크로키의 잡도리

크로키의 잡도리

Follow

Posts

크로키의 잡도리

Blog Automation (2) Looking at Pipeline 3, I concluded that "Generating images with Gemini 3.1 Flash" is the best approach, and I am currently using it. My original plan was: 1) Download and use images from the article body. 2) As a fallback, use Vertex AI Search to search for and use relevant images from English news sites. However, perhaps due to limitations in the script, there was a limit to the number of article images I could download. When I ran it through Vertex AI Search, I only got a bunch of images unrelated to the article's topic. (Google has a custom search API, but it is scheduled to be discontinued soon, and new users will not be able to use it..) In the end, I switched to Gemini 3.1 Flash Nano Banana and am using it with just enough satisfaction, though I am still looking for a way to save the original article images..! » Writing Quality Honestly, humans can usually tell immediately when reading text written by AI. There were many repetitive words and the endings were consistently "~합니다". I didn't know what to do, so I just asked Claude to analyze 16 actual popular blogs and organized the rules. I set up a mix of five different endings, specifying the ratios, and made it regenerate if the same ending appeared in two consecutive sentences. I put AI clichés like "I will look into it" and "I will find out" into the block list. Honestly, the catch is that creating these rules took more time than writing the logic for writing articles... I'm still not entirely satisfied with it, but I think continuous modifications will be necessary. » Semi-automation is the answer for Shorts/Reels. While full automation is technically possible, once a video is uploaded, editing it is difficult and bothersome. Also, I felt there was a significant gap in quality between semi-automation and full automation, so I went with semi-automation. I configured the pipeline to generate narration using Gemini TTS, extract word-by-word subtitle timing using Whisper, and then create a project JSON file. As mentioned above, the Telegram bot sends me DMs while running the pipeline. Among them is an "Edit" button; clicking it opens the DesignCombo web editor and automatically loads the project. (I am currently using DesignCombo as the web editor because I found it to be the best option while searching on GitHub!) The structure is such that after fine-tuning and rendering there, clicking the upload button publishes simultaneously to YouTube Shorts and InstaReels! One tip I can give you is to make sure you create a basic format! You need to set up the fonts, image placement, text placement, background image templates, etc., in advance so you don't have to repeat the same tasks over and over. » So, is it possible to monetize a blog? Honestly, I haven't been doing this for very long, so I'm not entirely sure, but based on the structures I've looked into, it seems to be structured in three branches: connecting to AdSense on Blogger, Coupang Partners Affiliate, and YouTube monetization. Of course, monetization requires a lot of viewers, but for now, I'm just trying it out, considering the satisfaction of having done one more thing as my reward... haha. However, since the structure involves consistently accumulating four videos a day, I think meaningful traffic will start coming in once the content exceeds 100 episodes! I believe one-person media is ultimately a game of volume, so I'm considering mass-producing content by modifying the Shorts/Reels templates and editor. » Lastly, for those who want to try doing the same thing: If I may offer just one piece of advice to those looking to build a similar system, I recommend simply using Claude Cowork if you only plan to post on Blogger or your own site. However, if you also want to upload videos to Korean blogs, YouTube, and Instagram, I think it would be best to request your LLM to design a system using a combination of Task Scheduler + Python + Claude Code CLI. My SoloMedia project isn't finished yet, so I'm constantly revising it every day to make it better little by little. However, seeing the "Published" notification on Telegram every morning and actually seeing the content uploaded to the blog does give me a sense of accomplishment haha. Now that the blog is somewhat established... I plan to step up my efforts to mass-produce Shorts and Reels!

