Forem: 咕咚

After a Week with Doubao, I Switched Back to WeChat Input

咕咚 — Thu, 29 Jan 2026 15:01:29 +0000

After using Doubao input method for a few days, I decided to switch back to WeChat input.

Last week, I wrote an article about how WeChat input has been getting better and better, particularly praising its massive improvements in voice input. The article got recommended and attracted a lot of attention and discussion.

But among all the comments, a surprising number were promoting Doubao input method.

Some said Doubao is the top 1 for voice input, others praised its fast response and dialect support.

When I shared the article on Xiaohongshu and V2EX, the comments were similarly consistent — everyone was pushing Doubao.

I figured since everyone recommends it so highly, I should give it a try.

So I downloaded Doubao and used it as my primary input method for a week.

But today, I finally couldn't take it anymore.

1. The Dealbreaker: Cross-Device Copy

Even though Doubao's voice input is excellent (this article's draft was actually written using Doubao voice-to-text), I had to switch back.

The reason is simple: cross-device copy-paste.

As an indie developer and blogger, my workflow depends heavily on phone-computer collaboration.

After writing a draft on my phone using voice, the next step is to send it to my computer for editing and formatting.

With Doubao: I have to copy the text to WeChat's file transfer assistant, or send it to Feishu, then open it on my computer, copy it again, and finally paste it into my editor. This process is too fragmented.
With WeChat input: Copy on phone, paste on computer with Ctrl+V. Done.

This seamless experience — once you get used to it, there's no going back.

I really hope Doubao adds a cross-device clipboard feature, but it currently doesn't have one.

2. Voice Input: It's a Tie

Many say Doubao's voice recognition is "god-tier" and much better than WeChat's.

After a week of deep testing, my feeling is: it's good, but not overwhelmingly better.

Both are now optimized extremely well and fully meet daily voice input needs. Doubao might be slightly faster in post-processing corrections, but this tiny difference isn't enough to make me give up the convenience of WeChat's ecosystem.

Speaking of ecosystem, WeChat input has the home-field advantage:

Emoji suggestions: When typing in WeChat chat, it automatically suggests emojis. This is huge for me as someone who uses a lot of emojis.
Custom phrases: WeChat input makes adding custom phrases very convenient — just copy any text and convert it directly. Doubao requires manual entry one by one, which is more tedious.

As for people complaining about WeChat input's large installation size (200M vs Doubao's 140M), I don't think it matters with modern phone storage. If it works well, I'm happy to accept the extra 60M.

3. A Pleasant Surprise: PC Voice Input

Just today, I saw a screenshot in a group chat that solidified my decision to return to WeChat input.

It was a new feature in WeChat input's PC version: press a key to do voice input directly on your computer.

Yes!

In my last article, I wished that WeChat would add PC voice input. I didn't expect it to come so soon — maybe my article reached the WeChat input product manager 😄.

In contrast, Doubao doesn't even have a PC version yet.

Voice input on the computer is really important.

Especially now that we frequently talk with AI (ChatGPT, Claude). If you rely solely on typing, getting your thoughts out is too slow. At home, I'd much rather speak to my computer — it's extremely efficient. This is also why I've been advocating for "conversational programming" — speaking is the most natural form of interaction.

Although I'm currently using Zhipu input on my computer (giving Zhipu a plug here since CodeParty helped me a lot, they deserve this free promotion), and the experience is quite good.

But I believe once WeChat's PC voice feature rolls out widely, my primary computer input method will likely return to WeChat.

4. Why Did Voice Input Suddenly Explode?

After all this input method experimentation, I've been thinking about a question: Why has voice input suddenly become essential in the AI era?

I think there are two main reasons:

First, the interaction target has changed.

Before, we mainly chatted with people — short messages. Now we interact with AI, needing to provide lots of context and prompts. In this scenario, speaking is the most natural, highest-bandwidth interaction method.

Second, technology has reached a tipping point.

Thanks to large language models, voice-to-text (ASR) is now incredibly accurate. Whether it's WeChat, Doubao, or Zhipu, recognition rates have reached astonishing levels.

Technology has matured, demand has exploded, and voice input has become a new battleground.

Final Thoughts

My one-week "defection" experiment has ended.

Although I switched back to WeChat input, this doesn't mean Doubao is bad. On the contrary, I think it's a great thing that Tencent and ByteDance are competing in the input method space.

No competition, no progress.

As users, we benefit. I look forward to them continuing to compete and improve the experience.

I'm Gudong, and good tools deserve to be seen by more people.

By Gudong, indie hacker building inBox Notes.

Two Nights, One Tutorial: The Hard Work Behind Sharing Knowledge

咕咚 — Wed, 28 Jan 2026 15:22:35 +0000

Yesterday's article surprised me.

I wrote a nearly 4,000-word guide on using Claude Code in China. After publishing, the view count wasn't impressive—barely 2,000.

But the share count exploded. Nearly 200 shares and climbing. This might be my highest share rate ever.

It confirms something: people love practical content.

When we see a hands-on tutorial, our first instinct is to save it or share it with a friend. Whether we'll use it immediately doesn't matter—"storing it away" feels valuable.

Honestly, seeing all those shares made me happy. Writing tutorials is genuinely exhausting work.

1. Writing Tutorials Is Hard

Friends who know me well know I prefer sharing reflections.

I actually resist writing tutorials. You have to extract tacit knowledge from your brain and break it down into explicit steps:

Is this screenshot necessary?
Can a beginner follow this logic?
Need to verify if the process has changed.

That article took over two hours. I just wanted to lie flat when done.

So why write it?

Partly because I promised a friend (and my wife kept nudging me to honor that promise). But mostly, I genuinely wanted to share Claude Code.

But today, I want to talk less about the tutorial itself and more about a reflection on "how to learn."

2. Keywords > Tutorials

I don't know how many of those 200+ sharers actually followed the tutorial step by step.

Actually, it doesn't matter.

For readers, I think the greatest value of my articles isn't those thousands of words of steps—it's that you gain a "keyword."

Like Claude Code. Like CodeParty.

I strongly believe: in the AI era, knowing "what exists" matters more than knowing "how to do it."

Through my article, you learn about a tool called Claude Code that can change how you program. That's enough.

As for specific installation and configuration—my guide might not be the best. Take that keyword, search for it, ask DeepSeek or ChatGPT:

"What is Claude Code? How can I use it?"

AI can explain it more clearly and give you even more detailed steps.

