Last30Days AI Research Skill

Why it's ranked here

What lifts this above a search wrapper is the plumbing around ranking. Source outcomes are typed into ten distinct states, so a rate limit or an auth failure never reads as a source having nothing to say. The machine-readable research contract is versioned and locked by a checked-in golden snapshot. A discovery run where nothing clears the evidence bar prints an honest empty result and names the closest weak signal instead of ranking noise. Coverage is gated at 84 percent. The cost is real: most of the source table needs credentials, cookies, or a helper binary on the path before it does anything at all.

What's good

Retrieval quality has explicit defenses. A grounding check requires the head word of the topic to appear before final ranking, so a viral off-topic post cannot ride engagement into the brief, and the check is deliberately conservative so its own misses cost nothing. The limited budget for fetching comments goes to posts that already passed grounding. The evidence bar for discovery is absolute rather than relative to the current pool, which is the usual way such bars rot. There is also a local evaluation harness measuring overlap, retention, precision and ranked gain against a baseline revision, with a separate model acting as judge.

Tradeoffs

Free out of the box means Reddit, Hacker News, Polymarket and GitHub. X needs either browser session cookies or a paid provider key, TikTok and Instagram need a third-party scraping key, and several sources activate only when their helper binary resolves on the agent subprocess path, which the project documents as a recurring setup failure. One documented Reddit search path routes through a paid model provider's web search tool. HTML publishing uploads pages to a hosted service and they are public by default. The quality evaluation runs five fixed topics by default, and its own notes call the ranking metrics comparative rather than truth.

How to use it well

Best fit is anyone who needs the community's current read before a call, a purchase, or a build: sales prep, competitor watch, meeting research, prompt technique hunting. It belongs at the front of a workflow, feeding a versioned machine-readable export to whatever writes the actual document. Budget half an hour for credentials and browser auth, or accept the thin free tier. It is not monitoring and it is not an archive. The window is recent by design, so it will not tell you what a company shipped two years ago, and it makes no claim on paywalled or private sources.

Observed

License

MIT

Language

Python engine, plus a Go MCP server wrapping it

Runtime requirement

Python 3.12 or newer; the project manifest declares no third-party runtime dependencies

Install surface

harness plugin marketplace, the Agent Skills installer, or a drag-in bundle for Claude Desktop

Interfaces

slash command inside agent harnesses, direct command-line invocation of the engine, and an MCP server exposing one research tool

Machine contract

versioned JSON export for research plus a separately versioned export for discovery, locked by a golden snapshot test

Failure reporting

per-source outcome states that separate a clean empty result from rate limits, auth failure, timeout, unreachable endpoints and schema drift

Tests

pytest suite in a top-level tests directory with a coverage floor configured in the project manifest

Documentation

README maintained in several languages beyond English, with separate concept, configuration and contributor guides