AI Guides › Step-by-step guides
By Nigel Guy · 5 min read
You ask an AI what people are saying about a tool, a rival or a trend, and it answers fluently from whatever it absorbed during training. It reads as current. It is often a year old, and nothing on the page tells you so. The usual fix, "search the web", gives you a handful of editorial pages, not what practitioners are actually arguing about this month.
The rule: never ask an AI about a fast-moving topic without handing it dated, sourced evidence from a fixed window, and make it cite which source each claim came from.
The open-source /last30days skill does the gathering. This guide covers installing it, running it, and turning its output into decisions.
mvanhorn/last30days-skill repository on GitHub (MIT licence). Per its README it searches the last 30 days across Reddit, Hacker News, Polymarket, GitHub, X, YouTube, TikTok and others, ranks by engagement, and writes a brief. It is not an Anthropic product.In a Claude Code session, run:
/plugin marketplace add mvanhorn/last30days-skill
Claude Code prints "Successfully added marketplace" when it works. Other routes in the README, which I did not test: npx skills add mvanhorn/last30days-skill -g for other agent hosts, a .skill upload under Customize > Skills on claude.ai, and a .mcpb bundle for Claude Desktop (the README lists Windows Claude Desktop support as pending).
Run /plugin, open the Discover tab, find last30days and press Enter. Read the "Will install" list, then choose a scope: user (every project), project (everyone in a repo) or local (this repo only, just you). For a research tool, user scope is usual. If the summary says "Run /reload-plugins to activate", run it.
Confirm it loaded by typing / and looking for last30days in the list. The exact command name can carry a plugin prefix depending on version, so trust what the list shows.
Run /last30days with a topic you care about. The README says the first run launches a setup wizard that can unlock extra sources. Skip anything that needs a paid key. Keys live in ~/.config/last30days/.env or the macOS Keychain, according to the README.
Briefs are saved as Markdown in ~/Documents/Last30Days/ by default. Change that with the LAST30DAYS_MEMORY_DIR environment variable or --save-dir <path>.
/last30days Claude Code returns a general roundup. A narrower topic returns something you can act on:
/last30days switching from [TOOL_A] to [TOOL_B] for [YOUR_USE_CASE]
Flags documented in the README: --quick or --deep for scope, --days <N> for the window, --as-of <DATE> for a historical lookback, --diagnose to check which sources are authenticated, --emit=json for machine-readable output.
A brief is evidence, not a verdict. Paste it into a fresh conversation with this prompt.
You are a careful research analyst advising [YOUR_ROLE] at [YOUR_ORGANISATION].
Context: I am deciding [THE_DECISION], and the options are [OPTION_A] and [OPTION_B]. Below is a brief of what was published and discussed during [DATE_RANGE]. Treat it as your only source of current facts.
Goal: tell me what the evidence supports, what it contradicts, and what I still do not know.
Steps:
1. List the five to eight most decision-relevant claims in the brief, each with the source type it came from (forum thread, repository, video, news, prediction market).
2. Mark each claim as repeated across independent sources, or single-source.
3. Separate hard facts (releases, prices, dates) from opinion and sentiment.
4. Note anything that looks promotional, outdated or contradicted elsewhere in the brief.
5. Give a recommendation with a confidence level, and the two facts that would change it.
Format: a table of claims (claim, source type, repeated or single, fact or opinion), then a short recommendation.
Rules: use only the brief; do not fill gaps from memory; if you need an input I have not given you, ask before answering; do not invent figures or quotes.
Before answering, check that every claim in your recommendation traces to a line in the brief.
BRIEF:
[PASTE_BRIEF_HERE]
Fill in your role, decision, options, date range and the pasted brief.
--diagnose and confirm the sources you expected are active, rather than assuming.--days 7 and again at 30. Findings that survive both windows are stronger than ones that appear only in one.| Play | Topic phrasing | What you decide |
|---|---|---|
| Tool switch check | [TOOL_A] vs [TOOL_B] for [USE_CASE] |
Whether recent complaints affect your use case |
| Pre-meeting briefing | [PERSON_OR_COMPANY] |
What to ask about, from public activity only |
| Content gap scan | [YOUR_TOPIC] questions |
What people are asking that nobody has answered |
| Launch watch | [COMPETITOR] launch |
Whether to respond, wait or ignore |
| Prompt refresh | [MODEL_OR_TOOL] prompting tips |
Which tips are recent, versus recycled |
yt-dlp. Check --diagnose rather than assuming.