I don’t know how you spend this boring time on the way to and from work. I used to watch plays and listen to stories. But after a long time, I found that my brain refused to stop. I often listen to stories or watch dramas, but my mind doesn’t know where it has gone. In the end, I don’t know what I am listening to or watching.
Later, I liked chatting with AI. So this idea that was a bit boring but self-interesting was born on the way from work on Friday.
##First challenge: AI’s “unreliable” advice
At first, AI’s feedback was relatively traditional. It recommended that I deploy my own model to train in recognizing rock, paper, scissors. Take 50 photos of stones, 50 photos of scissors, and 50 photos of cloth. Different angles, different sizes, and preferably different hands to train the accuracy of model recognition. I suddenly grew two at a head. If I had this ability, would I still ask you to help me program? Wouldn’t it be nice if I went to ByteDance to find a job?
##Lightweight: The wonderful use of Cloudflare Worker
** Light application, the most important thing is to be light. ** At first, I also wanted to use my own Cloud Virtual Machine to deploy Python as a backend. But I just happened to apply for Cloudflare to host my own domain name. On second thought, why not use Cloudflare’s worker to deploy JS to implement my idea? As an experimental game that won’t be played by many people, isn’t this the perfect choice to save money?
##Hard work: First experience of QWen3 Coder
Let’s do it as we say. Just as QWen3 Coder was released, there were praise and criticism online for a while. I dare not install Qwen3 Coder locally for fear that my bank card balance will be cleared by accident. So, I put my needs in the web version of QWen3:https://chat.qwen.ai/ My original goal was simple:
An html page and open the camera. Click a button and take a photo and send it to the backend javascript.
A javascript receives incoming images and asks an AI model to identify them. In parallel, let another AI model decide whether stone, scissors, or paper.
Regardless of strategy and fault tolerance, run the process through first.
To deal with such a small demand, the Qwen3 Coder looks like a knife to kill chickens. The code is generated by three steps and five divides by two. He also treated me like a new melon egg, instructing me on how to upload to Cloudflare’s Pages and Workers, and then how to connect the two. Don’t tell me, I really need it. Because I really don’t know how to.
##Troubleshooting Journey: Collaboration between Claude Code and Gemini CLI
Generating code is one thing, making it run is another. Since I have no concept of coding, I just used the configured Cluade Code and Gemini CLI to do this job. In this debugging journey,_Claude Code is absolutely responsible. It is indeed more efficient and accurate in correcting an error than Gemini CLI. The price is that the token costs money, but the price of Kimi is affordable.
The convenience of using Claude Code is that it will modify the code itself with your permission. If it is a locally developed application like python, it can also call the shell to run it, grab error messages by itself and continue to analyze it. Perfect closed loop.
The trouble with my current debugging is that I can only upload the modified code to Pages or woker.js over and over again, and then refresh it to see the effect. Efficiency is really not high. Of course, you can install a miniflash to run it locally. But I find it troublesome. Anyway, Cluade Code will regenerate the code. I can endure the work of copying/pasting to refresh the page.
###The “obstacles” encountered
A big piece of stone stuck before it started running was that I used SiliconFlow’s free model THUDM/GLM-4.1V-9B-Thinking for Gesture Recognition. But in order to ensure that there is no randomness in the recognition of pictures, the temperature is set to 0 in the code. Then the call to the API kept returning 400 errors. I don’t understand anything, but it seems that Cluade Code and Gemini CLI are equally equal in this regard. I thought of a little trick, and I asked them to write a Python locally, using the same call parameters. Then can’t you be able to correct errors locally in closed loop? At this time, the strengths of Claude Code were reflected, and it actively tried to change the model. Changing to Pro/Qwen/Qwen2.5-VL-7 B-Instruction runs the recognized process. Good accuracy.
The second small stone is that I chose a reasoning model to make the decision to punch, because I personally feel that if I choose MOE, the same prompt word will lead to the same gesture every time I return. But as a result, reasoning’s model opens up a thoughtful model, directly lengthening the feedback time to worker timeout. I still wrote a Python program to simulate the process in JS, and found that this Reasoning model really thought a lot. I thought in my mind for a moment that I should choose a fist, and for a moment that it should not be that simple. I want to think again. So he decisively revised the prompt and told it not to think too much and answer immediately. The speed was immediately shortened to less than 10 seconds. After testing it multiple times, it will not always be the same choice.
##Improvement and Outlook
The work after running through is improved. I started with the UI first and changed it to a retro arcade style. It looks interesting, a bit like dreaming back to the 1980s. ! AI Screenshot
Then I want to tune the logic of AI punching. Because it’s completely random now, I try to bring in past records and let AI make decisions. However, it was found that this greatly increased the thinking chain of the Reasoning model, and the feedback speed even reached 1-2 minutes. So I went back to my original random pattern. But no matter what, this step is already a game.
##Write at the end
The emergence of AI, especially Claude Code, has greatly facilitated ordinary people like you and me to write small applications that we need in our lives and work. Each of us should give it a try. It will bring great convenience and fun to our life and work.
I want to set a small goal for myself and challenge me to complete 100 boring but interesting light applications. Do you have any suggestions?