Stop getting confident-sounding answers you can't verify.

Ask Claude to "research something" and you get a fluent summary — with no way to tell which claims came from a real source, which numbers are stale training data, and which are made up. For a business decision, that's a problem.

Research-to-Report replaces vibes with a pipeline. Every run follows the same auditable process: scope the question into sub-questions; search from multiple angles so the report isn't hostage to one SEO bubble; grade every source T1 (primary/official) to T4 (promotional junk), with money-related numbers requiring T1/T2 evidence; report contradictions instead of averaging them; and produce a report with a TL;DR of conclusions, [n] citations throughout, a mandatory Contradictions section, a mandatory Limitations section, and a numbered source list with tiers, dates, and access dates.

Four built-in report templates: landscape scan, comparison/recommendation, literature brief, status report. The skill picks the right one for your question.

Who it's for: consultants, indie hackers validating markets, PMs writing option memos, anyone who has to defend their research to someone else.

Honest limits: it reads the public web — it won't access paywalled databases, and it can't make a thin evidence base thick (it will tell you when evidence is thin — that's the point).

No API keys — works with plain Claude Code (WebSearch/WebFetch).