Notes

Astro is joining Cloudflare. I always had a feeling Astro would be acquired — and now it’s happening. Astro has been my favorite framework for building static websites, and it’s my choice for my personal websites now. It reminds me of when I first discovered static site generators like Jekyll — just basic build-time templating and composition, making creating a blog very easy. Features like content collections and islands are genuinely innovative. I hope it stays productive and keeps its simplicity.

Ralph Wiggum as a “software engineer”. The AI field is evolving so fast like your math classes in high school, that if you miss a week, you’re suddenly lost. For me recently, it’s Ralph, a new pattern for coding agents that pushes them to a higher automation level.

Its name comes from a character called Ralph Wiggum in the show The Simpsons, who somehow captures the spirit of this technique.

To get familiar with Ralph, I skimmed (and watched) these materials, in addition to the original post by Geoffery Huntley:

• Matt Pocock’s walkthroughs: Ship working code while you sleep with the Ralph Wiggum technique, and 11 Tips For AI Coding With Ralph Wiggum
• Greg Isenberg’s video: “Ralph Wiggum” AI Agent will 10x Claude Code/Amp
• Ryan Carson’s article on X: Step-by-step guide to get Ralph working and shipping code

In short, Ralph is a technique that runs your coding agent sessions in a loop. It pushes the typical coding agent workflow — you give it a task, watch it work, and then a new task based on its output — forward by making the agent itself assess the outputs and decide what’s next. Back in 2025, we’ve got the agreement that an “agent” is simply an AI program running tools in a loop to achieve a goal¹. Ralph extends that idea naively: It’s a bash script running agent sessions in a loop to achieve a goal.

To run agents the Ralph way, you basically need the following harnesses:

• A bash script that simply runs your coding agent in a for loop
• A PRD file that lists and tracks the tasks, commonly organized as prd.json
• A progress note that the agent appends to when completing tasks, providing relevant context to the next agent session, commonly organized as progress.txt

These elements reveal what’s truly valuable about the Ralph idea: It formalizes a context engineering approach when tackling large scale development requirements. And that’s why Ralph differs from just using a single agent session for all tasks. Every time the session completes a task, it checks the tasks in prd.json, appends notes to progress.txt, and usually makes a git commit. Then a new agent session starts with the context window cleared, so the files the last session updated serve as the only memory of the Ralph loop.

Rough notes here. If you’re interested in the details, check the materials above. It’s indeed a new idea in the field and the community will explore it further to see if it’ll truly stand out.

1. Simon Willison’s well-known article: I think “agent” may finally have a widely enough agreed upon definition to be useful jargon now ↵

Last weekend, I reorganized my personal websites. Here’s the new structure:

Website	URL	Description
Personal landing page	https://zlliang.me	Brief introduction and navigation
Zilong’s Tech Notes	https://tech.zlliang.me	My tech exploration journal (English)
Zilong’s Days	https://days.zlliang.me	My life reflection journal (Chinese)

In this post, I explain my motivation for splitting the sites. After this change, I’m ready to mark version 1.0 of my writing framework. The permalinks will be stable, and I’m excited to share more technical and personal topics in these spaces.

Today I spent a day trying to add i18n support to the website. I brainstormed ideas and documented them in a GitHub issue. I also tried to design and implement a translation key system and a new routing system, and wrote a lot of code.

In the end, I realized it makes both the site and my writing workflow more complicated than I’d like. Direct i18n support doesn’t feel like the right move right now — it adds friction and mental overhead, and I want to be able to just start writing when an idea comes up.

Since the website’s structure is entirely under my control, I want to design a content organization that genuinely fits my own writing habits while still being open and readable to different audiences. I don’t want to add structural complexity to the site just to satisfy a sense of “everything must be unified.”

So I’m going to park this issue for now. The site will stay focused on technical writing and public English content. Anything that doesn’t fit yet will live in my private Notion workspace, and I’ll revisit it later when it makes sense.

I just updated the license of this website. Now it’s dual-licensed: code under MIT, content under CC BY 4.0. Previously I used CC BY-NC-SA 4.0 for content, but decided to go more open — fewer restrictions, more sharing.

Here’s the commit: zlliang/zlliang@af99933.

Happy New Year! Here’s a quick recap of my New Year’s break:

I finished watching The King of Internet Writing, a video podcast by David Perell about what we can learn from Paul Graham’s writing. I’m ready to write more, and better, in 2026.

On New Year’s Eve, I was traveling with my partner in Chongqing. We ate spicy hotpot and walked through the hilly streets!

Chongqing's cityscape on New Year's Eve

I’m planning to add new features to this website. I created two GitHub issues, following Simon Willison’s approach to building features¹:

• #71: Support bilingual content with AI-assisted translation
• #72: Support search across notes

1. Simon Willison’s blog post: How I build a feature ↵

Andrew Kelley's growth curve

I’m just reminded of the growth curve Andrew Kelley showed in his talk A Practical Guide to Applying Data Oriented Design. I watched it several days ago and realized I’m now at a plateau of my own. Back in university, I built a solid foundation in web frontend and landed a job in it. Now I’ve hit another bottleneck, eager to jump to the next stage. Andrew found his trigger in a book; I’m still looking for mine. But two strategies are already in my mind: writing and starting my own projects. I bet that producing and creating will push me to evolve.