I do the same when browsing Twitter or GitHub. I rarely follow others' tutorials step by step. I'm scanning for keywords. Once I spot something valuable, I grab that word and figure it out my own way (usually by asking AI).

Tutorials are just guides. Keywords are the keys.

3. Why Do I Still Write?

Since everyone can ask AI, why did I bother writing those 4,000 words?

Because of validation.

Tool experience can be "distorted." Feeling great when first starting doesn't mean it's actually great. When I first started, I was excited and wanted to share, but I didn't—my emotions were clearly skewed from objective reality, and my understanding had gaps.

I used Claude Code for over a month. I developed the SlideNote browser extension with it. I rewrote the inBox Notes PC version. I stumbled through pitfalls and actually built things with it. Only then did I feel comfortable throwing this "keyword" out to you.

I don't want people to waste time.

That's why I honestly discourage beginners at the start: if you're new, Claude Code has a learning curve. Try Trae first.

4. Final Thoughts

Writing tutorials is tiring. The numbers might not look impressive. But seeing so many shares proves it has value.

If I find better tools or discover better "keywords" in the future, I'll probably keep writing detailed tutorials while complaining "this is too exhausting."

After all, actively seeking and sharing—that's the joy itself.

Hope everyone finds their own new continent through these keywords.

By Gudong, indie hacker building inBox Notes. I write about AI-assisted development and the indie journey. Find me on GitHub or X/Twitter.

Claude Code Complete Setup Guide

咕咚 — Tue, 27 Jan 2026 14:27:18 +0000

If you follow AI programming, you've probably heard of Claude Code—Anthropic's official AI programming assistant.

It's powerful, but comes with barriers for some users:

Issue	Official Claude	Alternative
Price	$20/month (~145 RMB)	From 20 RMB/month
Payment	International credit card	Alipay, WeChat
Network	Requires VPN	Direct connection

This article teaches you how to use Claude Code with alternative model providers, walking through the complete setup from installation to实战.

1. What is Claude Code?

AI programming has evolved through three stages:

Code Completion (like GitHub Copilot) — You write, it completes
Chat Assistance (like ChatGPT) — You ask, it answers
Intelligent Agents (Claude Code) — It plans, uses tools, completes complex tasks

Claude Code is not a "tool", it's an "agent".

It doesn't just write code, it can:

Manipulate files, run commands
Understand entire project structures
Read information across projects
Remember context and conversation history

1.5 Is Claude Code Right for You?

Heads up: Claude Code has a learning curve.

It requires:

Comfort with command line (terminal)
Willingness to tinker with environment setup
Acceptance of occasional errors and debugging

If you're a complete beginner just wanting AI help with coding, I recommend starting with more beginner-friendly tools that have graphical interfaces.

But if you're willing to cross this threshold—Claude Code opens a new world.

Its capabilities far exceed traditional tools. Once you use multi-window, Skills, and MCP features, there's no going back.

2. Preparation: Getting API Keys

2.1 Register Account

Visit your chosen model provider's platform and complete registration.

2.2 Get API Key

After logging in, go to the API Keys section and create a new API Key.

Keep it secure — don't share it with others.

2.3 Choose a Plan

Model providers typically offer multiple tiers:

Plan	For Whom	Price
Lite	Casual users	Entry-level pricing
PRO	Heavy users	Higher quotas
MAX	Professional users	Maximum quotas

3. One-Click Installation: Coding Tool Helper

3.1 Install Claude Code First

Step 1: Open Terminal

Search for "Terminal" on your computer and open it.

Don't be intimidated — think of it as a cool-looking AI chat box.

Step 2: Run Installation Command

Paste this command in your terminal and press Enter:

npm install -g @anthropic-ai/claude-code

If you encounter errors:

If you have Cursor or other AI programming tools installed, try installing through them
Ask any AI for help troubleshooting

Step 3: Alternative Installation

If npm doesn't work, use the official script:

macOS / Linux / WSL:

curl -fsSL https://claude.ai/install.sh | bash

Windows:

irm https://claude.ai/install.ps1 | iex

Step 4: Verify Installation

After installation, run:

claude --version

If a version number appears, installation was successful.

3.2 Run Coding Tool Helper

Coding Tool Helper is a configuration assistant that helps load your model provider's settings into Claude Code.

npx @z_ai/coding-helper

You'll see a friendly interface — follow the prompts to paste your API Key and complete configuration.

4. First Use

4.1 Start Claude Code

Create an empty folder, enter it, and start:

mkdir todo-demo
cd todo-demo
claude

You'll see a welcome screen — AI is ready to help.

Select "Yes Progress" to agree.

4.2 Let AI Start Working

Tell it what you want to do:

Help me build an HTML todo app

Claude will:

Analyze requirements
Plan the implementation
Start writing code
Request your authorization when creating files

4.3 Authorization

When Claude needs to create files or run commands, it will ask for confirmation:

1. Yes
2. Yes, allow all edits during this session (shift+tab)
3. Type here to tell Claude what to do differently

Use arrow keys to select and press Enter to confirm.

If frequent confirmations interrupt your workflow, there's a solution:

4.4 Danger Mode (Optional)

If you trust the environment, skip authorizations:

claude --dangerously-skip-permissions

Not recommended for beginners.

5. Project Initialization: Help AI Understand You

For new projects with existing code, run:

claude
> init

This scans the entire project and generates a claude.md file — like an "employee handbook" for AI. For empty projects, you can wait until later or specify your requirements upfront.

Tip: Let AI Address You by Name

Add this to claude.md:

**IMPORTANT: Always address the user by name.**

The user's name is **Gudong**. When responding to questions or providing updates:
- Start responses with "Gudong," or "Alright Gudong,"

After this, AI responses will look like:

Gudong, I've completed the log simplification...
Alright Gudong, this is because...

This also serves as a context indicator — if AI stops using your name, context may have been lost and you should run /clear.

6. Common Commands

Basic Commands

Command	Purpose
`/exit`	Exit Claude Code
`/clear`	Clear conversation context and reload
`/help`	Show help information

Startup Parameters

Parameter	Purpose
`claude --compress`	Auto-compress context on startup
`claude --dangerously-skip-permissions`	Skip authorization confirmations
`claude --no-progress`	Don't show progress bar
`claude --debug`	Show debug information

Context Management

When conversations get long, Claude automatically compresses summaries.

