Roxy Lau

Professor at CUC | Visiting Scholar at UC Berkeley
Founder of AI tools for academic research and writing

[Roxy Lau @BossBunnyAI]

Over the past two weeks, we’ve been working on a new experiment at Figpad. Creating a scientific poster usually takes one or two days, but surprisingly little of that time is spent on science. Most of it goes into extracting figures, reorganizing content, and tweaking the layout. We wondered if AI could handle that part instead. Today, it can read a paper, extract the key figures, and generate a solid first draft of a conference poster in about five minutes. It’s not about replacing researchers—it’s about giving them back the hours they lose to formatting. Building toward that future has been a lot of fun. https://t.co/ghJOt4s5NZ

[Roxy Lau @BossBunnyAI]

Iterated 500+ times on prompts and finally got an AI-generated academic poster that’s actually usable.

Before: ~2 days to design one poster
Now: AI drafts the first version, I refine it manually, done in ~1 hour

AI doesn’t replace academic design judgment.
It gives researchers a much faster starting point. #figpad.ai

[Roxy Lau @BossBunnyAI]

I finally got Hermes to write an 80/100 literature review.

This morning, I was using Hermes during my commute to clear emails and schedule meetings. Feeling pretty proud of myself, I told my family, “My productivity is insane now. I’ve already finished half my work before getting to the office.” A few minutes later, I realized I had missed two subway stops because I was too busy chatting with Hermes. The time AI saved me was immediately spent riding the subway.

In the afternoon, I started training Hermes to become a better research assistant. I taught it how to search for papers, download them, read them properly, and then gave it a set of literature review writing skills to learn from. What surprised me was how quickly it improved. It didn’t just follow instructions—it kept refining its workflow and getting better with each iteration.

I then gave it a research topic I’m very familiar with and asked it to write a literature review based on full-text papers, format it, and send it to my email. The result was better than I expected. Not perfect, but a solid 80/100.

One thing I learned from this experiment: the future bottleneck isn’t AI. It’s knowing what to teach AI. There are thousands of Skills available now, but not all of them are good. A Skill that works for medical research may be terrible for social science writing.

Treat your AI like a junior researcher. Don’t let it learn everything. Teach it the right things.

[Roxy Lau @BossBunnyAI]

GPT Image and Nano Banana can now generate impressive scientific figures. But the more I test them, the more I feel the real bottleneck is not image generation. It is prompt writing.

Most researchers do not need a prompt like “make a beautiful scientific figure.” They need a prompt that explains the science: what the mechanism is, what happens first and next, which cells/molecules/materials must appear, what should be labeled, how the layout should work, and what the model should avoid.

That is the difference between a figure that looks nice and a figure that actually communicates research.

So we open-sourced 140+ research figure prompts with real output examples across graphical abstracts, mechanisms, workflows, lab apparatus, microstructures, systems/networks, cross-sections, ecology diagrams, and journal-cover-style visuals.

The goal is simple: help researchers spend less time staring at a blank prompt box, and more time communicating their science clearly.

Repo:
https://t.co/7oVovuxdT2

Use it, fork it, or contribute prompts from your own field.

[Roxy Lau @BossBunnyAI]

One of the happiest moments as a teacher is not simply seeing students graduate.

It’s seeing them discover what they truly enjoy and find a path they want to pursue.

Congratulations on your graduation and on finding work you’re genuinely passionate about. The future is just getting started.

[Roxy Lau @BossBunnyAI]

Nature says ChatGPT can help with academic writing.

But the key is not “write my paper for me.”

The real value is context-first revision:
preserve author intent, improve clarity, and avoid overstating claims.

So I turned this workflow into an Agent Skill:

Academic Writing Polisher

https://t.co/iCARtpDhg5

[Roxy Lau @BossBunnyAI]

勇气和智慧是人类最伟大的赞歌！

[Roxy Lau @BossBunnyAI]

昨天在准备周末分享的PPT。刚开始的时候，觉得很难受。那种感觉好像是在期末考试里考了70分，老师邀请我去给全班同学分享学习经验。是的，心里有种「我不配」的感觉。后来完成了自我说服。也许那些刚起步，考试不及格的人需要我。突然自洽了

[Roxy Lau @BossBunnyAI]

中午听到一个朋友说，他在期待明年的春天。他看到学校办公楼前挖了一个池塘，然后自己去研究了一下水质、环境，联系了某宝，预购了一大批青蛙幼苗。 他说：“明年我们可以听到蛙声一片了。” 听完以后，我觉得他「活过来了」，开始浪漫和鲜活。过去的很长时间，他情绪不佳。
我也开始期待了，期待他躲过所有摄像头，去成功的放入幼苗。也愿意和几个朋友一起，协助他，突破「保安」「摄像头」的监控。