크로키의 잡도리

Blog Automation (1) I’d like to share my experiences with those of you looking to automate your blogs! Since this is quite a long post... please read it when you have the time! -——————————————————- I personally built a system where an AI writes blog posts four times a day. I’ve been so busy lately working on this and living my real life that I haven't had time to think about it. While there are still areas that need improvement, most things are sorted out. However, I think I will need to continue refining the tone, image, and persona. It is a one-person media engine that automatically runs four times a day, handling everything from collecting trends to publishing blog posts and generating shorts. I started with Claude Code, but since I wasn't on the Max x20 plan and tokens were being consumed quickly, I switched to Codex CLI. I wanted to write AI news and Vibe Coding educational content every day, but writing four long posts manually each day was realistically too much of a hassle and time-consuming. So, I started this project with the thought, "What if I automated everything from collection to publishing?" I thought it would be simple at first, but publishing methods differ by platform, AI-generated text is immediately obvious, and there are copyright issues with images. I think it took me over a week of daily revisions to reach its current form. » This is how the pipeline works: Windows Task Scheduler runs the Python script four times a day. (Vibe Coding Education Series Part 1, AI News Part 3) One cycle of the pipeline consists of 6 to 7 stages: 1) It collects trends from Naver DataLab, Google Trends, and 14 RSS sources. 2) It selects only the single most impactful news item, and the Codex CLI writes an in-depth article. 3) It generates images using Gemini 3.1 Flash. 4) It converts the content to blog HTML. 5) It publishes simultaneously to Blogger and Naver Blog. Upon publication, it even sends an indexing request via the Google Indexing API. 6) JSON for Shorts is also automatically generated. + Additionally, once the JSON for Shorts is generated, an "Edit" button is added to the Telegram message. Clicking this button opens a local video editor where you can make edits and render the video. After rendering, clicking the "Post to Instagram + YouTube" button automatically posts the content to the connected account API. + Furthermore, the entire process is configured to send real-time notifications via a Telegram bot, making it quite convenient as you can immediately identify where a failure occurred. » Reasons for Choosing Local Instead of Claude Cowork Actually, I initially considered Claude Cowork. It’s convenient when things run automatically in the cloud, isn't it? However, realistically, there were too many limitations. Tistory and Naver Blog are prime examples. Since the official APIs for both platforms were discontinued a long time ago, you have to launch the browser directly using Playwright. Also, Tistory frequently resets its cookie and login authentication, so I just gave up on it. Also, since Cowork is cloud-based, it works well for platforms like Blogger, but local browser automation is not possible. Therefore, tasks such as registering for the Task Scheduler, accessing the SQLite DB, and processing image files all require local dependencies, so I proceeded with them locally. In conclusion, running Claude Code CLI directly locally in the beginning offered much more freedom than Claude Cowork. I only switched the article writing part to Codex CLI later, due to the token limit on my Claude Max subscription. If you write four articles a day, you reach the limit very quickly ;( » The part where I struggled the most with Naver Blog automation.. This was the part where I struggled the most. Standard web automation methods didn't work at all with Naver SmartEditor ONE.. Since the editor wouldn't recognize text inputs via DOM manipulation or execCommand, a bizarre bug kept occurring where copying to the clipboard with pyperclip and pasting with Ctrl+V resulted in a strikethrough appearing across the entire text. I even tried typing it out character by character using `page.keyboard.type()`, but it was the same. So, I added a script to click the strikethrough button once the post is finished on Naver Blog, selecting the entire body content (Ctrl+A), and pressing it once; that solved the problem. That's it.... The images weren't simple either; it turned out that it is stable to handle this by clicking the image button within the editor, capturing the file chooser event to upload a local file, and clicking the publish button with a bounding box mouse click rather than a CSS selector.

