Development Diary with AI

Additional improvements I have recorded the improvements made since Part 3. Improved database structure The first problem I felt after deployment was the lack of scalability. At first, I directly entered the category values into the channel as enums, but it was very inconvenient every time I added or changed a category. Separating the category table makes management much easier. Pitfalls of Image Collection Logic When we kept asking the agent to improve the video collection part, errors kept occurring instead. In particular, when trying to make imperfect improvements like a human, there were many cases where recent videos from certain channels were omitted. Eventually, I went through a process of manually reviewing and correcting all channels to ensure they were imported normally. As I was digging into something that wasn't working, I realized that there were some useless parts of the code that were similar to people. Cost Optimization: Switch to Gemini The task of extracting scripts from audio files has been changed from the existing service to Gemini. I thought that only OpenAI provided it as an API, but Gemini also provides it by default. After a quick price comparison request to Perplexity, I switched to Gemini. The price was a bit cheaper, and the switching process was faster than I expected. It also seemed to be a bit faster (feeling?), so it was a satisfactory decision. First public appearance and reaction Achievements at geeknews Thankfully, when I shared it with geeknews, I got a really positive response. We got 1,300 hits and 60 signups in just one day. The next day, I visited over 400 times and enjoyed the 2-day DAU goal of 100. I received a request in the comments to add RSS XML functionality, so I immediately created an rss.xml file and set it to automatically update whenever a new feed is added. https://shortv.replit.app/rss.xml Thank you to all the geeknews people who sent us so much support :) The analytics provided by Replit are good for viewing, but they don't provide enough data to analyze. I think we'll have to add GA in the future.

Continued in Part 2. Additional improvements After Part 2, I recorded the minor improvements made. Improved usability and added features + Mobile view compatible After deploying and testing, we started noticing usability issues on mobile devices. Since we had to view it on a mobile web browser anyway, we simply hid the search box for the mobile Safari PWA . + Added slide sweep action motion In the process of improving the Carousel feature in the Trending section, we experienced the limitations of Replit, which was unable to capture more detailed parts than expected . Key issues: * Issue where the width of the item card changes depending on the slide button * Problem with the position where the card stops after slide motion being misaligned After about 7-8 revisions, we were able to agree on the current state. The sweep action motion modification in the mobile view was also less satisfactory than expected. + Add slide filter I attached the website and code for reference and asked for it, and this part was surprisingly easy to solve. I was able to confirm once again the effectiveness of presenting existing services as references. Prompt Summary: * Made identical to Substack's slide filter button * Attached code for the slide filter button part + Added logic to exclude live video As we were running it, we realized that XML also included information on live and upcoming videos . Since these videos were not subject to summary, filtering was necessary. After asking Perplexity, they confirmed that this could be solved via the YouTube Data API, and improved the logic to only retrieve registered regular videos using the YouTube Data API . + Added 42 channels In addition to the channels I personally mainly watched, I added channels from various fields through searches and recommendations. In the process, I realized that the boundary of “information” on YouTube is more ambiguous than I thought . There were many cases where I added and then deleted channels, and I felt the need for continuous curation .

Continued from Part 1. Deploying Improvement Work In the previous article, I introduced the process of creating an MVP based on Replit. After that, I made some improvements and summarized the process up to deployment. Improved usability through UI modifications + Improved page structure We completed the implementation of core functions and made modifications that took user accessibility into consideration. Previously, the feeds I collected were a closed structure that only I could see, but now we have added a page that can be shared and viewed together. Prompt Summary: * Changed from existing Home → Feed * New Home page dedicated to recommended videos - Only display videos with at least 1 upvote - Sort by latest recommendations * Viewable without logging in * When you first access, go directly to this page As a result, the site's character has become clearer, with the first screen shown to users changing from a feed to a curated selection of recommended content. + Add channel button Since a public page was created, we thought there would be an action to “agree with other users’ choices.” We concluded that Upvote was the first function and adding a channel was the second. Prompt Summary: * Placed in the same style next to upvote * Click to add to user's channel list and add related videos * Disable button if already added * Deletion is performed on the channel management page. This feature will serve as a foundation for naturally building a user's activity history while improving the video curation experience. + Add trend slide At the top of the new homepage, we've added a trending area that features recently popular videos. Prompt Summary: * Conditions: Within the last 7 days + Most upvotes + In case of a tie, based on number of views * 8 cards in total, 4 can be slid at a time * Leave space empty if not available * If there are less than 4, the slide button is disabled.

