GoatWang

Father of Oysterun

[GoatWang @wanghsuanchung]

《換回 Claude Code 一週後》

自從上週把主力換回 Claude Code，我才真正開始把 Oysterun 累積的 Bug 全部排進開發排程。Codex 最近的表現，真的不行。

以前 Codex Agent Team 的交付完成度大概只有 60%，現在 Claude Code 已經可以到 95%。

最明顯的差異是：只要是我已經確認、交代要處理的問題，Claude Code 都能穩定交付。

Codex 面對比較複雜的問題，常常從規劃階段就開始鬼打牆，實作結果也容易跟原本的規劃產生落差。

另外一個很重要的指標是驗收。

我會要求 Agent 把環境和驗收步驟都準備好，讓我可以實際檢查所有交付。Claude Code 現在已經能讓我在兩小時內，驗收完一個三人團隊一週的工作；Codex 準備的交付，我可能要花一整個週末才能驗證完。

總之，在今年十月的比賽，Claude Code 在長 Context 的一致性和指令遵循能力上，完全勝過 Codex。

我現在已經有一套能讓我有信心同時開十個以上 Session 的 Agent Team。

接下來剩下的，真的就是財力問題了。

[GoatWang @wanghsuanchung]

拿ChatGPT對話與醫生互動…

今天去看牙齒，醫生建議我做一個支撐金屬，然後再套牙冠。

我聽完後就很自安然的跟chatGPT開始聊天，我就發現有一個東西叫做Furrule，就是在套上牙套時，牙套能不能部分套在自己的牙齒上（或是全部套在支撐金屬），可以看圖，文獻統計這會影響失敗高達兩倍（10% and 20%）。

但是，我需要知道我牙根斷裂的深度，以及最終會不會可以保留Furrule，才裝上牙冠。

但是，要在醫生面前討論（質疑）他的決定，真的是很需要勇氣啊。

不管了，不問的話，做了智障決定，我一輩子都會困住。

還好後來醫生很有耐心的回答我的問題，真的是好人醫生平安。

[GoatWang @wanghsuanchung]

《正確使用 LLM 的方法：牙醫診所排除廣告檢索》

請服用以下prompt：

請使用電腦操作功能控制瀏覽器，搜尋 PTT 和 Dcard 上推薦的台北植牙診所或醫院。
請交叉比對不同來源或不同使用者的分享；同一位醫師或同一家診所、醫院，至少需要被兩位不同的使用者推薦，才納入結果。
請盡量排除廣告與業配，運用自然語言處理（NLP）分析文章和留言，辨識並過濾疑似廣告內容。如果無法確定是否為廣告，先保留並視為真人分享。

p.s. 台北有網路上有推薦的植牙診所真的好少，拜託大家推薦給我。

[GoatWang @wanghsuanchung]

《新時代的軟體世界》

大家不要再卷 AI 開發了。產品做得好就會有人用的時代已經過去了；這年頭成功的路徑是噱頭、行銷與資本，好的產品已經是前提。

產品開發過程不要急，也不要先發布，要藏好，準備放大招。

然後開始砸錢，準備大量的 marketing 內容、聯繫領域網紅，再安排好一夜爆紅。

只要你的產品真的解決了一些問題，或是創造話題，就會有一堆人不知道為什麼開始跟風發文、評價了。

[GoatWang @wanghsuanchung]

把原本 OpenRouter 的 token 消耗轉到 Codex 訂閱後，這幾天看到帳單數字，心情舒坦不少。

[GoatWang @wanghsuanchung]

PitchYourOwner - 這次FutureMode黑客松的作品

現在的交友軟體，不是關注性就是關注錢，社交媒體更是充滿對立、廣告、讓人浮躁的短影片。

對我來說，現在要找到一個滿意的社交體驗，或是一場讓我可以回味的對話，幾乎已經遙不可及。

我期待PitchYourOwner幫助你我找到正在
- 解決相同問題、或
- 探索相同議題、或
- 有相同興趣嗜好
的人。

只要平常有在跟AI對話，無論是爬山、攝影、舞蹈、或是各式話題甚至是專業領域的追求，三分鐘就可以開始，尋找你的soul mate。

讓你的Agent把自己推銷出去吧！

https://t.co/nmi50eEh42

[GoatWang @wanghsuanchung]

Everyone Has Their Own Typeless Build—but How Do We Actually Compare Them?

Everyone seems happy with their own Typeless build
but there is no consistent benchmark for comparing models.

Oysterun needs voice input too,
so I recorded 36 real coding-agent commands and tested local, CLI, and cloud E2E options.

Full report:
https://t.co/FxfBSGN58O

P.S.
BYOK Voice Input feature on Oysterun will be available for next week.
Context-aware dictionary correction is planned within a month for better recognition accuracy.

[GoatWang @wanghsuanchung]

Oysterun has reached an important infrastructure milestone: the Tunnel Service is finally ready.

The Host no longer needs a static public IP, and the Client can still find it through a verified, encrypted connection path. Under the hood, we integrated NAT hole punching, WebRTC, E2EE, ICE, STUN, TURN, and the necessary Relay fallback.

But none of that technology is what matters. As long as your computer can get online—whether it’s a laptop, desktop, or VM—you can launch Oysterun with one click and start a secure remote connection.

After a week of intensive testing, we’ve finally reached a balance of cost, stability, and security that’s ready for release.

The new version is planned for next week.

[GoatWang @wanghsuanchung]

Lately, I’ve been thinking I need a small local model just to answer all of Codex’s security questions.