크로키의 잡도리

I’d like to share my experience with those of you looking to automate your blogs! I personally built a system where an AI writes four blog posts a day. I’ve been so busy lately juggling this with my real life that I haven't had time to think about it. While there are still areas that need improvement, most things are sorted out now. However, I think I will need to continue refining the tone, image, and persona. It is a one-person media engine that automatically runs four times a day, handling everything from collecting trends to publishing blog posts and generating shorts. I started with Claude Code, but since I wasn't on the Max x20 plan and tokens were running out quickly, I switched to Codex CLI. I wanted to write AI news and Vibe Coding educational content every day, but writing four long posts manually was realistically too much of a hassle and time-consuming. So, I started this project with the thought, "What if I automated everything from collection to publishing?" I thought it would be simple at first, but publishing methods differ by platform, AI-generated text is immediately obvious, and there are copyright issues with images. I think it took me over a week of daily revisions to reach its current form. » This is how the pipeline works: Windows Task Scheduler runs the Python script four times a day. (Vibe Coding Education Series Part 1, AI News Part 3) One cycle of the pipeline consists of 6 to 7 stages: 1) It collects trends from Naver DataLab, Google Trends, and 14 RSS sources. 2) It selects only the single most impactful news item, and the Codex CLI writes an in-depth article. 3) It generates images using Gemini 3.1 Flash. 4) It converts the content to blog HTML. 5) It publishes simultaneously to Blogger and Naver Blog. Upon publication, it even sends an indexing request via the Google Indexing API. 6) JSON for Shorts is also automatically generated. + Additionally, once the JSON for Shorts is generated, an "Edit" button is added to the Telegram message. Clicking this button opens a local video editor where you can make edits and render the video. After rendering, clicking the "Post to Instagram + YouTube" button automatically posts the content to the connected account API. + Furthermore, the entire process is configured to send real-time notifications via a Telegram bot, making it quite convenient as you can immediately identify where a failure occurred. » Reasons for Choosing Local Instead of Claude Cowork Actually, I initially considered Claude Cowork. It’s convenient when things run automatically in the cloud, isn't it? However, realistically, there were too many limitations. Tistory and Naver Blog are prime examples. Since the official APIs for both platforms were discontinued a long time ago, you have to launch the browser directly using Playwright. Also, Tistory frequently resets its cookie and login authentication, so I just gave up on it. Also, since Cowork is cloud-based, it works well for platforms like Blogger, but local browser automation is not possible. Therefore, tasks such as registering for the Task Scheduler, accessing the SQLite DB, and processing image files all require local dependencies, so I proceeded with them locally. In conclusion, running Claude Code CLI directly locally in the beginning offered much more freedom than Claude Cowork. I only switched the article writing part to Codex CLI later, due to the token limit on my Claude Max subscription. If you write four articles a day, you reach the limit very quickly ;( » The part where I struggled the most with Naver Blog automation.. This was the part where I struggled the most. Standard web automation methods didn't work at all with Naver SmartEditor ONE.. Since the editor wouldn't recognize text inputs via DOM manipulation or execCommand, a bizarre bug kept occurring where copying to the clipboard with pyperclip and pasting with Ctrl+V resulted in a strikethrough appearing across the entire text. I even tried typing it out character by character using `page.keyboard.type()`, but it was the same. So, I added a script to click the strikethrough button once the post is finished on Naver Blog, selecting the entire body content (Ctrl+A), and pressing it once; that solved the problem. That's it.... The images weren't simple either; it turned out that the process requires clicking the image button within the editor, capturing the file chooser event to upload a local file, and that the publish button needs to be clicked via a bounding box mouse click rather than a CSS selector for stability. »

크로키의 잡도리

[ Aster + USD1, WLFI Week 1 Rewards Distributed! ]☀️ Week 1 rewards for the Aster + USD1 futures market have been distributed, with a total of 625,000 WLFI tokens distributed! ✔️ A total of 625,000 WLFI tokens were distributed as rewards for USD1 perpetual futures trading, and the trading volume for Week 1 exceeded $2.66 billion in just one week. ✔️ The rewards received this time apply only to futures trading rewards; it is said that the higher the trading volume and holding amount, the higher the reward weighting. Additionally, a separate monthly incentive of 2.5 million WLFI tokens is scheduled to be distributed to USD1 holders! ✔️ You can check your transaction history in Aster Spot > Trading History, and the Week 2 reward phase has already begun. The WLFI event is always feeding us generously, so I'm going to keep digging until it feeds me..! Original Tweet

크로키의 잡도리

☄️ [ PICK ME : TG Giveaway Platform Launch! ] ☄️ To all you channel operators out there, what do you think is the most annoying part when running a channel event?! Verifying participants' channel membership, filtering out fraudulent entries and bots, the raffle, announcing results, and so on! We’ve created a platform that automatically handles all of these processes! And for the participants, it’s not easy to keep up with multiple events running at once! In that case... 'Pick Me,' filled with the hope of winning, is the perfect fit! Of course, events organized by individual projects and joint events created by multiple channel owners are also possible! ━━━━━━━━━━━━━━━ ✔️ Create Event (Done in 3 Minutes) - Create an event by simply entering options such as title, prize, and duration. - Differentiate between Personal Channel and Project categories. - Supports thumbnail and banner image uploads. - Set conditions for simultaneous membership in multiple channels. - Set conditions for simultaneous membership in multiple communities. ✔️ Automated Verification System - Telegram bot double-checks channel/community membership status (when clicking the join button and when selecting winners). - Participants who are not members are blocked from participating. - Twitter account linking requirement. - Discord server membership requirement available. - EVM/SOL wallet address collection requirement. ✔️ Prevention of Fraud - Telegram Premium user filter. - Profile picture filter. - hCaptcha bot blocking. - IP-based multi-account detection. ✔️ Draw & Results Announcement - Select Pure Random / Weighted Random. - Set up reserve winners. - One-click draw execution. - Select option for automatic channel announcement. - Participants can check results directly. ✔️ Admin/Channel Leader Tools - Download Participant List CSV - Event Participant Status Dashboard - Check Real-Time Participant Count ━━━━━━━━━━━━━━━ Currently in beta, it is available for free. There is no limit on the number of events Channel Leaders can post. Since we just launched, there are no events posted yet, but we look forward to the not-too-distant future... We designed this so that Channel Leaders and visitors can use it easily and simply! We ask for your interest! (If you would like to try it out, please DM me and I will guide you! / I will contact you when I wake up in EST time!) 🔗 Web : pickm3.com 🔗 Inquiries & Partnerships : @min0119

