last30days — recency-grounded multi-source research

/last30days is an installable Agent Skill (by mvanhorn, MIT) that researches what people
actually say about any topic across Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub,
Bluesky and the web — scored by upvotes/likes/odds, then synthesized into one cited brief.

It is NOT bundled with 8sync — it ships its own Python engine (scripts/last30days.py) and is
installed through the host's skill mechanism. This doc tells you when to reach for it and how it is
invoked; the authoritative runtime contract lives in the upstream SKILL.md inside the install dir.

1. When to use it

Use this over a raw WebSearch loop whenever the answer depends on recent, engagement-weighted,
multi-platform community signal (Reddit comments, X threads, YouTube transcripts, Polymarket odds)
rather than editorial SEO results.

2. Install (engine is separate — do this once per machine)

# Any of 50+ Agent Skills hosts (codex, cursor, gemini-cli, opencode, claude-code, ...)
npx skills add mvanhorn/last30days-skill -g          # -g = global, available in every project

# Claude Code native plugin (auto-updates):
#   /plugin marketplace add mvanhorn/last30days-skill
#   /plugin install last30days

Reddit (+comments), Hacker News, Polymarket, and GitHub work with zero config. The first
/last30days run launches a 30-second setup wizard that unlocks the keyed sources below.

3. Sources & optional keys

4. Invocation rules (when YOU are the host model running it)

• Run the engine — do not answer from WebSearch alone. A prose summary built only from a few
  WebSearch calls is NOT valid output; the Python engine is the skill.
• No topic given → ask once for a topic. Do not research a blank query.
• Named-entity topics REQUIRE a query plan (--plan <file>): you are the planner — generate the
  JSON plan upstream and pass it; never inline apostrophes in --plan.
• Pass the engine's emoji-tree footer (✅ All agents reported back!) through verbatim — a
  response without it did not actually run the skill.
• Output is a single cited brief, not a Sources: link dump. Citations are inline [name](url).

Full flag reference, output LAWs, and the per-source key matrix: see references/cheatsheet.md.

5. Citation convention

Echo the brief's inline [name](url) citations verbatim. Research files are saved under
LAST30DAYS_MEMORY_DIR (default ~/Documents/Last30Days/); cite the saved *-raw.md path when the
user wants the durable artifact.

Homepage: https://github.com/mvanhorn/last30days-skill