research-pro-maxlisted

Universal research engine for any question that needs real evidence — technical, library/API, codebase, security, product, market, competitor, pricing, academic, legal-adjacent, company due diligence, social/trend, or content research. Use whenever the user says research, investigate, find out, dig into, compare options, evaluate, "is X worth it", "what's the best X", "how does X actually work", "who is X", "should we use X", "should I build this", "research kro", "pata karo", "check kro sahi hai ya nahi" — before any build/buy/adopt/migrate/publish decision that is expensive to reverse, and whenever an answer must be sourced rather than recalled from memory. It names the venues that can actually settle the question before searching, opens primary sources, verifies each claim against the line that supports it, triangulates, and delivers an answer-first brief with per-claim citations, confidence, and named gaps. Prefer this over ad-hoc web searching any time the answer will be acted on.
xniperbuilds/research-pro-max · ★ 0 · AI & Automation · score 70

Install: claude install-skill xniperbuilds/research-pro-max

# Research Pro Max A working system for one question: **what is actually true here, how do I know, and what should be done about it?** Most "research" output fails in ways that are invisible until someone acts on it. This skill exists to make those failures impossible to reach by accident. Built from primary sources: Anthropic's published multi-agent research architecture, the 2026 literature on deep-research agent failure (citation hallucination taxonomies, source-attribution evaluation, trajectory-level hallucination), Caulfield's SIFT and the lateral-reading research, PRISMA and evidence-hierarchy methodology, CIA structured analytic tradecraft, OpenSSF dependency evaluation guidance, NN/g research-method selection, and standard market/competitive intelligence practice. Sourcing with dates: `references/sources.md`. --- ## Prime directive > **Every load-bearing claim carries a source that was actually opened, and a confidence that was actually earned. Research ends in a verdict, not a link dump.** Two halves, both mandatory. **Provenance half.** A citation is a promise that someone read the thing. Search snippets, titles in a result list, and recalled knowledge are *leads*, not sources. The 2026 measurements are worth internalising, because they point at a failure most people defend against in the wrong place: - Frontier models keep **link validity above 94%** and topical relevance above 80% — so "the link works" proves almost nothing. - The same models score only