research

name: research description: Investigate an open-ended question across the web by orchestrating multi-source search and reading the actual pages. Use whenever the user wants to find out how people use / feel about / discuss something, gather opinions or real-world patterns, survey a topic, or asks to "research X", "调研 X", "看看大家怎么用/怎么看 X". Not for SEO/keyword research (use the SEO skills) or single-fact lookups (search directly).

Run open-ended investigation across the web for the user. The job is to find what real people say and do about a topic, by spreading across sources, iterating on queries, and reading the actual pages — not by trusting search-result previews.

This skill orchestrates; the actual networking is done through other skills. Load them as needed:

Google — general web, articles, forums, threads → use web-access
Reddit — community discussion, firsthand experience → use web-access (navigate into the site)
X — live sentiment, fast-moving takes → use x-search

Goal

Surface the strongest real-world signal on the question: how people actually use it, what they like/hate, recurring patterns, disagreements.
Base every conclusion on content read from inside the source page/post, with citations.
Be honest about signal strength; say when evidence is thin rather than inflating it.

Method

Core search-query discipline: go broad→narrow (add one modifier at a time), rewrite weak/empty/off-target queries instead of stopping (try aliases, abbreviations, alternate wording, common misspellings; do multiple rounds), and stop once the answer is clear or signal quality is established. On top of that:

Start broad, then narrow. Open with the core entity/topic in plain terms. Only add modifiers once you see what the broad pass returns. Don't open with a hyper-specific query unless the user already gave a narrow target.
Spread across the three sources by default (Google + Reddit + X). They surface different things — articles vs. community threads vs. live takes. Don't conclude from a single source unless the others genuinely have nothing.
Iterate, don't give up. Weak/empty/off-target results mean rewrite the query, not stop. Try aliases, abbreviations, alternate wording, common misspellings, X-native forms (hashtags, handles) before lowering confidence.
Adapt source weight to the topic, without assuming what the topic is. The question can be about anything — a product, a person, a place, a health or money or life decision, a cultural trend, a how-to, a controversy. Let the topic decide where the richest signal lives: some questions live in long-form articles (Google), some in lived experience and community threads (Reddit), some in real-time reaction (X). Probe all three, then dig deeper where the signal actually is. Make no default assumption that a topic is technical or any other domain.

Click-in discipline (core rule)

Search-result surfaces are only a candidate list, never a source of conclusions:

Google SERP snippets, the Reddit search/subreddit listing, and the X search stream are entry points. They tell you what to open, nothing more.
Open each promising result and read the real content: enter the Reddit post and read the OP body + top comments; open the article and read the body; open the X post and read the full thread + notable replies; for non-trivial Google hits, click through to the page.
Never quote, summarize, or draw a conclusion from a title, snippet, or list preview alone.
Be selective about what you open so this stays affordable: from each listing, pick the few most relevant/highest-signal items, then read those in full. This selectivity is the same broad→narrow move — cast wide to find candidates, go deep on the best.
Don't guess sub-page URLs. Get URLs by reading or clicking elements on the page (per the web-access rule), not by constructing them.

Signal quality

Prefer firsthand experience and concrete examples over reposted framing or summaries.
Treat many near-identical posts/articles as one repeated claim, not independent confirmation.
On X, weigh engagement as rough signal; don't build broad conclusions on a handful of low-engagement posts unless the user asked for early/niche signal.
State limitations explicitly: thin signal, one-sided sources, mostly derivative content.

Output

Leave the exact format to your judgment — fit it to the question. Whatever the shape, it must:

synthesize themes and patterns rather than dump a list of links
group overlapping findings under one conclusion instead of repeating per source
cite the specific posts/articles/threads you actually read
make signal strength and any limitations visible
follow the project reply-language rules (default 简体中文; keep English quotes, titles, product names, identifiers untranslated)

Avoid

concluding from SERP/listing previews without opening the page
opening everything indiscriminately and burning tokens — select first, then go deep
starting with long over-specified queries
giving up after one weak query instead of rewriting it
leaning on a single source when the others were never tried
counting near-duplicate content as multiple independent signals
dumping raw results with no synthesis