I don’t know what still newer marvels will make writing twice as easy in the next 30 years. But I do know they won’t make writing twice as good. That will still require plain old hard thinking.

— William Zinsser, On Writing Well

William wrote this in 2006. Nearly 20 years later, LLMs have made producing text easier than ever — yet concerns about AI-generated junk are growing too. Good writing remains rare and precious. It still needs to be written and rewritten, again and again, by humans.

How uv got so fast¹. I haven’t followed the Python ecosystem for maybe five years. But I know uv has taken off. I have it on my Mac, and it’s my go-to when I occasionally want to play with Python. It feels like pnpm or Cargo — fast and modern.

I assumed Rust was the main reason uv is so fast. Turns out, that’s actually the least important factor. Years of PEP standards made uv possible in the first place. Intentionally limited compatibility with pip, plus smart language-agnostic optimizations, did most of the heavy lifting. It’s the design choices, not the language choice, that really matter.

1. Discussion on Hacker News ↵

I just added pagination to note pages like /notes and /notes/types/link, as the number of notes grows. Each page now shows up to 20 notes, and a tiny pagination indicator lets you navigate between pages without scrolling endlessly. I used Astro’s built-in pagination feature. Here’s the commit: zlliang/zlliang@649a5df.

Happy holidays! At the end of 2025, I’m starting a blog. I’ve already written several entries and feel confident I can keep it going.

Here I talk about my past attempts, the writers who inspired me, the motivation, the topics to cover, and the approach I’m taking. It’s my version of a blogging manifesto.

I gradually realized that a unified formatting rule set is needed when working with multiple AI chatbots and agents.

Output formatting styles vary from model to model. For technical topics, I’ve found that Claude tends to output responses like complete documents, starting with an h1 heading and loves to use horizontal rules to separate sections; Gemini usually skips to h3 headings directly without h2 ones, which in my opinion is not a good practice.

Here are examples I tried on OpenRouter, prompting “Explain the Python programming language.”

Even worse, from my experience, outputs from different versions of the same model series (e.g. GPT-5 and GPT-5.2) can vary greatly in terms of formatting.

To address this issue, and to unify output styles of different tools I’m using (ChatGPT as my daily driver, Gemini for work, and Amp as my coding agent), I drafted a minimal formatting guide as follows:

The following formatting rules MUST BE FOLLOWED.

Shared formatting rules:

• Use consistent formatting within the same response
• Insert spaces between English words and CJK characters
• Always specify the language for syntax highlighting when using fenced code blocks
• Never number headings (e.g., ## About me, not ## 1. About me)
• Never horizontal dividers (<hr> or ---) unless they add clear structural value, especially not directly before headings
• For list items, do not use a period at the end unless the item is a complete sentence

For chat responses:

• Use “Sentence case” for chat names (auto-generated chat titles) and all section headings (capitalize the first word only), never use “Title Case” in such circumstances
• Use heading levels sequentially (h2, then h3, etc), never skip levels; Introductory paragraphs may be needed before the first heading in chat responses; Never use h1 for chat responses
• Avoid filler, praise, or conversational padding (for example “Good question”, “You’re absolutely right”)

For document generation and editing:

• Use “Title Case” for top-level headings (e.g. h1), typically only once in a document, and “Sentence case” for section headings (capitalize the first word only)
• Use heading levels sequentially (h2, then h3, etc), never skip levels

I apply these rules to the custom instructions setting in ChatGPT and to AGENTS.md for my coding agent.

Custom instructions setting in ChatGPT

GLM 4.7 and MiniMax M2.1. Chinese AI labs first caught the world’s attention with the DeepSeek models in late 2024. Then in the second half of 2025, we saw a wave of Chinese open-source models like GLM 4.6, MiniMax M2, and Kimi K2. People like these models for their low price, open weights, and solid performance — just slightly below state-of-the-art proprietary models¹.

Today, the updated GLM 4.7 and MiniMax M2.1 dropped on the same day. As public holidays approach in the US, Chinese AI labs keep pushing forward, making good use of the time 😉

AI is a rapidly changing field. I’m not one to chase every new model release, though I do find myself following this topic more recently. I’m still learning these concepts and trying to find a pragmatic way to use AI tools. I use ChatGPT as my daily driver, Gemini for work (my company subscribes to it), and Amp as my coding agent.

I may not post about every model release in the future, but here are the models on my radar:

• Proprietary models: GPT (OpenAI), Claude (Anthropic), and Gemini (Google)
• Open-source models: DeepSeek, Kimi (Moonshot AI), GLM (Z.ai), and MiniMax

1. For example: Vercel CEO Guillermo Rauch’s post on X, T3 Chat creator Theo’s post on X ↵

TIL: When editing Markdown files in VS Code, you can paste URLs as formatted links via the markdown.editor.pasteUrlAsFormattedLink.enabled setting.

This setting was first introduced in June 2023, with the release of VS Code 1.80.

This is a nice quality-of-life feature. I used to type brackets, parentheses, and URLs manually, always wishing for a simpler way. I’m now using the smart option, which “smartly creates Markdown links by default when not pasting into a code block or other special element.”

Here’s a quick demo:

Paste URLs demo