Research is a system of questions, not a search bar.
Category · AI
Plans, sources, methods, and synthesis in language Claude can act on. For literature scans, interviews, competitive landscapes, and defensible conclusions.
A fuzzy question gets a fuzzy answer. Sharpen these and Claude researches the right thing, at the right depth.
State what you really want to know as a specific, answerable question, not a topic.
Geography, timeframe, population, sub-questions. Bounds keep the research focused and finite.
A quick orientation vs an exhaustive, cited report. Say which, they're very different jobs.
How current it must be. Flag when only the latest data will do so Claude searches the live web.
A decision memo, a literature review, a blog post, the use shapes depth, tone, and format.
The verb sets the job. Scoping, deep-diving, and fact-checking are different asks.
Get the lay of the land before diving, key players, debates, sub-topics, terms.
Exhaustive investigation of one question across many sources, with citations.
Confirm a specific claim against primary sources and flag what can't be confirmed.
Pull multiple sources into one structured comparison or coherent view.
Condense a long source or body of work. Specify length and what to keep.
Have Claude attack the argument or find weaknesses in a body of evidence.
Keep tabs on a topic over time. Pairs well with a scheduled recurring brief.
Naming source quality gets you evidence, not vibes. Tell Claude what counts.
Primary = original (study, dataset, filing); secondary = reporting on it. Prefer primary for claims.
Vetted by experts before publication. The bar for scientific claims.
Posted before peer review, fast but unvetted. Useful, but flag the caveat.
The body that owns the data (a statistics agency, regulator, the company itself).
An independent second source confirming a claim. One source is a lead, not a fact.
When it was published and where it came from. Old or anonymous = lower trust.
Who funded or benefits from a finding. Affects how much weight it carries.
A traceable link or reference so you can verify it yourself.
How Claude finds material. Steering the search shapes everything downstream.
For anything current, tell Claude to search rather than rely on training data.
Cast wide across many sources before narrowing. Ask for multiple independent queries.
Have Claude vary the search language, jargon, plain terms, opposing framings.
Restrict to trusted domains (gov, edu, specific journals) when quality matters.
Constrain to a time window so stale results don't crowd out current ones.
Trace a claim back to its origin; chase who cites a key paper.
The lenses that separate findings from noise. Name them and Claude reasons, not just collects.
Confirm a finding from multiple independent angles or methods.
Two things moving together isn't one causing the other. A classic trap.
How many, selected how. A vivid anecdote isn't data; a tiny study isn't proof.
The background frequency a claim should be judged against. Guards against scary-sounding stats.
Selection, publication, and framing bias in how evidence is gathered and presented.
Build the strongest version of the opposing view before judging. Beats a strawman.
Ask Claude to state how sure it is and where the evidence is thin.
Keep what's established separate from what's interpreted or argued.
Naming the deliverable gets you a usable artifact instead of a wall of text.
The answer first, in a few sentences, with detail below. Lead with the conclusion.
A structured survey of what's known, where sources agree, and where they conflict.
Each source listed with a short note on what it says and how reliable it is.
Options as rows, criteria as columns. Makes tradeoffs scannable.
Bullet the takeaways and, crucially, what remains unknown.
A full write-up where every claim links to its source. The gold standard.