LegolasWang

An extraordinary Individual. Appreciate well-thought simplicity. Always deliver the best. Focus on creating things that enhance people’s life.

[LegolasWang @LegolasWang]

🧵 Opus 4.7 is a fundamentally different base model. The vibes have shifted — feels less conversational, more GPT-like. But here's the thing: it's not worse. It demands a completely new way of working with AI.

1/ With 4.6, you could still "chat" your way through tasks. Opus 4.7 punishes that. It over-asks, over-qualifies, gets verbose. That's not a bug — it's demanding precision from YOU.

2/ The image understanding and generation capabilities are a massive leap. Genuinely on par with GPT-Image-2 level surprises.

3/ The real shift: stop writing Skills. Start writing Agents.

Skills = rigid workflows, step-by-step instructions
Agents = closed-loop systems that think, verify, iterate, and deliver

4/ Skills still matter — but only as tools INSIDE your Agent. The mistake is treating your entire prompt as a Skill. That's why 4.7 feels "weird" to people — they're using 4.6 patterns on a 4.7 model.

5/ Claude Design is clearly built on 4.7's new capabilities. And the pattern is obvious: if your Agent can't nail it in one shot, just restart the conversation. Multi-turn debugging is mostly dead.

6/ Biggest takeaway:

❌ Don't let Skills constrain your Agent
✅ YOUR job is defining the Agent — goals, acceptance criteria, the loop
✅ Skills just provide guardrails
✅ Codify experience into Agents, not Workflow Skills

The leap from 4.6 → 4.7 isn't incremental. It's a paradigm shift in how you should structure your AI workflows.

[LegolasWang @LegolasWang]

Walking is underrated. Today I wandered by the river, caught the cherry blossoms at their best, and listened to the first lecture of Stanford CS153p right after it dropped.

The lineup this year is stacked: Jensen Huang, Ben, and Garry. It starts with something simple but important: take care of yourself, and be ready for big change. Even using AI for the assignments is allowed — as it should be.

This year, we’ll be sharing more through our own products, while also opening a new line to help more individuals and businesses. The best discoveries shouldn’t stay private. We want to bring the new experiences and methods we’ve found to more people ready to grow.

[Garry Tan @garrytan]

Lots of engineers think AI codegen is only good enough to do little bug fixes here and there

If you tell them you can ship 15k LOC of ai gen code to prod they think you have lost your mind.

It’s so so early. And also those engineers are living in 2025.

[LegolasWang @LegolasWang]

Flow state is basically a context window.

When the context is fully loaded, a few short prompts are enough to get you back in.

Sleep acts like automatic compacting: some details are lost, but the thread is still there.

Wait 3–5 days, and the forgetting curve wipes the context. You have to cold-start again.

That’s why interruptions are so expensive: they don’t just break attention, they destroy context.

Humans really are just biological AIs with tiny context windows 🤣

[LegolasWang @LegolasWang]

@garrytan So true

[LegolasWang @LegolasWang]

人的一个心流，其实就是一个 Context 上下文窗口。窗口不大不小，应该是远小于 1M Tokens

在心流中，上下文信息完整。几个简短小的聊天，就能激活心流。睡一觉，大脑自动 Compact，再次开始心流就会少很多信息，但勉强也能接上。

隔断三五天，记忆曲线生效，上下文信息彻底丢失，基本重新需要重启。这个聊天就废了，得重来。

心流中，人脑在努力记住上下文，因此一被打断就得重来。心流中不被打扰的情况下，人脑能勉强 Context Switch 几个不同任务，但多了不行。

这么想想看，人，就是个 AI 🤣

[LegolasWang @LegolasWang]

@garrytan Happy birthday 😊

[Jesse Pujji @jspujji]

https://t.co/BQiXJVUT17

[Marc Andreessen 🇺🇸 @pmarca]

My information consumption is now 1/4 X, 1/4 podcast interviews of the smartest practitioners, 1/4 talking to the leading AI models, and 1/4 reading old books. The opportunity cost of anything else is far too high, and rising daily.

[Prajwal Tomar @PrajwalTomar_]

https://t.co/vXr1MQ45wa

[Tw93 @HiTw93]

https://t.co/RTkJ3NQKeS

[Zara Zhang @zarazhangrui]

Remember: by default, nobody cares about you

Most people consistently overestimate how much other people care about them

[Elvis @elvissun]

https://t.co/DotZ3V6XhJ

[LegolasWang @LegolasWang]

3) These capabilities exist in Claude Code too, but OpenClaw wraps them into a cleaner model. Setup is rough for beginners, but once set up it’s super flexible—24/7 auto-support + multi-channel tooling flows feel natural. A very fun new toy 🪀

[LegolasWang @LegolasWang]

🦞 and Zara talked about OpenClaw today. Lots of people are asking how it differs from Claude Code. I spent a day with it and still feel it can do a lot of the same things, but the framework itself has real potential and is a fresh, fun way to think about agent systems.

[LegolasWang @LegolasWang]

2) The standouts: built-in heartbeat always-on; very high permission model for wild workflows; and as a comms Gateway, it nudges you to connect email, Feishu, Discord, Twitter as one system.

[LegolasWang @LegolasWang]

Haha, OpenClaw is set up and running 😄 I just gave the shoutout in Telegram, and my tweet went out right away. I’m now using Codex 5.3 Spark for my posting pipeline—finally, OpenAI and this model are a great match. It feels like the whole system is now focused on faster reasoning, grounded in first-principles design. And speed is everything when token throughput matters for Agent workflows.

[LegolasWang @LegolasWang]

I finally got OpenClaw and everything finally clicked. Turning chaos into smooth flow, one happy day at a time ✨

[LegolasWang @LegolasWang]

I just switched from photo chaos to a real yearbook. FlipYear turns my favorites into a 3D flip story, one day at a time—private, local, and surprisingly satisfying. Feeling good 😄