This is the next time I used Replit after MGX.dev and Lovable. To begin with Replit This project used Replit with the goal of creating a quick MVP. I had seen Replit's founding story in the past and tried a simple project with Free Tier, but this time I got to try some pretty serious API integration and LLM utilization. I was thinking about paying for Lovable again, but I found a discount code and paid right away. Setting the topic I didn't want to repeat a topic I had already done, so I set a new topic and started: building a YouTube summary service . This goal is to build a YouTube summary service . We created a service called shortV . When you open YouTube, you tend to get lost in the algorithm swamp, leaving behind the videos you originally intended to watch. To escape this 'time-wasting' and to quickly select and view only useful videos, I created a service called shortV . Introduction to services included in the prompt "It provides AI summaries by subscribing to videos from YouTube channels that users specify. It gives users the freedom to select videos they want to watch through previews before watching them, and even goes beyond the YouTube platform itself." Setting Targets and Goals ShortV is a tool for those who want to consume informative content efficiently. It provides a summary based on the video script, and the goal is to generate timestamps for each topic so that you can immediately check only the sections you need. Different from before UI design attempt

This article is based on the content recorded through chat with GPT while working. Trying to build a web service from scratch is always interesting, but it is also a complex and cumbersome task. In particular, since you have to handle various technology stacks such as front-end, back-end, and infrastructure, there is a high barrier to entry for individual developers or initial planners. Among these concerns, the platform I recently discovered is MGX.dev. Here, AI-based multi-agents divide roles like an actual team and develop web services according to user requests. Thanks to this, I was able to approach everything from planning to design and implementation much more easily. In this article, I will briefly share the outline of the project I attempted through MGX.dev, the development process, and my thoughts on it. Project Overview The service I planned has the following goals: "An article-based RSS feed curation service that gives you insight" Hand-curated selection of trusted blogs and news sources Using LLM, summarize the core of each content in 3 lines in Korean and English Provides key content in a topic-centered, core-organized format Lowering language barriers with machine translation Providing information to help users think critically and make independent decisions The project combines a basic RSS reader with AI summarization and language translation capabilities, with the goal of helping users quickly identify and select only the essential information in an age of information overload. Defining core functions As of version 0.1, the following key features have been established: RSS Aggregation: Collect RSS from specific websites Storage and Processing: Store the collected feeds in a database. LLM Summary: Generate a 3-line summary and key points in Korean/English for each feed

This article is based on what I recorded in a chat with Gemini while working. I would like to share my experience of developing an RSS-based web service using Lovable, an AI development platform. This project, which is integrated with Supabase, GitHub, and OpenAI API, is more than just a side project; it is a good example of the potential of development automation using the latest tools. Why Lovable? Lovable is a cutting-edge AI-based development tool that has the following advantages over existing Cursor and Windsurf: Direct integration with Supabase and GitHub: The setup process is automated, so developers don’t have to do complex connections themselves. API keys, table settings, and deployment processes are all automated. Low barrier to entry: You can easily create prototype-level web services without separate backend knowledge. Automatic setup based on up-to-date information: Eliminates the lack of documentation that previous tools had and provides seamless integration with platforms like Supabase. Summary of development flow 1. Start the project: Implement a basic web service UI in 2 minutes After running Lovable and connecting to GitHub, the basic UI structure is automatically configured as soon as the project is created. The home screen in the form of a list of posts is set up, and it only took 2 minutes to get there. 2. Supabase connection and data structure creation The Supabase project is automatically created and Lovable automatically connects it (basic account creation is required). After that, Lovable automatically inserts sample data and a table (SQL-based) for storing RSS data, completing the database configuration. This part was very attractive to me. Previously, when using Cursor and Windsurf, I had to set it up myself, but for me, who was not a developer, it was more difficult than I thought. However, Lovable solves the Supabase and DB settings just through chat, so it is very helpful for beginners. 3. RSS feed integration Since the first prompt request already had the basic RSS parsing logic in place, I was able to fetch the feed right away by simply entering three website URLs. Sample websites: Paul Graham Blog , Sam Altman Blog , Flarup Blog From here, the structure for saving feeds to Supabase in real time and outputting them from a web service has been completed.