You can also manually clear:

/clear

Or restart:

/exit
claude

7. Multi-Window Mode: True Multi-Threading

This is my favorite feature.

Open two terminal windows, same project:

Window A: Develop settings page
Window B: Develop about page

They work independently without interfering.

You can open 6, 8 windows simultaneously — as many as you can keep up with.

Git Worktree Mode

Advanced users can combine with Git Worktree.

Git Worktree lets you check out multiple branches of the same repo into different directories:

# Create a worktree for feature development
git worktree add ../feature-a origin/feature-a

# In new window, enter that directory
cd ../feature-a
claude

This way:

Window A: Fix bugs on main branch
Window B: Develop features on feature-a branch

Two completely independent, non-interfering windows.

8. Skills: Turn Repetitive Tasks into Shortcuts

Skills encapsulate repetitive tasks as "shortcut commands".

8.1 Skill Example: Explain Code

Create an explain-code skill:

mkdir -p ~/.claude/skills/explain-code

Write SKILL.md:

---
name: explain-code
description: Explains how code works with plain language and diagrams
---

When explaining code, please:
1. Use real-world analogies
2. Show flow with ASCII diagrams
3. Walk through code execution step by step
4. Point out common pitfalls

After this, when you ask "how does this code work?", AI will use this skill to answer.

8.2 Skill Example: Code Review

mkdir -p ~/.claude/skills/review

SKILL.md:

---
name: review
description: Reviews code, identifies potential issues and improvements
---

When reviewing code, please check:
1. Whether naming is clear
2. Whether there's duplicate logic
3. Whether error handling is comprehensive
4. Whether there's room for performance optimization

8.3 Real Example: Turn GitHub Projects into Tools

The AI era is different.

Those amazing GitHub projects used to be "visible but unusable" for beginners.

Take yt-dlp, the 143k star video download tool. You know it's powerful, but the command line and environment setup scare people away.

Now with Skills, you can tell Claude Code:

Help me package yt-dlp as a Skill

It analyzes the project, writes the wrapper, handles dependencies. In a few minutes, you have a working video download tool.

Not just video downloads. Format conversion, webpage-to-APP, password tools... all those verified GitHub projects can become Skills.

The stability of these open source projects far exceeds AI's临时 written code.

More About Skills

One skill, one purpose — that's the principle.

Think LEGO blocks: each piece is simple, but together they create infinite combinations. Better to create several "small and beautiful" skills than one "big and comprehensive" skill.

You don't even need to write code yourself. Just tell Claude Code your requirements, and it can generate the complete skill code.

There's even an official skill called create-skills (the skill that creates skills) — it asks you questions step by step, then automatically creates the skill for you.

9. Image Handling

Claude Code supports multiple ways to process images:

Method 1: Paste Directly

Press Ctrl+V in the terminal to paste a screenshot.

Method 2: Send Path

> Check this screenshot /path/to/image.png

Method 3: Project Images Folder

Put images in your project's images/ folder, AI can read them directly.

10. MCP: Plugin Ecosystem

MCP (Model Context Protocol) is Claude Code's plugin system.

One use case: Figma design to code.

After installing Figma MCP, send the design link to AI and it generates corresponding HTML/CSS code.

Not covered in detail here — search "Claude MCP" for more information.

11. Recommended Workflow: Understanding New Projects

The official documentation has a great workflow:

# 1. Enter project directory
cd /path/to/project

# 2. Start Claude
claude

# 3. Request project overview
> give me an overview of this codebase

# 4. Deep dive
> explain the main architecture patterns used here
> what are the key data models?
> how is authentication handled?

This workflow helps you quickly get up to speed with unfamiliar projects.

12. Troubleshooting

429 Error (Usage Limit Exceeded)

Usage limit reached for 5 hour. Your limit will reset at 2025-10-18 03:03:23

Means your quota is used up. Wait for reset or upgrade your plan.

Other Issues

The official documentation is well-written — check it first when encountering problems.

In Closing

Claude Code is not a "tool", it's your "intelligent collaboration partner".

It:

Understands entire project structures
Plans complex tasks
Calls various tools
Remembers context and history

From "passive response" to "active collaboration" — this is the qualitative leap in AI programming assistants.

To get started, the barrier is actually quite low:

Register for an account with your chosen model provider
Use coding-helper for one-click configuration
Create an empty folder, type claude

The rest of the journey, it will walk with you.

Related Reading:

Claude Code: It's Not a Tool, It's an Intern — Understand the core concepts
What is Vibe Coding? Or, "Conversational Programming" — Understand the philosophy
10 Minutes, I Built an App by "Talking": Conversational Programming Guide — Complete beginner's workflow
Sidebar Notes: I Built a Browser Plugin with Claude Code — Real-world case study

By Gudong, indie hacker building inBox Notes.

Claude Code: It's Not a Tool, It's an Intern

咕咚 — Mon, 26 Jan 2026 14:07:06 +0000

If you've been following me for a while, you've probably noticed that my programming tools keep changing.

From Trae to Gemini to AI Studio, I've tried many AI coding tools, sampling them like an ancient herbalist tasting herbs.

But for the past month, I've almost exclusively used one tool: Claude Code.

Friends often ask me: "Gudong, what exactly is Claude Code? How is it different from Trae or Cursor? Can non-programmers use it?"

Honestly, it's hard to explain in one sentence.

Today, from a software engineer's perspective, I want to skip the complex installation and purchasing tutorials and simply discuss: what kind of creature this is, and why I think it changes the game.

(Note: While this is a technical topic, I'll try to keep it accessible. Even if you don't code at all, you can still follow along—because what it can do goes far beyond writing code.)

1. What is Claude Code?

First, don't be fooled by the name. Despite the "Code" in its name, it's not just a coding tool.

It's an "AI intern" living in your command line.

Think of it this way:

Traditional AI (ChatGPT, etc.): Like a consultant. You ask questions, it gives advice. But you still have to do the actual work (copy-paste, modify files, run commands). It has no hands.
Claude Code: Like an intern. You give it permissions, and it actually does the work.

You can directly tell it: "Compress all images in this folder," "Analyze this project's code structure," or "Read this Excel spreadsheet and calculate the total."

It will directly manipulate files and run commands on your computer, then hand you the results.

It doesn't just "talk"—it actually "does things."

2. Why is it different?

You might ask: "Trae and Cursor can write code too. What's the difference?"

The difference lies in interaction method and permissions.

Traditional tools (VS Code, Cursor) have graphical interfaces. You need to click buttons, open files.

