Turn a question into a cited report with the deep-research skill
A single web search hands you the first plausible answer; deep research fans out across dozens of sources, tries to break each claim, and hands back a report with the citations attached. It is the difference between looking something up and actually investigating it.
The deep-research skill is a research harness: instead of running one search and summarising the top hit, it fans out across many queries, fetches and reads the sources, tries to disprove each claim it has gathered, and then synthesises the survivors into a written report with citations. The output is not a chat answer — it is a structured, sourced document you can hand to someone else. