I started ai-tldr.dev because I was tired of opening Twitter at 9am and losing 40 minutes before doing any work. The deal I made with myself was simple: every week, write one short digest of what actually mattered in AI — models, repos, papers, tools — and stop scrolling.
Six months in, the site has 332 releases logged, a few hundred regulars, and a backlog of mistakes. This is the honest retro: what I'd cut if I started today, what I'd keep, and what surprised me.
What I'd cut
Categories that sound smart but nobody clicks.
I had an "ecosystem" bucket for org changes and funding news. Click-through is the lowest of any category — about a third of what models or tools get. Turns out, engineers don't open a digest to read about a Series C. They open it to see if there's a new model worth trying on Friday. I'd kill the bucket or fold it into a footnote.
Cover images on every entry.
Spent three weekends building an automated cover generator. Engagement difference between "image" and "no image" rows: noise. I now ship without them and the only person who noticed was me.
Twice-weekly cadence.
I tried bumping to Tuesday + Friday for a month. Open rates per send dropped roughly in half, total weekly opens stayed flat. People have one slot in their head for "the AI thing I read." Don't fight that.
SEO-bait titles.
Early on I A/B'd "Top 10 AI tools you need this week" against plain dated titles like "AI digest — week of March 3." The plain ones won on clicks AND on retention. The audience is engineers; they can smell a content farm.
What I'd keep
One specific idea per week, even when nothing is happening.
The temptation when it's a slow week is to pad. Don't. A 4-item digest is fine. A 12-item digest of half-relevant stuff trains people to skim and unsubscribe.
Filter views over feeds.
The single highest-retention feature is the category filter — people land on the models view, the repos view, the papers view, and stay. A flat feed is for the homepage; everything else should let you carve.
Manual taste, not just an RSS firehose.
I tried automating ingestion fully. The result read like a Bloomberg terminal for AI: technically complete, emotionally dead. I now keep a human pass before publish — even 5 minutes. It's the difference between a tool people skim and a tool people return to.
Plain markdown output.
I write each digest as markdown first, in a single file, before any formatting. It forces the writing to stand on its own. If the post is boring as plain text, no amount of cover art will save it.
The thing that surprised me
I assumed people wanted more depth — long writeups, benchmarks, comparisons. They didn't. The single most common feedback was the same line in different words: "I just want to know what's worth opening." A digest is a triage tool, not a magazine. Once I accepted that, every product decision got easier — shorter entries, fewer categories, faster ship.
If you're thinking about starting one
A few rules of thumb from the trenches:
- Pick a vertical narrow enough that you can hold the whole week in your head. "AI" is too broad; "AI engineering tools and model releases" is workable.
- Write the first 8 issues before you tell anyone it exists. You will hate issue 3 and want to quit; that's normal.
- Don't add a paywall, a discord, a course, a newsletter platform migration, or a logo redesign in the first 6 months. Just ship the digest.
- Track one number: do people open issue N+1 after reading issue N. Everything else is vanity.
The repos view is where I personally spend the most time — https://ai-tldr.dev/?cat=repo — because that's where the actual signal lives most weeks. Models are loud; repos tell you what people are building with them. Your mileage will vary.
Six months from now I'll probably disagree with half of this. That's fine. The point of running the experiment in public is that you find out which of your opinions were just vibes.
— Alex
Top comments (0)