[Roxy Lau @BossBunnyAI]

这两天发现很多事情底层的方法论都是一样的。上周，我请教少楠，我很关心一个问题：“请问要怎么做，才能「成长」为你这个样子？” 我试图去找到一个方法，然后企图通过按部就班的学习，来达到最终的目的。但是，后来交流过多个前辈，我理解了，没有个一个课叫做「创业」，只能通过经历去学会。
回到十多年前，我问我的导师：“请告诉我，怎么写一篇博士论文。”他说：“你写完了，就会了，不是会了，才去写。” 我当时听完以后，我记得我非常崩溃，我背后蛐蛐导师在说「正确的废话」。
十年后的一天，我的学生问我：“怎么写论文呢？”试图让我说出1、2、3、4的路线。但是，我也只能回答：“你不会，但是你先去写吧。” 我在回答他的时候，我瞬间解答了我自己一直的困惑。是的，很多事情我现在都不会。但是去做了，然后就会了。不是吗？

[Roxy Lau @BossBunnyAI]

今天学生找我讨论论文。她很沮丧。她在论文写到一半的时候，陷入了自我怀疑：我这么写能毕业吗？我看了很多论文，就是输出艰难，我觉得自己不行，没有进展？我很理解她。我在写博士论文的时候，也曾经无数次怀疑过。尤其是看了大量论文，依然不能很好的输出的时候。那个时候，我看到了「荷花的故事」，给了我很多力量。现在把这个故事送给正在自我怀疑的你。

听过一个故事：
有人把荷花种在池塘里，
第一天，只有一小片叶子浮在水面。
第二天，也没什么变化。
第三天，依旧如此。
……
直到第 29 天，荷花才开了一半。
第 30 天，池塘里突然开满了荷花。

不是第 30 天荷花变厉害了，
是 前 29 天的积累终于有了回响。
⸻
🧠 写论文也一样

第 1 天找资料，
第 5 天推翻前面的方向，
第 10 天怀疑自己是不是不适合读研、读博，
第 15 天开始焦虑失眠，
第 20 天想要放弃……

但某一天，突然：
•思路清晰了
•框架完整了
•数据跑通了
•终于看到「成稿」的样子了

那天你会发现：

原来努力不会白费，它只是不会提前通知。
⸻
✨ 写给正在坚持的你
如果你正在写论文，不要急。
今天看似没什么进展，
但你做的每一次文献整理、每一个图表、每一句思考，
都在为最终的 “第 30 天” 积蓄能量。
这世界上，
没有一份论文是轻松写完的，
但每一份完成的论文，都值得骄傲。

坚持下去，
你会遇到那个突然开花的自己。

[Roxy Lau @BossBunnyAI]

最大的自由是敢于让所有人失望。

[Roxy Lau @BossBunnyAI]

Daily reminder: Enjoy the process of building your product! ✨

[Roxy Lau @BossBunnyAI]

记录一下citely迎来第一个终身用户。我们并未设立这个价位，但用户自己要求成为永久用户，3555元的价格还是用户和我商量后很快决定的。很神奇的感觉。

[Roxy Lau @BossBunnyAI]

周日和一个大厂的HR聊天，其中一句话对我触动很大，她说她们大厂的机制非常混乱，属于变开车边掉零件的那种。但是最大的优势是，这个车一直往前跑。没有停下来过。

[Roxy Lau @BossBunnyAI]

Citely更新一版，文献查询结果里新增一个more，提供更多参考。然后在部分文献里，提供摘要供大家直接参考，不用点击全文了

[Roxy Lau @BossBunnyAI]

@yihangGuo1 可以啊，➕v：littlefools

[Guangzheng Li @iguangzhengli]

Lee Robinson 从 Next.js 退出后，加入 Cursor 成为 AI Coding 的教育布道者，还是坚持一如既往的从零教学啊

当初学 Next.js 就是看了 Lee 视频感觉还挺有趣，现在转头去教 AI Coding 了

完全零基础教学，应该对非程序员接触 AI Coding 很有帮助👍

https://t.co/CVGhSTw5o5

[Roxy Lau @BossBunnyAI]

@iamtonyzhu I'm still at the first step : reply 😅

[Roxy Lau @BossBunnyAI]

For the first time since we launched https://t.co/1mCWajpbDM, I said no to an opportunity to be in the spotlight. I feel like our users are at capacity, so we need to quickly iterate based on their feedback first before we can focus on promotion.