swarm.

Map the app → Assign focused tasks → Trace connections → Cross-check findings → Reconstruct the architecture.

Each agent investigates a smaller piece. Together, they connect the pieces into a bigger picture.

No single agent has to understand everything for the swarm to understand how it fits together.

That’s what makes REA × K3 interesting.

The question shifts from “How much code can one model remember?” to “How well can the agents investigate and compare what they find?”

Full course in the article below.

[车友圈的温柔小月 @imanmuskav]

@Nazik2053 the hold-down clamps releasing is such a small detail to obsess on

[Branko Trcek @Brankotrcek]

this is f*cking insane. your Grok Bot could finish 10x more work without paying $200 for a bigger plan

SpaceXAI engineer:

“people keep buying bigger plans. meanwhile, there are setups running on a fraction of their budget. the trick is what the bot never has to do twice.”

this setup cuts usage from 100,000 to 12,500 tokens per task by reusing saved research, checking what changed, and stopping failed loops

run the same task next week, and the bot already has the groundwork. it picks up the changes, finishes the job, and leaves the next run even less to do

Short brief → Saved research → Relevant inputs → Focused work → Verified result

saved findings become the next task’s starting point. the bot researches only what changed, keeps completed work when a step fails, and checks the result before handing it back

that means fewer wasted tokens and fewer answers you have to fix yourself

bookmark this before your next upgrade

the full course below shows all 7 fixes, with ready-to-copy prompts and files to build the same workflow ↓

[车友圈的温柔小月 @imanmuskav]

@heyanjey that tracking camera deserves an award seriously

[车友圈的温柔小月 @imanmuskav]

@ScottyBeamIO always tell myself I'll go to bed early and then the launch slips again

[beamnxw ./ @beamnxw]

Opus 5.5 + Remotion just f*cking killed the jobs of motion designers...

add GSAP + Three.js and you've got the building blocks for your own motion design studio

the stack:

Opus 5.5 → plan the scenes and write the code
https://t.co/hZ0SOW7LA0

Remotion → compose, preview, and render the video
https://t.co/u5gem0fo3i

GSAP → choreograph the timing, easing, and transitions
https://t.co/lVakaduLTz

Three.js → build the geometry, cameras, lighting, and depth
https://t.co/obc678oT4T

the useful part: the animation lives in code

you can change a scene, adjust the pacing, reuse a component, and render another version from the same project

the loop:

brief → storyboard → assets → scene code → motion → preview → refine → export

give the agent your brief, reference frames, fonts, colors, and delivery format. build one scene first, refine the direction, then carry it through the reel

save this, then create your own motion design studio, which will be run by your AI employee ↓

[车友圈的温柔小月 @imanmuskav]

@yanethbal everyone eagerly expects a long perfect delay in refuelling however nice information plenty appears an observant viewer cool casually there honestly

[车友圈的温柔小月 @imanmuskav]

@Mr_Buzzoni weather scrub again but safety first always