크로키의 잡도리

☄️ [ PICK ME Telegram-Exclusive Giveaway Platform ] ☄️ To all you channel operators out there, what do you find most annoying when running a channel event?! Verifying participant channel membership, filtering out fraudulent entries, the raffle, announcing results, and so on! I’ve created a platform that automatically handles all of these processes. ━━━━━━━━━━━━━━━ ✔️ Create Event (Done in 3 Minutes) - Create an event simply by entering options such as title, prize, and duration. - Differentiate between Personal Channel and Project categories. - Supports thumbnail and banner image uploads. - Allows setting conditions for simultaneous membership in multiple channels. ✔️ Automated Verification System - Telegram bot double-checks channel/community membership (when clicking the join button and when selecting winners). - Participants who are not members are blocked from participating. - Twitter account linking requirement. - Discord server membership requirement. - EVM/SOL wallet address collection requirement. 🔹 Prevention of Fraud - Telegram Premium user filter. - Profile picture filter. - hCaptcha bot blocking. - IP-based multi-account detection. - Monitoring of winning concentration. 🔹 Draw & Results Announcement - Select Pure Random / Weighted Random. - Allows setting reserve winners. - One-click draw execution. - Automatic channel announcement (includes link to results page). - Participants can check results directly. Available 🔹 Admin Tools - Download Participant List CSV - Event Participant Status Dashboard - Check Real-time Participant Count ━━━━━━━━━━━━━━━ Available for free. No limit on the number of events. 👉 pickm3.com 📩 Inquiries & Partnerships: @min0119

크로키의 잡도리

I tried building an 'automatic prediction market betting bot' using Mirofish, which is trending these days. To be honest, I have absolutely no luck when it comes to trading, and I hesitated because I only have bad memories of getting burned by futures in the past. However, I went ahead and built it purely with the intention of creating a proper app! (Of course, this is also my first time using a prediction market, aside from mock trading... haha) To look at what Mirofish is, it is a 'swarm intelligence engine' that assigns different personalities, memories, and behavioral patterns to thousands of AI agents and allows them to freely discuss in a simulated environment (Twitter/Reddit). The AIs post, comment, and retweet with each other, engaging in fierce debates, and as the rounds pass, a single 'public opinion' strangely forms and changes. Based on this, Prediction markets like Polymarket are essentially 'crowd behavior markets' where people bet on whether "this event will happen," and the market price reflects the consensus of the crowd. This led me to wonder, "What if we bet on the discrepancy when the consensus of a virtual crowd simulated by AI differs from the actual market price?" Upon searching, I found that quite a few people had already created bots applying this principle! So, I quickly put together a bot that connects these two elements. 1) Collect real-time news 2) Inject data into the MiroFish engine 3) Simulate discussions and public opinion formation by AI agents 4) Extract 'probabilities' from simulation results 5) Automatic betting after comparing with the actual market price I believe the key point is betting when the AI crowd in the simulation says, "The real-world crowd is wrong." Although it is still in the early stages, I think it will be a quite interesting experiment, both technically and from an investment perspective. For now, playing it safe with just $500, haha.

크로키의 잡도리

[ Myriad Prediction Market Switches to USD1 Single Currency ☀️ ] ✔️ Key Details » Myriad, a prediction market platform based on BNB Chain » Starting Season 3, all transactions are integrated into the single currency, USD1 » Settlement Currency: Multi-chain » USD1 Single Currency » Transaction Structure: AMM > CLOB method introduced ✔️ AMM (Automated Market Maker) » A method primarily used in DEXs like Uniswap » No counterparty is required; transactions proceed through liquidity pools ✔️ CLOB (Central Limit Order Book) » A method used in stock markets and centralized exchanges like Upbit and Binance » Order book-based trading, allowing users to trade at their desired price It feels like the USD1 ecosystem is expanding faster than expected! Original Tweet

Loading..