Skip to content

/last30days: AI Agent Skill for Multi-Platform Research

Research on a single topic rarely lives in one place. A useful answer may require Reddit threads, X posts, YouTube transcripts, Hacker News debate, GitHub activity, and Polymarket bets — each with different access rules and ranking signals. /last30days is an open-source skill by mvanhorn that lets an AI agent search many of those places in parallel, rank what people actually endorsed, cluster related items together, and return one cited brief.

Why cross-platform research is hard for AI agents

A conventional search engine mainly rewards editorial authority and SEO. That can bury a Reddit discussion with thousands of upvotes beneath a single-author article, and it gives no weight to whether people liked, shared, or put money behind a claim. Meanwhile, large language models often cannot reach the platforms where current conversation happens: some pages are not readable natively, some APIs are paywalled, and a self-hosted model may have no search at all. Checking each platform by hand can take an hour or two and leaves dozens of tabs to reconcile.

/last30days addresses the gap as an agent skill rather than a standalone website. Installed into a host such as Claude Code, Codex, Gemini CLI, Cursor, Copilot, or another Agent Skills host — the project describes support for more than 50 such hosts — it gives the agent a shared way to query many sources, then synthesize the findings. It is open source and normally runs locally, although a web-based host changes that deployment picture.

From query to one cited brief

Start with a topic: a person, product, company, or comparison. The agent first resolves the relevant people and places — for example, an X handle, a GitHub repository, or a YouTube channel. It then searches several sources in parallel, often with more than one query per source. Results are scored for engagement, freshness, and relevance, and the search can go beyond headlines into comments, captions, and transcripts.

Related items are merged into clusters. If a festival is announced on Reddit, debated on X, and discussed for ticket prices on TikTok, those pieces appear as one story with combined context instead of three disconnected links. The final output is a single brief with citations for cross-checking. Once that context is loaded, follow-up chat with the agent can be more accurate because it is grounded in the gathered material.

A concrete illustration from the captured example: asking about Lionel Messi immediately after installation, using only the zero-config sources, produced a summary drawn from Reddit, Hacker News, and GitHub. That is a useful first test because it shows both the strength and the limit of the out-of-the-box setup — coverage expands only after the remaining sources are configured.

Setup: what works immediately and what needs configuration

Reddit, Hacker News, GitHub, and Polymarket are available straight after installation. On first run, a setup wizard helps configure additional sources such as X, YouTube, TikTok, Techmeme, and arXiv. Some are free; others may work better with an API key.

The wider source list includes X, YouTube, TikTok, Instagram Reels, Digg, LinkedIn, StockTwits, Threads, Pinterest, Bluesky, Perplexity, and the web. Each adds a different signal: YouTube transcripts surface deep reviews, arXiv provides recent research without paywalls, GitHub shows stars and discussion around code, and Polymarket reflects where people are staking money rather than merely expressing an opinion.

Key features and a practical decision checklist

People-scored ranking: likes, upvotes, and staked money carry weight, so endorsement by many users can outrank a lone editorial voice.

Comments coverage: Instagram, Reddit, and YouTube comments are included by default, material that typical search results often miss.

Cross-source clustering: posts, discussions, and news about the same topic are grouped into a single unit for easier review.

Shareable output: the skill can produce a self-contained HTML brief that opens in any browser.

Trend monitoring: a bundled watchlist script supports scheduled runs, and a briefing script produces periodic digests.

Decision checklist — choose the local skill when you want direct control over each connector and are comfortable maintaining separate API keys, browser authentication, rate limits, differing data formats, and proxy or anti-bot handling. Consider a managed scraping service when you would rather have one subscription, prebuilt scrapers for search engines and social platforms, automatic proxy rotation and headless browsing, and scaling handled for you. The trade-off is control and local operation versus ongoing setup and maintenance.

Use cases worth trying first

Pre-meeting preparation: gather recent posts, video essays, and research papers on a field into one summary.

Hiring signals: examine a company's active job titles to infer where its focus is shifting.

Emerging topics: an open-ended query about what is exploding in a field can surface fresh discussion from Hacker News, Digg, Reddit, and X before it becomes mainstream.

Keeping up with a topic: a query such as /last30days Kardashians pulls together news, Reddit, TikTok, X, and Polymarket so you can follow a fast-moving story across platforms without opening each one separately.

Tool comparison: compare projects using repository features, star counts, and issues alongside community discussion.

Current affairs: build a calm, non-sensational brief on a conflict and its consequences, such as energy prices, without manually sorting news results.

Pre-travel checks: confirm whether a venue is operating, note planned expansions and typical wait times, and capture local opinion.

Fast learning: generate a primer on a fast-moving tool grounded in the latest updates rather than an outdated tutorial.

Bottom line

/last30days is best understood as a research layer for an agent: it finds and ranks current, people-backed discussion across many platforms, then turns it into a brief you can interrogate further. Its local, open-source design gives you control, but each connector remains your responsibility to configure and maintain.

If that maintenance burden is the blocker, a managed service can supply the data-access layer — prebuilt scrapers, proxy rotation, and headless browsing under one account — while the agent still handles synthesis. Start with the zero-config sources, run one familiar topic, check the citations, and only then decide whether the additional connectors or a managed alternative are worth it.

Sources and the signal each contributes
SourceWhat it adds
RedditUnfiltered discussion backed by upvote counts, not always ranked by search engines
X (formerly Twitter)Hot takes and the latest posts by experts, celebrities, leaders — direct from the source
YouTubeDeep-dives and detailed reviews, with searchable transcripts to get just the highlights
TikTokHighly engaging content with millions of views/likes that may not be picked up by search engines
Instagram ReelsThe influencer perspective, backed by likes for the visual signal
Hacker NewsArguments and likes from where developers hang out
PolymarketSignals of what people are putting their money on, not just opinions
GitHubWhere code gets written; search engines don't surface projects by stars or discussion
DiggCurated story clusters from the AI world
arXivAccess to the latest research from the scientific world without steep paywalls
TechmemeThe latest editorialized tech news when editorial quality is needed
LinkedInPosts and articles giving signals from the professional world
StockTwitsInsights from the social layer of trading not found from market data alone
ThreadsLike X/Twitter but an entirely different crowd
PinterestA visual discovery layer that is already curated
BlueskyAnother X/Twitter analogue with a crowd that loves decentralized social media
PerplexityAccess to raw search data and research
WebWeb pages including blogs and product pages, used as necessary

Coverage and configuration requirements vary by source; some need API keys or authentication.