AI Daybook - Today's AI news, shared every morning at 6 AM (formerly Daily AI Brief)
Translated by AIView original
Screenshots
Screenshot 1 of 4
The Daily AI Brief is now AI Daybook, with a new look.
After the last intro, I read the news I had made myself and found a lot to improve.
- Yesterday's stories kept showing up on the front page, so every day looked about the same
- Most of the items were plain news, so the open source projects, models and papers developers actually care about almost never showed up
- Robot stories had no fitting section, so some got filed under 'Agents'
- I wanted to publish for overseas readers, but there was no English edition, and the name 'Daily AI Brief' was too common
So I changed the following.
- Front page is today's issue only
At 06:00 every day, only the stories I picked go on the front page. It is made up of a lead, subs, Today's Briefing, and the Open Source & Paper Radar, and the top has a 'three-line summary' with the core of this issue.
- Open Source & Paper Radar
It also collects trending models from HuggingFace, GitHub topic search, and arXiv and HF Daily Papers. Repos, models and papers get REPO, MODEL or PAPER badges, and stars, downloads and upvotes are shown too. The Sunday issue runs a 'Weekly Radar' that bundles the top 10 from the week's radar.
- Reorganized into 6 sections
LLM · Agents · Physical AI · Open Source & Tools · Research & Papers · Industry & Policy. I created a new Physical AI section covering robotics, humanoids and autonomous driving, and added more sources for it.
- English Edition
The same issue is also published in English (/en). Articles from Korean outlets are gathered in a 'Korea Desk' to bring Korean AI news to overseas readers. You can switch between them with the EN / 한국어 links at the top.
- Service renamed to AI Daybook
I changed it to a more memorable name for the global launch.
How it was made
I worked with Claude Code again this time. This time I dug into 'why is the front page boring' first.
- Checking the problem with data
While checking the October 25 front page, I traced the problem through code and the DB. There were two causes. First, the front page treated every story published within 48 hours by its original publication time as a candidate, so yesterday's stories got mixed in. Second, the selection prompt had a 'AI news editor-in-chief' persona with a rule to exclude 'minor repo and version updates', so open source was getting filtered out. Of the 388 items published in the last 14 days, 368 were RSS news.
- Changing the page structure
I changed the basis of an 'issue' from original publication time to collection time (KST), so the front page is written only from that day's collected items. The LLM classifies sections. The 'format' of items, such as news, open source, model or paper, is decided by source attributes, so it is used for badge display and front page diversity control without LLM cost. Existing articles were reclassified into the 6 sections with a migration script. The selection prompt was rewritten as an editor-in-chief who looks at news, open source and papers together.
- English edition
The English edition does not translate the Korean. In the same LLM call used for processing, it reads the original and writes the English title and summary together. If English contains Hangul or numbers not in the original, code filters them out and leaves the English blank. About 1,200 past articles were backfilled at once. Each page has hreflang, a sitemap and an English RSS feed so it also shows up in English search.
- Cost stays the same
The operating structure is unchanged. GitHub Actions collects once a day, commits the SQLite file DB, and deploys static pages to GitHub Pages. Sources and published counts grew and English was added, but thanks to the Message Batches API (50% off) and batched processing, I keep the cost per issue under a $0.1 cap. Things code can solve, like merging the same event, are handled by code without calling the LLM.
- Design and way of working
I kept the newspaper tone of DESIGN.md. The larger front page, radar and badges were checked first as mockup HTML and then implemented. Extra ideas like the three-line summary, Weekly Radar and linking follow-up coverage were worked on in parallel by sub-agents in their own worktrees, then merged, verified and deployed.
- Built with
Comments 0
No comments yet
Be the first to comment.