"Independent Development Diary 8: Are the Three AI Products I Created Starting to Gain Traffic?"

Personal Life

I did a lot of things this week, and I've already forgotten what I did. My body really can't take it anymore, so I made a trip to the hospital and spent over 10,000 Thai Baht for allergic rhinitis and bronchitis. Fortunately, I have insurance; otherwise, I really wouldn't be able to afford medical treatment. I think the air conditioning here is set too cold; in schools, it's set to 22 degrees Celsius, and Thai people don't feel cold at all, which is quite amazing. Plus, it's the rainy season now, and it rains a lot, so the temperature is already much lower than usual.

Recently, I've started to seriously learn about operations. Once I got involved with traffic generation, I saw many fascinating things.

Observation Position

I learned what an observation position is. It means someone is starting from scratch to create something, and you can observe them closely while they do it. One person charged 599, but later said it was too popular and got reported, so they only accepted over 20 people. Then, posts appeared about how to earn over 10,000 in one hour. As for what happened afterward, I have no idea.

Many people might think that if he can't deliver, then he can't fulfill his promise. However, in reality, many people impulsively forget about it afterward. Some feel that causing a rift over a few hundred Baht would affect their standing in this circle, and later, the big shots wouldn't want to work with them anymore. Back in 2020, I joined such a group, and after a while, there was no delivery, and then I saw the big shots working on other groups, like a learning delivery group that was popular at the time.

I Replicated XX

Recently, Grok has become very popular, and they launched a virtual girlfriend character from the anime world, with whom you can chat. Someone made a video using Jimo and open-sourced it on GitHub, claiming they replicated Ani, which gained a lot of exposure. Another case is that I replicated a Douyin (TikTok), which was essentially just recreating some of Douyin's interfaces, and the wallet could input numbers, making it seem like you earned a lot on Douyin. After being criticized by serious people, they claimed it was a replication and started telling stories about how they made millions on Douyin.

First of all, I don't really want to comment on this behavior since I wasn't involved; what if it's true? However, I have noticed that many people involved in traffic generation are genuinely serious about it. From a money-making perspective, they indeed earn more than those who are genuinely developing independent products. But I still can't achieve that through unethical means.

I told a friend about this, and he said that a lot of knowledge seems worthless to you just because you understand it. For example, replicating a virtual character—someone who doesn't understand would learn what Jimo is and how to create videos, which is valuable knowledge. You know too much, so you think these things have no value.

After thinking about it, I realized this is something I need to reflect on. If I want more people to see it, I need to make the content simple enough. But I still hope to be more sincere.

StickerAI

Last week, StickerAI underwent many iterations. First, there was a comprehensive SEO adjustment for the entire site, adding many pages that I previously thought were unnecessary because I wanted Google to crawl the pages. However, I still haven't figured out Google Search Console, as the website address keeps showing invalid crawl.

The lab added a text-to-sticker generation feature. If someone purely wants to generate expressions from images, it’s also possible for some to create IP characters from scratch. Another significant reason for this addition is that when you search for Sticker AI, most of the results are related to this feature's SEO. So if I can get indexed, it will likely increase traffic. As for keywords, I haven't learned that part yet.

Additionally, the lab added an image gallery. Previously, there was an issue where images generated in the lab couldn't be retrieved once you left the page. Now, with the gallery added, all images from the lab can be found there.

Also, the text-based expressions now support the Indonesian language. As for why Indonesian is supported, I will explain that later. Furthermore, we added PromptPay (a local payment method in Thailand). I will also discuss this later, and many small optimizations won't be mentioned separately.

