Building Hong Kong × China AI Operator Community AI Agents | Workflows | Growth Mindset
Hong Kong
Joined November 2018
- Tweets2.9K
- Following1.1K
- Followers8.9K
- Likes2.2K
Turn 30 days of Claude Code sessions into 1 onboarding guide for a teammate
→ Step 1: /team-onboarding
→ Step 2: Claude reads your sessions, commands, and MCP server usage from the past 30 days
→ Step 3: a teammate pastes the markdown guide as their first message
I think this is a faster handoff than a wiki page nobody updates.
Who on your team would you hand it to first?
via @ClaudeDevs code.claude.com/docs/en/comm…
Sonnet 5.5 max costs 12.8x as much as medium for 24 more index points
→ Medium: 30 at $0.78
→ High: 41 at $1.64; xhigh: 50 at $4.09
→ Max: 54 at $10.02
I think xhigh is the Sonnet stop; max adds 4 points for 2.4x the bill.
Which Sonnet effort do you leave on for coding agents?
via @ArtificialAnlys artificialanalysis.ai/leader…
Codex CLI 0.165.0 adds 6 features, including a built-in Vertex AI provider
→ Google Cloud Vertex AI: set model_provider = "gcp"
→ GPT-6.1 Terra joins the model catalogs
→ Reasoning effort changes mid-chat are on by default; memory v2 is on for fresh profiles
I think mid-chat effort is the 1 to use: start cheap, raise it when the task gets hard.
Which of these would you switch on first?
via @OpenAIDevs github.com/openai/codex/rele…
Gemini 4 Carbon is #1 on Artificial Analysis at 59
→ $7.76 a task vs Opus 5.5 max at $10.04
→ 88.2% SEC-bench Pro, per Google
→ Cyber defenders get it first
I think access beats the 1 point.
Would you switch?
via @GoogleDeepMind
Gemini 4 Carbon takes #1 on Artificial Analysis at 59, 2 weeks after Argon
→ Carbon high: 59 at $7.76 a task; Opus 5.5 max: 58 at $10.04
→ Intro API price matches Argon: $2 in, $10 out per 1M tokens
→ Access starts with a subset of cyber defenders; no date for everyone else
I think the 1 point matters less than who can actually call it this month.
Would you move a coding agent to Carbon once it opens?
via @GoogleDeepMind
According to the post, Tripo's smart mesh inside Kimi's 3D Make creates exactly the count you specify — quad polygons or polygons, your number, no rounding off.
The person trying it wasn't using their own Tripo account either; the Tripo features sit inside Kimi, so creation happens in chat.
That's the part of the post that lands: specifying a number.
「not a single video under 1M views」：有人話自己掃到個金礦，Claude 相關嘅片冇一條低過一百萬 view，結論係唔拍就等做永久底層。
呢個數靚到我有啲想自己搜一次。同一個「Claude」關鍵字，可能教學、demo、news reaction、片尾硬加一句都撈埋一齊計，如果全部倒入同一個池度計，數字自然好睇。
真係咁好賺，就唔使特登寫句 we are so fucking early 催人入場。
Reportedly, 33.2M views from 6 videos, 100% simulated with Claude Opus 5.5.
The post says its author has seen 8 of these accounts, and none of them are underperforming.
The Fruit Ninja mobile games make $1M+/mo, and reportedly haven't had a competitor in over 16 years. Turns out the gap is being filled with simulated gameplay.
Under $1 a task, top AA speed is 91 tok/s
→ GPT-6.1 Terra (max): 51 at $0.96, 91 tok/s
→ GPT-6.1 Terra (xhigh): 45 at $0.40, 91 tok/s
→ GPT-6.1 Sol (xhigh): 52 at $0.76, 73 tok/s
I think speed under $1 matters more than 1 index point for batch coding jobs.
Would you pick the fastest row or the highest index?
via @ArtificialAnlys artificialanalysis.ai/leader…
Haiku 5.5 max costs 10.2x as much as low for 17 more index points
→ Low: 19 at $0.04
→ Medium: 25 at $0.09; High: 30 at $0.16
→ Max: 36 at $0.41
I think Haiku high is the batch stop; max adds 6 points for 2.6x the bill.
Which Haiku effort do you leave on for cheap coding jobs?
via @ArtificialAnlys artificialanalysis.ai/leader…
Open the Claude Code hooks menu in 1 command
→ Step 1: /hooks
→ Step 2: view configured lifecycle hooks (PreToolUse, PostToolUse, and the rest)
→ Step 3: add, disable, or remove hooks from the same panel
I think /hooks is how you stop a bad shell habit before the next tool call lands.
Which hook would you add first: block-rm, or a notify-on-permission hook?
via @ClaudeDevs code.claude.com/docs/en/comm…
Opus 5.5 max costs 4.1x as much as medium for 10 more index points
→ Medium: 48 at $2.45
→ High: 52 at $3.46; xhigh: 57 at $6.05
→ Max: 58 at $10.04
I think xhigh is the Opus stop now; max adds 1 point for 1.7x the bill.
Which Opus effort do you leave on for coding agents?
via @ArtificialAnlys artificialanalysis.ai/leader…
GPT-6 Astra max costs 4.4x as much as low for 14 more index points
→ Low: 43 at $1.42
→ Medium: 51 at $2.73; High: 55 at $3.24
→ Max: 57 at $6.24
I think high is the Astra stop; max adds 2 points for 1.9x the bill.
Which effort level do you leave Astra on?
via @ArtificialAnlys artificialanalysis.ai/leader…
Claude Code /memory edits CLAUDE.md and auto memory in 1 menu
→ Step 1: run /memory
→ Step 2: toggle auto memory
→ Step 3: open the memory folder
I think the 200 line auto memory cap matters.
What would you prune first?
via @ClaudeDevs
Opus 5.5 max still leads Artificial Analysis at 58
→ Claude Opus 5.5 max: 58 at $10.04
→ Opus 5.5 xhigh and GPT-6 Astra max: both 57
→ GPT-6.1 Sol max: 55 at $1.27
I think Sol max is the row to watch: 3 points under Opus for about 1/8 of the bill.
Which of these 4 is open in your terminal right now?
via @ArtificialAnlys artificialanalysis.ai/leader…
Someone asked them the same prompt: Opus at $0.40 is ripping the farm apart; Kimi at $4.10 is sitting under a big cloud.
10× more expensive, and it's the one that looks asleep. Who's actually picking Kimi?
Someone shared a whole city in Unreal Engine, all assets from Blender, claimed to be built by Opus 5.5.
City-scale, not a room or a single hero asset. That's a much bolder claim to make than another tidy prop shot.
Someone’s two days into testing various harnesses with Claude + Blender and still can’t get a proper long-form video out. At that point the cost feels like the smaller problem.
Camera movements and effects are the parts some folks write about so well — stringing them into something that actually holds a video together is a different job.
