Verified against Claude · 2026-07-28
Get a web-search research brief where every claim traces to a real, checkable source
A prompt for Claude's web search feature that requires an inline citation on every substantive claim, a source-quality breakdown of primary versus secondary sources, and an explicit list of what search could not actually verify, instead of a fluent answer that blends real results with plausible filler.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Use web search to research this — do not answer from what you already know without checking, even if you're confident. This needs to hold up if someone clicks through every citation. RESEARCH QUESTION Has remote work adoption been shown to measurably change output per knowledge worker, according to research published since 2023? SEARCH SCOPE Peer-reviewed studies, government labor statistics, and reporting from established business/economics outlets — not opinion pieces or single-company case studies presented as general findings. Prefer primary sources and recent material within this scope over aggregator or listicle-style pages that themselves cite no one. GEOGRAPHIC OR INSTITUTIONAL DIVERSITY If every source found so far comes from the same outlet, institution, or point of view, say so explicitly and search further before concluding — a single institution's framing repeated across several of its own follow-up pieces or syndicated copies is not independent corroboration, even if it looks numerically like several sources agreeing. EXCLUDED SOURCE TYPES Exclude LinkedIn posts, company blog claims about their own product's impact, and any source that does not name its methodology. WHAT TO PRODUCE 1. A direct answer to the research question in three sentences or fewer, up front, with each substantive claim carrying an inline citation to the specific source it came from — not a claim followed by a general "sources suggest" with no pointer to which one. 2. A short section titled SOURCE QUALITY, one line per source actually cited: name, approximate recency, and whether it is a primary source (the original study, filing, or statement) or a secondary one reporting on something else — a reader deciding how much to trust this needs to know which kind of source is doing the work. 3. A section titled WHERE SOURCES DISAGREE — if two sources you found say materially different things, name both explicitly and state the disagreement rather than picking a side silently or averaging them into one smoothed claim. 4. A section titled COULD NOT VERIFY — anything relevant to the question that search did not turn up a checkable source for. This includes things you personally believe are true from general knowledge but that the search results themselves did not actually support; say so rather than quietly asserting it as a searched-and-confirmed fact. CONSTRAINTS - Never cite a source for a claim it does not actually support just to avoid leaving a sentence uncited — an unsupported claim belongs in COULD NOT VERIFY, or should be dropped, not stapled to the nearest plausible-looking source. - If search results are thin, contradictory, or mostly low-quality for this specific question, say that plainly at the top rather than presenting a confident-sounding answer built on weak footing — thinness of evidence is itself part of the answer. - Prioritize anything published in the last 18 months; note the publication date on every cited source explicitly. OUTPUT LENGTH Under 400 words total, not counting the source quality list.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Explicitly instructing Claude to use search rather than answer from confident prior knowledge matters because for a well-known topic type the model has real training-data familiarity, and left unconstrained a 'research this' ask can quietly blend actual retrieved snippets with recalled-from-training claims into one seamless paragraph, with no way for the reader to tell which sentence came from an actual click-through-able source and which came from the model's own prior. The SOURCE QUALITY section, distinguishing primary from secondary sources, matters because a search-grounded answer is only as good as what it actually retrieved, and a bibliography of secondary sources reporting on other secondary sources — a common shape for search results on a popular topic — reads as well-cited while tracing back to zero or one primary source; naming the type explicitly surfaces that instead of letting a citation count stand in for evidentiary weight. COULD NOT VERIFY as a mandatory section, rather than an implicit gap, prevents the output from reading as more complete than the search actually supports: a synthesis that only reports what it found suggests by omission that it covered the question fully, and naming what search did not turn up — including anything the model privately believes from training but cannot point to a source for — closes that gap explicitly rather than leaving it invisible. The instruction that thinness of evidence is itself part of the answer directly counters the model's default incentive to always produce a confident, well-organized-sounding response regardless of how weak the underlying material actually was, which is precisely the situation where a fluent wrong-feeling-right answer is most dangerous, because it does not read as uncertain even when the search behind it genuinely was. The instruction against mistaking repeated framing from one institution for independent corroboration targets a failure specific to how web search results are actually shaped: results for a topic often surface an outlet's original piece plus several follow-ups or syndicated copies of that same piece, which can look like multiple confirming sources while really being one source republished, and treating that as broad agreement inflates confidence the evidence does not actually support.
Verified against
Claude Sonnet 4.6 (web search) · 2026-07-28
Changelog
- 2026-07-28 — Initial publish, verified against Claude Sonnet 4.6 with web search enabled.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