But Claude Code is a command-line tool (CLI). No fancy interface—just a terminal window.

Something magical happens here: because it's a command-line tool, it naturally has "god-level access" to your operating system.

It can call git to commit code.
It can call node to run scripts.
It can access files anywhere on your computer at any time.

This gives it tremendous freedom. It's no longer trapped in a software window—it directly takes over your operating system.

3. How does it "remember" things?

Anyone who has used AI knows its biggest weakness: forgetfulness. Context gets lost as conversations grow.

Claude Code has a clever approach: documentation-first.

When you start a project, running init makes it scan the entire project and write everything—what the project does, what technologies it uses, its structure—into a file called claude.md.

It's like writing an "employee handbook" for your AI intern.

Before it starts working, it reads this handbook. No matter when you return, it can immediately pick up where it left off, never losing context.

4. A Programmer's Dream: Cross-Project Operations

This is my favorite feature.

I'm currently developing both the Android and PC versions of inBox Notes—two completely separate project folders.
With traditional tools, I'd need two windows. Not only don't they communicate, but I'd have to act as the messenger.

But in Claude Code, I can tell it:

"Read the Android project's logic, and based on that, write a sync feature for the PC version."

Because it runs in the command line, it can cross folder boundaries and read information from another project.

I can even have 6 command-line windows open simultaneously:

Window A: Writing PC code
Window B: Writing product documentation
Window C: Analyzing old Android logic

This is true multi-threaded work.

Once you get started, you might only wish you had three heads and six arms to interact with multiple AIs at once.

5. Powerful for Everyone: Not Just for Code

Here's the important part. If you're not a programmer, is this useful to you?

Absolutely.

Claude Code's core logic is: "Give me a folder, I'll give you a world."

You can create an empty folder on your desktop, throw in whatever you want to process:

Throw in bill screenshots → Tell it: "Analyze this month's expenses, make a spreadsheet."
Throw in meeting notes → Tell it: "Organize into a weekly report, arranged by timeline."
Throw in materials → Tell it: "Help me plan my annual goals, generate monthly folders."

It acts like an all-purpose secretary, helping you manipulate files, organize data, and generate charts.

That's why I say it's not just a coding tool—it's a productivity tool.

6. The Hardest but Most Powerful: Skills

Finally, let's talk about Claude Code's killer feature—Skills (technically MCP, or Model Context Protocol, but let's skip the jargon).

Simply put, skills are just "shortcuts" that encapsulate things you repeatedly do.

Here's my example:
I often need AI to explain obscure technical concepts. But every time, I have to say: "Please explain simply, use real-life examples, don't get too academic..."

Too tedious.

So I created an "explain skill."

Now when I say explain 智能体 (AI agent), it automatically triggers my preset workflow:

Act as a science communicator.
Explain in plain language.
Must include a real-life example.

That's a skill.

Similarly, you can create an "upload skill" (automatically upload images to a server), or a "translation skill" (automatically translate documents into bilingual format).

Anything you repeatedly do can become "muscle memory" for Claude Code.

For space reasons, I won't cover how to create skills in detail here. If you're interested, just search "Claude Skills" and you'll find many tutorials.

In Closing

After using it for a month, I've found my workflow completely transformed.

I used to spend 100% of my time writing code.

Now, I spend half my time coding, and the other half directing it to write documentation, do research, and organize materials.

If you want to try it, the barrier is actually quite low:

Install Claude Code.
Create an empty folder.
Tell it: "I want to make [your idea]."

Whether it's building an app or making a PPT, as long as you take that first step, it will accompany you on the rest of the journey.

Related Reading:

SlideNote: A Browser Plugin I Built with Claude Code — Claude Code in action
What is Vibe Coding? Or, "Conversational Programming" — Understanding the philosophy
10 Minutes, I Built an App by "Talking": Conversational Programming Guide — A complete beginner's walkthrough

By Gudong, indie hacker building inBox Notes.

A Relaxed Saturday - Health Check, Basketball, and PC Sync

咕咚 — Sat, 24 Jan 2026 23:45:44 +0000

Weekends are always something to look forward to.

Today is Saturday. Honestly, I think Saturday is the happiest day of the week.

Because you rest and play for a day, thinking "Hey, there's still another day ahead."

Sunday is also a rest day, but as soon as you think "I have to work tomorrow," that vague sense of dread starts creeping in.

I woke up early this morning.

I had a health checkup scheduled, so I rode my bike to the software park. Most tech companies provide this benefit—an annual health check for employees. Counting the days, I hadn't done one for two years, so today was the day to catch up.

Weekend health checks are crowded. I went pretty early, but the place was still packed.

Finally got it done around 11 AM.

On the way back, riding my bike, I unexpectedly discovered a basketball court.

Beijing has many parks, and basketball courts are everywhere. Just around where I live, I've already discovered 4 courts on my own explorations.

The one I found today was in a park. I rode in to look around and spotted it. The court is beautiful, with colorful cement flooring and benches for resting.

When I arrived, there weren't many people. Two people were playing, and they had an extra ball. I asked to borrow it and shot around by myself for about an hour.

I don't play much these days—only once this week since it's too cold in winter.

For the past two years, I've been practicing my shooting form. There's been some improvement, but I'm still a practice player with a long way to go for real games.

When they were leaving, I returned the ball and rode back.

I took some photos on the way back. I really like these clean streets.

Approaching Baiwang Mountain, the road looked beautiful—the end of the road was a mountain.

After spending too much time in the city, you rarely see this kind of view in Beijing. It reminded me of Linzhou in Henan province, where the plains end and the Taihang Mountains begin.

I love this feeling, so I stopped and took a photo. At the end of that road was Baiwang Mountain.

I didn't go back to my apartment at noon. I went to the office instead.

My apartment is too cold—better to stay at the company where it's warm.

I spent the afternoon working on the PC version of inBox Notes.

Last week, I used AI to write the cloud sync feature for inBox Notes PC. You heard that right—all the code, all the product design, was done by AI.

This afternoon was mainly for testing, syncing data across multiple devices.

Watching data flow smoothly between devices was satisfying. The core cloud sync functionality is now complete, with just some detail optimizations remaining.

Oh, I also worked on something else today.

I wanted to list my browser extension (SlideNote) on the Microsoft Edge Edge Add-ons marketplace.

Previously, it was only listed on the Chrome Web Store. Users asked if I could list it for Edge too.