When I was creating the gallery, I used production data because there were no local images, and I ended up seeing many sticker packs with headscarves. I researched a bit and asked GPT; based on usernames and Google statistics, I basically confirmed that these users are from Indonesia, so I added an Indonesian sticker pack option. The reason for adding Thailand's local payment method originated from my discovery that the website often had some Stripe statistics, but when I checked Stripe, I didn't see any orders (including failed ones). Initially, I didn't pay attention, but later I found it happening frequently, so I pulled data from Google statistics and discovered that this data came from Thailand (not excluding errors due to local testing). So, I decided to add a Thai payment option to test its authenticity.

There are still many things I haven't figured out, such as whether to create a quick import feature for sticker packs, like for WhatsApp and Instagram. Also, there's the idea of inviting users to register for points and adding a profit-sharing system, along with an advertising comic feature that I've always wanted to implement but feel like no one would use.

However, this week, I should pause the new feature iterations for StickerAI and continue to improve SEO, considering external links, and I might need to write some articles, etc.

PromptPlan

From my personal perspective, most of my daily planning has shifted to PromptPlan. However, I'm still considering whether to add a regular to-do feature because sometimes I just want to quickly jot down a task temporarily and plan it elsewhere later.

Another thing is the iOS version. Sometimes I come across information or ideas on my phone that I want to turn into tasks for later processing. Currently, I just throw them into Notion, but sometimes I forget. If there were an iOS version, it would be convenient, but creating an iOS version might require technical migration, such as using Tarui to completely replace SwiftUI for multi-platform support, which is quite daunting to think about.

Additionally, PromptPlan has added some small features, such as recently added tasks, today's tasks, and the ability to set completion times for tasks. You can also directly delete projects, and some conventional to-do features will be re-integrated. I also optimized the language check; it no longer relies entirely on AI for judgment but preprocesses user input. If the Chinese content exceeds 5%, it is considered Chinese; otherwise, it is English. The prompts given to AI now explicitly specify the output language.

Currently, the issue with PromptPlan is that there is no user feedback. The few paying users I have never seem to complain. Those who come in asking for everything to be completely free seem to be a bit annoying, so I just block them. The core features are already free; you just need to fill in the API key. Foreign users see that you are continuously iterating and leave supportive comments; I haven't seen anyone come in to criticize.

The problem I'm facing now is that as tasks accumulate, some may have changed priority, resulting in a screen full of high-priority tasks. I might consider limiting the number of high-priority tasks in the prompts and breaking down priorities into five finer levels, P1-P5, which would help manage multiple tasks better.

Another thing is that the task automation isn't as good as I imagined. Currently, creating tasks, updating statuses, and deleting tasks basically meet expectations. However, when I told AI that there are too many tasks now and asked it to reorganize, tasks that were originally part of one project might get split into three separate projects. My intention was more about readjusting priorities and merging tasks, but it seems the semantic understanding isn't very good.

One solution I thought of is to stop adding so many tasks each time I plan. Instead, I could treat each planning session as one task and create sub-tasks within that task. This way, it would effectively resolve the issue of whether to create one task or multiple tasks, as AI might have a different understanding than us unless I can input it very clearly.

Another thing is to reduce the burden of prompts. Currently, each planning session involves about 7,000 tokens for input and 1,500 tokens for output, which is quite heavy for AI, causing each planning session to take about 30 seconds, which doesn't meet my expectations for this product.

TraderAI

Yes, I’ve started a new project again. However, this product is not aimed at users for now and may be open-sourced directly after completion. This product mainly aims to use AI for quantitative trading. As someone who doesn't understand quant trading or investing, I'm curious whether AI can outperform the market. I started this project because my investments in recent years have been a disaster, losing a lot.

Currently, the program will fetch some technical indicators, long-short ratios, fees, and situations of large investors, and then let AI analyze whether to go long or short. During backtesting, the data looked quite good, but once it went live, it became a mess, with many unexpected situations not considered. I'm still patching and testing it. I can't share too much for now; I will formally introduce it once it's open-sourced.

Here’s a backtest result chart.

I also welcome everyone to communicate with me about everything related to independent development, and I especially encourage everyone to download and use my products 😄.