Research any topic from the last 30 days with last30days
last30days is an agent skill that searches Reddit, X, YouTube, Hacker News, Polymarket, GitHub and more in parallel, then has your agent write a cited brief of what people actually said in the last 30 days.
TL;DR
/last30days <topic> runs a Python engine that pulls posts, comments and engagement from many community sources, scores and merges them, and hands the evidence to your agent to synthesize. It installs as an Agent Skill on Claude Code, Codex, Cursor, Copilot, Gemini CLI and others, and the free tier needs no API keys.
Why this repo matters
The author's pitch in the README: he built it to keep up in AI, because "the training data was always months behind what the community had already figured out." A model's memory is old; Reddit, X and YouTube are not. The skill turns "tab by tab, thread by thread" research into one command, and the answer is grounded in dated, cited items instead of the model's recall.
Best practices
- Name a specific topic. The README's flow starts with "a person, company, product, technology, 'X vs Y'" and resolves who matters (X handles, subreddits, GitHub repos) before searching. A named entity gives that step something to resolve.
- Start on the free tier. Per the README's keys table, Reddit (with comments), Hacker News, Polymarket and GitHub need nothing; add keys only for the sources you miss.
- Match depth to the question. The engine exposes
--quick(lower latency) and--deep(higher recall) retrieval profiles. - Treat the output as cited evidence. SKILL.md's output laws require inline citations and an engine footer that reports which sources answered; check the
Partial Coverageblock (orsource_statusin JSON) before trusting a thin brief; the footer only has counts. - Ask follow-ups instead of re-running. SKILL.md says existing findings are reused for follow-ups unless you ask for a new topic or a freshness check.
- Know what leaves your machine. The README's data-handling section says reranking calls an external model only if you configured a reasoning-provider key; otherwise a local scoring model is used.
How-to
Where do I start?
1. How do I install it?
The README lists one command per host. For Claude Code (the recommended path, since the marketplace handles updates):
/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30days
For Codex, Cursor, Copilot, Gemini CLI and other Agent Skills hosts:
npx skills add mvanhorn/last30days-skill -g
-g installs to your user directory so every project sees it. OpenClaw users run clawhub install last30days-official. The skill needs python3 and node; SKILL.md requires Python 3.12+ for the engine.
Next: Run my first research
Try it
Pick a tool you are about to adopt. Run /last30days <tool> vs <alternative> and check the Partial Coverage block and footer counts first: which sources reported? Then ask one follow-up ("what do users complain about most?") and check that the answer cites a thread, not a guess.
Pitfalls
- Thin sources, thin brief. Without keys you get the free tier only. A
partialorskipped-unconfiguredsource status (reported inPartial Coverage/source_status, not the footer) means the absence of evidence is not evidence of absence. - Login-backed sources need consent. SKILL.md asks before reading browser cookies (for X, or the optional Reddit session) and treats a skip as a refusal. Do not expect X results without a token or login.
- Not offline. Source data comes from live public pages and APIs; only the optional store and library are local.
- Check the rerank privacy setting. If you set a provider key such as
OPENAI_API_KEY, your topic and retrieved snippets go to that provider for relevance scoring, per the README.
Next steps
- Repository and README
- CONFIGURATION.md: save directory, store, trend monitoring
- CONCEPTS.md: the project's vocabulary (engine, keyless path, entity grounding)
- JSON export reference
- SKILL.md