I tried it today—technically no problem, since the underlying engines are compatible and the code runs fine.

But when registering, I discovered that Edge developer account is paid.

Not too expensive—about 116 USD per year.

But then I thought: this is a free, open-source extension. Paying out of pocket just to list it in a store felt wrong.

So I added a note in the GitHub documentation: I won't be listing it on the Edge store for now.

If you want to use it, just download the installation package and install it in "Developer Mode." It works the same.

That was my Saturday.

A relatively relaxed day. Happy weekend everyone.

Open Source & Free: My WeChat Cover Image Generator

咕咚 — Fri, 23 Jan 2026 00:42:46 +0000

Designing cover images for WeChat Official Accounts? Gudong Cover has you covered.

1. The Origin

A few days ago, I wrote an article about how I use AI to generate cover images.

All my WeChat cover images are AI-generated, but I let the AI generate HTML code. I have my own set of prompts, so all my cover images share a consistent style.

Later, people often asked me how I made these cover images.

I thought about sharing my prompts. But I discovered that wouldn't work with DeepSeek. The prompts only work in IDE tools like Trae or Claude Code. For most non-technical users, that's too high a barrier.

2. The Solution

Then I had an idea: I could build a web app specifically for generating cover images for WeChat Official Accounts and Xiaohongshu (Little Red Book).

I just needed to bake those prompts into the app, letting AI generate HTML-formatted cover images based on the prompts.

This way, ordinary users could use it directly, solving the problem of needing IDE or programming tools.

That morning, I quickly developed this web app using AI Studio.

3. Try It Online

👉 https://cover.gudong.site/

Open the webpage, enter your article title or content, click generate, and you're done.

The interface is simple. On the left is an input box for your article content or title.

Once you enter your content or title and click generate, AI automatically extracts key information and generates a cover image in a specific style based on the prompt settings—what text to display, what background to use, and so on.

For many non-technical users who write for WeChat, you can completely use this tool to design cover images. I've also provided a download button—just click to download, or take a screenshot.

4. Dual Mode Support

Gudong Cover supports two modes, optimized for different platforms:

WeChat Mode (2.35:1)

Landscape ratio, minimalist dark style. Perfect for technical articles, deep insights, and knowledge-sharing content. Eye-catching in the WeChat feed without being flashy.

Xiaohongshu Mode (3:4)

Portrait ratio, high-saturation colors, bold headline style. Better suited for mobile browsing, emphasizing emotional appeal and visual impact. Great for recommendations and experience-sharing content.

5. Open Source & Free

This tool is open source and free to use.

💻 Source Code: https://github.com/maoruibin/GudongCover

Currently, the site uses my own Gemini and DeepSeek API keys. I don't expect too many users, so I'm providing free quota for now.

Users both domestic and international can use it. I also support custom API keys—if you have your own API key, you can configure it.

Technical Highlights:

Powered by: Gemini 3 Flash
Generation speed: < 2 seconds
One-click export to high-resolution PNG

I welcome everyone to try it and recommend it to others who might need it—friends who run WeChat accounts, friends who create content on Xiaohongshu.

If you have any feedback, please let me know. I'll continue to improve it.

6. Future Plans

My WeChat articles are all formatted using WeiMD, so I've integrated Gudong Cover directly into my WeChat formatting tool.

Now when formatting WeChat articles, you can send the article to Gudong Cover to generate a cover image, making the entire workflow much smoother.

That's the story of Gudong Cover. Thanks to AI for making this possible.

I'm Gudong. Good tools deserve to be seen by more people.

Related Reading:

24.87 Yuan in My Account

咕咚 — Wed, 21 Jan 2026 15:36:26 +0000

Hello everyone, I'm Gudong. Today I want to share my recent "massive earnings" from my WeChat Official Account 💰💰💰

Yesterday, I opened my WeChat Official Account助手 and checked my revenue: 24.87 yuan.

For me, this is indeed a proper "fortune."

1. Daily Life with Traffic Monetization

Last year, I enabled the traffic monetization feature for my WeChat Official Account.

Simply put, you add an ad space at the bottom of your articles, and you earn money when people read them.

For a long time after enabling it, my daily revenue was just a few cents, sometimes less than a yuan.

So I almost never looked at this section, and didn't care much about it.

2. An Unexpected Surprise

But seeing 24.87 yuan still made me quite happy.

As for this fortune, I have to thank Tencent. A few days ago, my article "微信输入法，越来越香了" was recommended by WeChat, and the view count is now approaching 20,000+.

This is the first time I've seen such results since writing over 300 articles. Previously, my article views were mostly in the hundreds, sometimes under 100, and occasionally breaking 1,000 was considered good.

After this article was recommended, I truly felt the power of traffic.

When the WeChat Official Account gives you traffic, it's like getting on a highway—the push-back feeling is real.

Every time I opened the WeChat Official Account助手, there were dozens of unread notifications. This was unprecedented for me.

3. A Form of Positive Feedback

I see this as a reward, a form of positive feedback.

For a long time, I wrote quietly without much recognition. The positive feedback mainly came from everyone's likes and comments, as well as my own habit of treating it as a fixed routine. But this extra monetary reward still makes me happy.

Happiness isn't about how much money—it's about the feeling of being seen.

4. Not Chasing Traffic

So will I start writing articles for traffic and hot topics?

Probably not.

I'll still focus on areas I care about: AI, programming, indie development, history, basketball... I care about quite a few things, but I won't write about something just because it's "trending."

I think we still need to focus on our inner selves—please ourselves first, then please our readers.

Writing for a WeChat Official Account means completely different approaches and directions for different authors. I think the hardest state to achieve is finding a balance between pleasing yourself and pleasing your readers—winning together. I haven't reached that level yet, but I think if I can achieve it one day, that would be pretty cool.

5. A Non-Utilitarian State

I've never liked doing things too utilitarianly.

Beyond utility, there are many things worth pursuing. Whether it's writing for a WeChat Official Account or making apps, you can actually do better in a non-utilitarian state.

Once you become too utilitarian, your actions tend to get distorted, and your eyes only focus on the results. But in a non-utilitarian state, we have more mental energy to enjoy the process. When we're less desperate for results, the results often turn out better.

24.87 yuan isn't much, but it's a signal—what you're doing is being seen, and people find value in it.

That's enough.

I'm Gudong. Slow down to go faster.

About Gudong

inBox Notes author | Indie Developer | AI Programming Practitioner

Love development, love sharing.

