That’s Old News
Ask AI to find you what’s hot, and the biggest trap is old news passed off as news.
A model’s knowledge has a cutoff date. It’ll talk your ear off about the “latest trends,” spit flying, when out in the real world that stuff went cold ages ago. Like a sent-down youth just back in the city after years in the countryside: the minute he opens his mouth, he’s quoting prices from years back.
There’s a recent open source project that treats exactly this condition. It’s called Last30Days, and the name is the house rule: only the last 30 days count.
So what is it
The project is called last30days-skill, by Matt Van Horn, who has quite the résumé: he co-founded June, the smart oven company, which later sold to Weber, and he was also one of the co-founders of the company that later became Lyft.
It opened for business in January 2026 and racked up 65,000 stars in nine months (65,225 when I looked), plus 5,600-some forks. MIT license, written in Python.
It’s an agent skill: install it in your AI assistant, type /last30days plus a topic, and it combs through the last 30 days of discussion across a bunch of platforms, then hands you a research brief with sources.
What it cures
The author is refreshingly upfront about it: he built it to keep up with the AI scene. This field changes every day, and the nerds on Reddit and X always hear first; a model’s training data always runs months behind the community.
Search engines won’t save you either. What Google turns up is mostly articles written by editors and pages pumped up with SEO. Reddit comment threads and posts on X are out of its reach. And every platform is its own little fiefdom, with its own API and its own login.
The idea here: you bring your own keys and your browser’s logged-in sessions, and let the agent string all those fiefdoms together in one go. As the author puts it, Google aggregates editors, and this searches people.
Where it looks
Four work right out of the box, nothing to configure: Reddit (comments included), Hacker News, Polymarket and GitHub.
The first run pops up a setup wizard, and in half a minute you can switch on a whole string more: X, YouTube, TikTok, Instagram, Bluesky, arXiv, Techmeme, plus web search. The README lists more than twenty sources, and even RedNote (Xiaohongshu) is in there (you have to run a logged-in RedNote service locally yourself, then switch it on by hand).
The ranking doesn’t care whose SEO is best. It cares whether real people are buying it: Reddit upvotes, likes on X, YouTube views, TikTok engagement, and bets placed with real money on Polymarket. A Reddit post with 1,500 upvotes outweighs a blog post nobody reads, and odds backed by actual cash beat an expert pulling a number out of thin air.
When the same story gets posted on Reddit, argued over on X and filmed on TikTok, it merges them into one item instead of reporting it to you as three pieces of news.
How to use it
In Claude Code, two commands install it:
/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30days
For Codex, Cursor, Copilot, Gemini CLI and the rest, 50-plus agents in all, use this one:
npx skills add mvanhorn/last30days-skill -g
Once it’s installed, type /last30days in the chat, followed by the topic you want to look into:
/last30days what is everyone saying about AI video tools lately
First it works out where to look for that topic: relevant subreddits, accounts, channels, GitHub repos. Then it searches all the platforms at once, scores what it finds on engagement, relevance and freshness, and finally pulls it all into one brief, with a source on every item.
Once you have the brief, there are three things to dig out of it:
- Hot tools: what everyone’s been using lately.
- User gripes: what still doesn’t work well, and what people are cussing out the loudest.
- Topics worth mining: what to write next, what your next video should cover.
If you haven’t even picked a topic yet, just ask it “what’s hottest in the AI agent world lately?” It switches to discovery mode and ranks 5 to 10 topics that are picking up steam, each with numbers from every platform and a ready-made /last30days command attached, so you can keep digging.
After one run, that conversation is halfway to being an expert. Keep asking it questions, or have it write prompts or drafts, and it all runs off what actually happened in the last 30 days.
One user in the README says that every time he wrote something, he used to spend an hour and a half just going through Reddit and X, one tab at a time. Hand that job to this thing, and you can go get on with your life.
Does it cost anything
- Reddit, Hacker News, Polymarket, GitHub: free, no key needed.
- YouTube: just install the free
yt-dlp. - X: log in to x.com in your browser and it borrows your browser’s session, for free; or you can put in an official API key.
- TikTok, Instagram, Threads, Pinterest, LinkedIn: one ScrapeCreators key covers the lot; the first 10,000 calls are free, then you pay as you go.
- Web search: a Brave key, 2,000 free searches a month.
- Perplexity deep research: a paid option, billed by usage on your own key.
Bottom line: it works without spending a cent, but if you want the full picture, bring your own keys.
No need to sweat the data either: by default everything runs on your own computer, and nothing leaves it except the calls to whatever search APIs you use. Briefs are saved to ~/Documents/Last30Days/ by default.
One last thing
However clever AI gets, it has one bad habit: the world in its head stopped on the day its training cut off. If you don’t show it anything new, it’ll serve you old news as news, with a perfectly straight face.
What this skill does, put plainly, is make the AI scroll its phone for thirty days before it opens its mouth. See what everyone’s talking about first, then decide what to say.
My only worry is that once it learns to scroll, it’ll turn out just like the rest of us: one quick look, and the sun’s coming up.