When I want to know whether a new tool is worth my time, Google is close to useless. What comes back is a layer of review sites copying each other, an affiliate listicle, and the vendor’s own page. What I actually want lives somewhere else: in a Reddit thread where somebody gives up after three weeks, in a comment under a YouTube video, in a post liked 400 times by people who genuinely run the thing.
Assembling that by hand costs me half an evening. There is now an agent skill that does it in one command, and this week it sits at 58,077 stars on GitHub. I went through it properly, because a tool growing that fast deserves a harder look than a round of applause.
What actually happened
On 12 August, Allie K. Miller appeared on Greg Isenberg’s podcast for a 48-minute episode about running a fleet of AI agents. She named /last30days in it. The next morning Matt Van Horn, who built it, posted about the mention on X. Miller replied to that post herself and immediately qualified her own compliment: most people, she reckons, only use it for research, and there is more in there than that.
That is the news, and it is small. The interesting part is what the compliment was about.
The mechanism: engagement as ranking
The project describes itself as a skill that “researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded summary”. The heart of it is what that sentence leaves out: no crawler, no index, no pagerank.
For each source, the skill pulls posts from the past 30 days and treats the platform’s own measurement as the ranking signal. Upvotes on Reddit, likes on X, views on YouTube, points on Hacker News, and on Polymarket the money people have actually staked on an outcome. It then clusters items covering the same event, and tags any cluster drawn from a single platform as single-source, with an instruction to treat it cautiously. Clusters carrying Reddit and X and YouTube together get the highest confidence.
That is an explicit position on what counts as true. Google mostly ranks what other sites link to. This ranks what people click, upvote and pay for. Both are proxies, and neither one measures whether something is correct.
Where the sources come from varies a great deal, and that is the side rarely mentioned in the recommendations:
| Source | How it arrives | What you need |
|---|---|---|
| Hacker News | Algolia search API | nothing, free and keyless |
| Polymarket | Gamma API | nothing, free and keyless |
| YouTube | yt-dlp running locally on your machine | brew install yt-dlp |
| GitHub | the gh CLI you already have | an authenticated gh |
| free path, with a paid scraper as fallback | optional ScrapeCreators key | |
| X | your own API key, or your own browser cookies | xAI key, xurl, or AUTH_TOKEN and CT0 |
| TikTok, Instagram | commercial scraper | ScrapeCreators key, 10,000 free calls |
The bottom two rows are the part I spent longer thinking about than all the rest.
The numbers, including the awkward ones
Measured on 13 August 2026, pulled from the GitHub API rather than copied off a screenshot.
- 58,077 stars, 5,045 forks
- Version 3.18.4, MIT licensed
- Installable on Claude Code, the Grok CLI, Claude Desktop, claude.ai, and through
npx skillson more than fifty other agent hosts
Stars are not users. It is the cheapest signal of approval GitHub has: one click, no install, no second visit. A repository that spends three weeks on the trending list collects stars from people who never run it. The 5,045 forks say a little more, though plenty of those are “look at this later” too. What I cannot find anywhere, and what would be the only honest measure, is how many installs run an actual query in a given week. The project does not publish that, and it is far from alone in this.
There is a second claim in Van Horn’s post that I am deliberately not repeating as fact. He writes that he handed Grok the 48-minute YouTube link and asked it to cut out the section mentioning last30days, and that it got it nearly right first time. I could not find any xAI announcement describing clipping from an existing YouTube video as a feature. What is documented covers generating and extending short clips. One user’s experience is not a product capability, and I am treating it here as unconfirmed.
The caveat
The default shifts from “what got published” to “what got popular”. For a question about a niche tool that is a clear gain, because the indexable layer there is mostly marketing. For a sensitive subject it inverts. The loudest voices on Reddit are not the best informed, and an X post with 3,000 likes can be a widely shared misunderstanding. The skill does label thin evidence honestly, but those labels land inside a summary that then gets read as an answer. The warning sticker rarely survives the copy and paste.
Responsibility for access lands on you, not on the tool. For X, the skill offers a path where you supply your own session cookies. That is your logged-in account reading at scale, under your name, inside terms you personally accepted. The project asks for explicit consent and documents it carefully, so this is not a complaint about the maintainer. It is still a different thing from an API key you can revoke. TikTok and Instagram run through a commercial scraper, which means your query travels to a third party you did not choose.
Some judgement quietly drains away. When I spend half an evening in threads, I also see who is answering, how angry they are, and whether the complaint is about version 2 while version 4 has shipped. A summary compressing that into “mixed reception, mostly around performance” is faster and thinner. The dangerous moment, I find, is not when the tool is wrong. It is when it is exactly good enough to make me stop reading.
What I actually do with it
I installed it where I already run skills in Claude Code, with only the free sources switched on: Hacker News, Polymarket, YouTube through local yt-dlp, and GitHub through the gh CLI that was already there. No ScrapeCreators key, and certainly no X browser cookies. That costs me TikTok, Instagram and part of X, and for my questions about developer tooling those three matter least.
I use it in one direction on purpose: to find what I would otherwise miss, not to be handed a verdict. I treat the output as a list of places to go and look. I open the sources it surfaces myself, and read the summary only afterwards. That sounds laborious and it is, but it is precisely the part I do not want to automate.
If you are still weighing up whether this class of tool is for you, I wrote earlier about the difference between a chatbot that answers and an agent that goes off and does it, and /last30days is a textbook case of the second kind.
Frequently asked questions
Do I need paid API keys to get started?
No. Hacker News, Polymarket, GitHub and YouTube all work without a key. YouTube does need a local yt-dlp install, which runs on your own machine. Keys only come in for TikTok, Instagram and full X coverage.
Is this a replacement for Google?
No, and it is not trying to be. It answers a different question: what are people saying about this over the past thirty days. For facts that are not tied to a date, or for documentation, an ordinary search engine is better and faster.
Why a thirty-day window?
Because the measurement it leans on decays. Upvotes, likes and views pile up around the moment of publication, and a two-year-old post with heavy engagement says little about the situation now. The window is not a technical limit, it is the whole idea.
Can I only use this inside Claude Code?
No. There are install paths for Claude Code, the Grok CLI, Claude Desktop, claude.ai, and through the Agent Skills CLI for more than fifty other hosts, including Codex, Cursor and Gemini CLI.
Sources
- Matt Van Horn on X, post about the podcast mention and the Grok anecdote — 13 August 2026: x.com/mvanhorn
mvanhorn/last30days-skillon GitHub, project description, install matrix and the Security & Permissions section — accessed 13 August 2026: github.com/mvanhorn/last30days-skill- The project’s
SKILL.md, version 3.18.4, for the source list, the clustering and the uncertainty labels — accessed 13 August 2026: SKILL.md on main - Star and fork counts via the public GitHub API — measured 13 August 2026: api.github.com
- “My top secrets to running an AI Agent Workforce”, Greg Isenberg with Allie K. Miller, 48 minutes — published 12 August 2026: youtube.com
Fact-checked on 13 August 2026. The star and fork counts come from the GitHub API and are a snapshot; by the time this publishes they are probably higher than what you see here. The install matrix and source list came from the repository itself, not from write-ups about it. Two things I could not verify myself. I did not listen to the 48-minute podcast episode, so the mention and its exact wording come from Van Horn’s post and from Miller’s public reply beneath it. And the claim that Grok cut the right segment out of that video in one command appears only in that post; I found no xAI documentation for such a feature, so I am treating it as unconfirmed. Figures on real usage are missing entirely, both for this project and for almost every comparable one.