Here, I continuously share valuable AI practices and development insights.

Follow me to explore "AI Programming" practices and reflections.

Related Reading:

No Going Back

咕咚 — Tue, 20 Jan 2026 13:55:59 +0000

I'm an Android software developer.

Back in the early days of my Android development journey, our primary tool wasn't Android Studio as it is today. We used something called Eclipse.

I used that tool for many years. At the time, it seemed perfect: rich plugins, and the ability to develop multiple projects within a single workspace.

But what happened next proved a brutal law in the tech world — "once you upgrade, there's no going back."

It's that irreversible feeling — once you've experienced more advanced productivity, you can never return to the old ways.

1. The 2014 Transition

Old-time Android developers might remember this history.

Before 2014, Eclipse was the standard for Android app development.

Back then, I thought Eclipse was great. Shortcuts, various plugins — I used them fluently, and everything worked smoothly.

Then, Google released Android Studio, a modern development tool based on IntelliJ IDEA, and pushed the Gradle build system.

Honestly, nobody wanted to switch. Everyone felt Eclipse worked fine for development. We were used to it, and nobody wanted to change their workflow.

Plus, early Android Studio had plenty of issues: lag, slow performance, large installation size, and complicated setup.

"It's too slow," "Why did the directory structure completely change," "Gradle takes forever to compile"... Many people (including me) resisted. Some even thought they'd try it for a while and then switch back.

But management was firm: "Everyone must switch to Android Studio."

After struggling with it for six months, something remarkable happened:

Nobody wanted to go back.

Once we got used to Gradle's powerful dependency management, Android Studio's intelligent code completion, and its robust debugging tools, looking back at Eclipse felt like looking at an antique from the last century.

Why couldn't we go back?

Because the new tool wasn't just a facelift. It represented more advanced compilation logic, more efficient engineering practices — it was the "future" of that era.

2. The 2026 "Must-Have"

Twelve years later, that feeling is back.

This time, the protagonist is AI.

A couple of years ago, many people were skeptical about AI-assisted programming: "AI-generated code still needs editing," "I still need to copy and paste," "I can write faster myself," "It doesn't understand business logic."

But in 2026, if you're starting a new project:

AI is no longer an "option" — it's a "must-have."

When I develop new projects now (whether inBox Notes or new tools), I absolutely never start by hand-typing void main() line by line anymore.

That's the "stone age" approach.

My current workflow:

Feed ideas to Trae or Claude, iterate through conversations, refine product logic, and produce product documentation
Discuss technical implementation with AI, then generate the MVP (minimum viable product)
I act as the architect — handling requirements, product validation, and marketing

If you asked me to turn off Claude Code, turn off AI assistants, and write a complete app by hand today...

It would feel like asking me to download Eclipse and write Android apps in 2026.

It's not that I can't do it. It's that the efficiency would be so abysmally low that it's incomparable with modern workflows.

3. Keep the Craft, Don't Be the Mule

Of course, I must add one caveat:

"Not writing code by hand" doesn't mean "not understanding code."

You still need fundamental programming skills, algorithmic logic, and architectural thinking. These are your craft. Because AI is still a black box — it occasionally hallucinates, and you need the expertise to spot its bugs.

But the act of "hand-coding" itself has transformed from creative work to repetitive work.

Humans should focus on more creative tasks — designing products, thinking through logic, optimizing experiences. As for brick-laying and wall-building? Leave that to AI, the tireless foreman.

4. The Unstoppable Tide

This isn't just about programming.

Other fields — education, policy, tools across every industry — face the same reality.

The adoption of new things never depends on individual preference.

When a new tool is proven to dramatically boost productivity and significantly reduce costs, it sweeps through like a flood.

Over the past three years, AI development has gone from "hesitant观望" to "unstoppable."

Since there's no going back, don't look back. Embrace AI. Learn AI. Use AI well.

By Gudong, indie hacker building inBox Notes.

Related Reading:

用了就回不去

咕咚 — Tue, 20 Jan 2026 13:40:08 +0000

大家可能知道，我是一名 Android 软件开发工程师。

还记得刚开始做安卓开发的那些年，我们的“吃饭家伙”并不是现在的 Android Studio，而是一直在用一个叫 Eclipse 的工具。

那个工具我用了好多年，当时觉得它简直完美：插件丰富，一个工程里能同时开发好几个项目。

但后来发生的一切证明，技术世界里有一种残酷的定律叫做——“用了就回不去”。
一种一旦体验过更先进的生产力，就再也无法回头的“不可逆感”。

1. 2014 年的“阵痛”

做安卓开发的老同学可能都记得一段历史。

在 2014 年之前，开发Android App 基本就是 Eclipse，那是当时的标配。

那时候我觉得 Eclipse 真好用，快捷键、各种插件用得飞起，也很顺畅。

然后，Google 发布了 Android Studio，基于 IDE 开发的现代化开发工具，并推行 Gradle 构建系统。

说实话，刚开始大家都太不愿意用，都觉得 eclipse 开发项目并没有什么问题，用习惯了，都不愿意更改自己的开发环境。

而且，一开始的 Android Studio 也有很多问题，卡顿、慢，安装包大都是问题，配置起来还费劲。

“太卡了”、“目录结构怎么全变了”、“Gradle 编译一次要半天”……很多人（包括我）内心是抵触的，甚至想的，先用一段时间，然后换回去。

但当时领导很强硬：“所有人，必须切到 Android Studio。”

后来硬着头皮用了半年后，奇妙的事情发生了：

大家谁都回不去了。

当我们习惯了 Gradle 强大的依赖管理，习惯了 Android Studio 的智能提示和强大的调试工具，再看 Eclipse，就像是在看上个世纪的古董。

为什么回不去？

因为新工具不仅仅是换了个皮肤，它背后代表着更先进的编译逻辑、更高效的工程理念，它代表着那个时代的“未来”。

2. 2026 年的“必选项”

时隔 12 年，这种感觉又来了。

这次的主角，是 AI。

前两年，很多人对 AI 编程还持观望态度：“AI 生成的代码还要改，还需要复制粘贴，不如我自己写得快”、“它不懂业务逻辑”。

但到了 2026 年的今天，如果要开始一个新项目，那么：

AI 已经不再是一个“可选项”，它是一个“必选项”。

我现在开发新项目（无论是 inBox Notes还是新工具），绝对不会再从 void main() 开始一行行手敲了。

那是“刀耕火种”时代的做法。

现在的流程是：

把想法喂给 Trae 或 Claude 不断交流，确认具体产品逻辑，输出产品文档
跟 AI 交流确认技术实现方案，生成 MVP（最小可行性产品）。
我作为架构师去做需求管理，做产品功能验证以及市场推广。