[GoatWang @wanghsuanchung]

Working at DeepSeek is peak romance.
This is basically China’s DeepMind.
(Playing Go for years without needing to make money somehow feels even more hardcore?)
Liang Wenfeng on how they actually operate:
1. 50/50
“Half the time is top-down real work — shipping V4, coordinating, etc.
But it can never eat more than 50% of people’s time.
The other half? Do whatever you want. Explore freely. No prerequisites. As long as the compute is there, you don’t even need to ask anyone.”
2. Stay loose
“Research needs a relaxed environment. Tighten the screws and exploration dies. People have to want to think about the problems in their free time.
We’re extremely focused — we do very few things — so overtime almost never happens.
You’ll notice a lot of our products are imperfect. We don’t rush to polish them. That’s deliberate culture.”
3. Beyond profit
“Keeping the team stable is the single core interest. As long as the team stays and keeps going, AGI will happen. Money and resources are not the bottleneck.
A lot of people aren’t here purely for the paycheck — they want to work somewhere that actually has a shot at AGI. That’s why our turnover has always been lower than peers.
We don’t have a perfect organizational model to copy. Unlike Bell Labs we still have to keep the company alive.
But history’s greatest companies all had something they cared about more than profit.”
Video: https://t.co/NCh2HaFOZi

[GoatWang @wanghsuanchung]

Video:
https://t.co/FiMpWCFqy7

[GoatWang @wanghsuanchung]

About Liang Wenfeng’s talk at an investor meeting...

1. It was leaked and has now been taken down across China.

2. He said: “Huawei’s 950 SuperPoD can fully match the performance of the GB200 and GB300.”

3. “Within one year, China will no longer be held back by the software ecosystem gap around CUDA.”

With AI, rebuilding NVIDIA’s software ecosystem is now foreseeable.

DeepSeek V3 has already removed CUDA and uses its own TileLang stack. It can be moved directly to Huawei hardware.

[GoatWang @wanghsuanchung]

I gave my Agent one YouTube screenshot and asked it to write a TODO.

It completed it instead: saved the video audio and English transcript, used Whisper for Chinese, found the cited research, and validated my point.

yt-dlp can still be IP-blocked, so reliability infra comes next.

Publishing and post management are already connected. A world where creators can focus only on content feels much closer.

[GoatWang @wanghsuanchung]

Finished my automated social posting system:

• Fully remote from my phone
• Post + engagement tracking
• 0 APIs—all browser-based
• Threads, X & Facebook connected

Buffer is $6/account, so I save $24. Build a few tools like this and the subscriptions pay for themselves 😂😂😂

[GoatWang @wanghsuanchung]

Four hours to eight hours.

My first Oysterun mobile UI could drain the battery in about four hours while I kept the app open. It pushed every tiny agent update over WebSocket.

I switched to long polling so the screen redraws only for progress worth showing.

In my everyday use, I stopped noticing the heat, and my phone lasted at least eight hours.

Rebuilding the interface was expensive. Rethinking it from first principles was worth it.

[GoatWang @wanghsuanchung]

What do you do when an agent has been running forever, but you don’t want to interrupt it?

Easy…

Ask another agent to check on its progress.

Cross-session communication skills are finally done—and it feels so good.

[GoatWang @wanghsuanchung]

https://t.co/3RyjNZJaJc

[GoatWang @wanghsuanchung]

I used to hate frontend development.

Now that I’m building my own product, I finally understand how much work it is. Good design makes a huge difference, and understanding user behavior takes a lot of thought.

Luckily, I’m also a full-time user.

Today, I added a feature that lets me jump between my own prompts while browsing agent conversations. Finding context is now insanely fast—and honestly, it feels great.

[GoatWang @wanghsuanchung]

Oysterun — Agent Team

Last night, I suddenly discovered that Oysterun was storing conversation history in JSON.

I had specified SQLite very early on, but somewhere along the way, a series of trade-offs led to JSON being used instead—and I completely missed it.

So I handed the migration work to my AI team: implementation, automatic migration, and verifivation.

When I checked this morning, the migration was already complete and had been tested against multiple edge cases.

Tasks like this—important backend work that doesn’t require any UI changes—are incredibly satisfying to delegate to an agent team.

Oysterun’s cross-session communication skill is now fully developed, so the whole team can collaborate smoothly across sessions.

You can inspect every agent’s conversation, step in at any time, and steer the work while it is running.

Once you turn on the loop, you really start to feel the power of loop engineering.

P.S. Yesterday’s marketing progress was delayed a little.

But the Agent Store demo is ready, and the demo video is coming very soon.

[GoatWang @wanghsuanchung]

https://t.co/3RyjNZJaJc

[GoatWang @wanghsuanchung]

Oysterun — Bringing AI Agents to Everyone

I built a stock research agent because I wanted my mom to use AI.

It taught me something unexpected.

AI agents don’t have an AI problem.

They have a deployment problem.

Claude Code.
Schedulers.
Web servers.
Notifications.
Terminal.

Every extra dependency eliminates another 90% of potential users.

GitHub is full of incredible AI agent projects.

People fail to build them because the stack is too complicated.

That’s why I built Oysterun.

My goal is simple: AI agents should be one click away.

Today, the first version is finally here.

For the past three months, Oysterun has been running my AI team 24/7.
Together, we built this first release.

While I sleep, it’s working.
While I’m at work, it’s working.
Even during my month-long trip across Europe, my Mac mini kept running back home in Taiwan.

Today, I can’t imagine building AI products without it.
This is only the beginning.

https://t.co/oA0jhRQt7j