如果现在让我关掉 ClaudeCode，关掉 AI 助手，让我纯手工去写一个完整的 APP。
这感觉，就像你让我在 2026 年，重新下载一个 Eclipse 去写安卓一样。
不是不能写，而是效率低到令人发指，是完全跟不上时代节奏的降维打击。

3. 手艺要有，但别做“苦力”

当然，我必须补充一点：
“不手写代码”不代表“不懂代码”。

基本的编程技能、算法逻辑、架构思维，这些手艺你必须保留。因为 AI 还是个黑盒，它偶尔会胡说八道，你得有能力一眼看穿它的 Bug。

但是，“手写代码” 这个动作本身，已经从一种创造性劳动，变成了重复性劳动。
人类应该去干更有创造力的事——去设计产品、去思考逻辑、去优化体验。至于搬砖、砌墙这种事，交给 AI 这个不知疲倦的包工头就好。

4. 无法阻挡的洪流

其实不止是编程。

其他领域，包括教育、政策、各行各业的工具迭代。

新事物的普及，从来不以个人的意志为转移。

当一个新工具被验证能大幅提升生产力，能极大地降低成本时，它就会像洪水一样席卷而来。

AI 发展这三年，已经从“犹豫不决”到了“势不可当”。

既然回不去，那就别回头了，拥抱 AI 学习 AI，使用好 AI。

关于咕咚

inBox Notes 作者 | 独立开发者 | AI 编程实践者

爱开发，爱分享。

在这里，持续分享有价值的 AI 实践与开发感悟。

关注我，一起探索「对话式编程」的实践与思考。

夜间模式展开收起优化：SlideNote 更新日志

咕咚 — Mon, 19 Jan 2026 14:50:27 +0000

在写今天的文章之前，先说个小插曲。

昨天写了篇《微信输入法越来越香了》，是有感而发。

这段时间经常用语音输入，电脑上用闪电说，但发现总有些小问题。

后来发现微信输入法在手机端的语音转文字特别准，说完会有个 AI 校正的过程。

就因为这一点，我觉得微信输入法真好。有感而发，就给"打了个广告"。

没想到这篇文章被微信推荐了，现在 1 万多阅读，成了我阅读量最高的一篇。

反正想来想去，微信怎么都不亏，自家推荐自家的产品。

今天要说的是 SlideNote 的更新。

SlideNote 是我上周发布的 Chrome 浏览器侧边栏笔记插件。具体可以看这篇：侧边笔记，常伴左右：我用 Claude Code 开发了一个浏览器插件

发布后数据比我预期的好：GitHub 80+ star，Chrome 商店 400+ 下载。

我之前给 inBox Notes做过一个浏览器插件，到现在才 40 多个用户。

SlideNote 用的人多，可能是因为这个需求确实存在。

GitHub 和微信上都收到一些用户反馈，所以这周修了几个 bug，加了些新功能，发布了 0.0.6 版本。

1. 夜间模式

之前一直是浅色，浏览器变黑后它也不变黑，有点突兀。

这次适配了夜间模式，会跟随浏览器的主题自动切换。

2. 侧边栏交互优化

展开收起的交互做了调整。

这个改动是和 AI 讨论后，它给了一个方案，我觉得挺好。

现在展开收起的按钮放在顶部。点击收起后，侧边栏会展示笔记的名称，再次点击就展开。

3. 后续计划

还有几个功能在计划中：

划词收藏：选中网页文字，快速收藏
数据导出：备份自己的笔记

暂时先放到后面的版本。

如果你也有在浏览器里记录临时信息的需求，欢迎试试 SlideNote。

有其他想法也欢迎反馈。

我是咕咚，好工具，值得被更多人看见。

关于咕咚

inBox Notes 作者 | 独立开发者 | AI 编程实践者

爱开发，爱分享。

在这里，持续分享有价值的 AI 实践与开发感悟。

关注我，一起探索「对话式编程」的实践与思考。

WeChat Keyboard Keeps Getting Better

咕咚 — Sun, 18 Jan 2026 14:18:53 +0000

I recently discovered something surprising: WeChat Keyboard's voice-to-text is actually better than the dedicated desktop voice input tool I've been using.

A Pleasant Discovery

I've been using a desktop voice input tool called ShanDianShuo for a while. It's an AI-powered input method that uses local LLMs for voice-to-text conversion. Initially, I thought it was great, but over time I noticed various issues:

It couldn't remember words I frequently used, like inBox Notes. Mixed Chinese and English recognition was often inaccurate. I frequently had to manually correct the output.

But after WeChat Keyboard's recent upgrade, all these problems were solved. Even when I don't speak clearly, it accurately transcribes my words.

I found out that last December, WeChat Keyboard released an update with significant AI optimizations for voice input. I can genuinely feel the improvement in transcription accuracy.

Using it is simple: just hold the spacebar on the WeChat keyboard to start voice input.

WeChat Keyboard is Now TopTier

I've actually recommended WeChat Keyboard before.

Looking at the current landscape, it has truly become the number one input method in China. It has quietly surpassed both Sogou and iFlytek former leaders where iFlytek was known for voice recognition and Sogou for user experience. I've seen user experience surveys placing WeChat Keyboard firmly in the first tier.

I've even recommended it to my family my wife, my dad, and my mom all use it now.

After recommending it to my dad, his favorite feature turned out to be the emoji integration: when typing in WeChat, entering words like haha or really automatically suggests relevant emojis that can be sent with one tap. He loves using this feature when replying in group chats.

As for me, my favorite feature is crossdevice copypaste.

Text copied on my phone can be pasted directly on my computer using WeChat Keyboard, and vice versa. As long as both devices are on the same network, this feature works seamlessly.

Another plus: WeChat Keyboard has no account system. No login required, no cloud sync to configure. Many people dislike being forced into account systems on their phones. With WeChat Keyboard, you just use it simple as that.

One Wish: Voice Input on PC

As great as WeChat Keyboard is, there's one regret: voice input only works on mobile, not on desktop.

If the desktop version supported voice to text, I'd switch to it immediately.

The demand for desktop voice input is strong. Many developers, engineers, and creators today have a greater need for voice input than typing. I hope the WeChat Keyboard team will add this feature to the PC version soon.

From a technical perspective, it shouldn't be too difficult it's already implemented on mobile. I'll be among the first to use it when desktop voice input becomes available.

I'm writing this hoping the WeChat Keyboard team sees this need.

Good Tools Deserve to Be Seen

WeChat Keyboard has no ads, feels lightweight, and is hard to switch away from once you're used to it.

With the recent AI optimizations for voice to text, it's gotten even better. I recommend everyone give it a try.

By Gudong, indie hacker building inBox Notes

当 AI 变成我的"超强员工

咕咚 — Sun, 18 Jan 2026 00:28:08 +0000

大家好，我是咕咚。

我最近用 AI 用得特别多。

买的是智谱的套餐，基础版。

一开始买了一个账号，每天使用额度有限，用着用着就超额了。超额之后这天就不能用了，得等两三个小时才能恢复。

后来干脆买了两个账号，一个超额了就换另一个。

但有趣的是，我的 AI 使用方式，这段时间发生了一个明显的变化。

之前很长一段时间，我用 AI 主要就是写代码、开发项目。做 inBox 笔记也好，做新软件也好，基本都是这个场景。

但这阵子，我发现自己越来越多的时间，开始用 AI 做一些非编程相关的工作。

分享几个我最近的实践。

1. 让 AI 跟审核员"吵架"

上次我写了篇文章《我让 AI 帮我跟审核员"吵架"》，讲怎么让 AI 帮我回复小米应用商店的审核邮件。

inBox 商家版上架小米市场，被拒绝了。我就把审核要求、拒信内容、我的情况全部整理好，喂给 AI，让它帮我写回复邮件。

它帮我回了两次。

后来还有人问我，审核通过了吗？还没有，小米那边还是不让上架，说要做备案。这个我暂时没继续折腾，但这个过程让我意识到一件事：

AI 可以帮我们处理很多与人交涉的事情，尤其是非面对面的、文字形式的内容，比如邮件、申请书。

关键在于：要给 AI 足够的上下文。

对方的要求越具体、越详细，我们就可以把这些信息完整地给到 AI，AI 才能更好地处理。

没有上下文，AI 就是瞎猜；有了上下文，AI 就是你的代笔人。

2. 用 AI 做述职 PPT

年初做年终述职，每个人要准备一个 15 分钟的 PPT，讲这一年做了什么。

我也是用 AI 做的。

怎么做呢？先把自己这一年做的事情列出来，越详细越好。我写成了一个文档，然后把这个文档给 AI，让它帮我生成 PPT 大纲和内容。

之前我写过文章《免费用 AI 创作 PPT，用这款软件就对了～》，讲用 NotebookLM 做 PPT，核心逻辑是一样的。

这个过程让我意识到：要让 AI 帮你输出，你平时要刻意记录信息。

通过文档的形式，把日常的工作、成果、数据记录下来。等到需要用的时候，才能把这些信息"喂"给 AI，让它帮你分析整理。

没有记录，就没有输入；没有输入，就没有输出。

3. 改造开源项目做信息机器人

我还用 AI 做了一个信息自动化工具。

我关心篮球，也关心 AI，所以我做了一个机器人，每天定时爬取这些信息，生成简报推送到飞书。

这个项目不是我从零写的。我是找了一个开源项目，在它的基础上做二次改造。

改造整个过程都是让 AI 帮我忙的。说实话，我对自动化流程这东西一知半解，但我知道大概是什么回事。

我把项目下载到本地，让 AI 帮我分析项目结构，然后告诉它我的诉求，让它帮我改代码、加功能。

这其实是一个很重要的能力：如何把别人的项目拿来改造成自己想要的样子。

用 AI 辅助，你可以快速理解一个陌生的项目，然后按自己的想法去定制。

开源世界有那么多好东西，不用白不用。

4. Chrome 插件上架：全流程 AI 辅助

前两天我发布了 SlideNote，一个浏览器侧边栏笔记插件。

开发过程是 AI 辅助完成的，这个不意外。但有意思的是，后面的上架流程——填写表单、准备材料——也全部是 AI 做的。

Chrome 应用商店上架需要很多东西：图标、封面图、介绍文案、推广图，而且每个都有特定的尺寸要求。

我把 Chrome 商店的要求全部整理给 AI，让它帮我生成符合规范的材料。

比如 128x128 的图标、440x280 的宣传图、还有简短描述和详细说明。AI 生成的图片尺寸精准，文案也符合商店的风格。

我在文章《侧边笔记，常伴左右：我用 Claude Code 开发了一个浏览器插件》里详细介绍了这个过程。

这就是 AI 作为"超强员工"的另一个场景：它可以帮你搞定那些繁琐、但又必须按规矩来的事情。

5. 设计类工作

我公众号的所有封面图，都是用 AI 生成的。

这个也花了不少 token。

还有做 inBox 笔记 PC 端的时候，整个 PRD、设计图、设计方案，都是让 AI 帮我生成的。

这些事情，我以前是完全不擅长的。

现在有了 AI，我可以让它去做这些我以前做不了、或者做不好的事情。

6. 角色升维

总结一下。

AI 对于个人开发者、对于像我这样的"一人团队"来说，可以做我以前能做的，也可以做我以前不能做的。

它让我完成了一种"角色升维"。

以前我可能只是一个写代码的。现在我可以把自己放到一个全局统筹者的视角：

方案设计，AI 可以做
UI 设计，AI 可以做
开发，AI 可以做
测试，AI 可以做
运维，AI 也可以做

人在这个过程中，更像是一个验证者、一个传话筒。

验证 AI 的方案是不是 OK，把用户的需求接收到、反馈给 AI，让它继续调整。

但这么说，不代表人的价值变小了。

恰恰相反，人的价值、人的观点、人的价值观变得更重要了。

产品的灵魂，不是 AI，是人。

AI 只是一个工具，一个超强的"员工"。它能做出什么样的东西，取决于你的灵魂、你的价值观、你的审美。

你是老板，AI 是员工。

做好决策，把控方向，剩下的，让 AI 去做吧。

我是咕咚，用好 AI，从今天的一个小场景开始。

关于咕咚

inBox 笔记作者 | 独立开发者 | AI 编程实践者

爱开发，爱分享。

在这里，持续分享有价值的 AI 实践与开发感悟。

关注我，一起探索「对话式编程」的实践与思考。

排版 by weimd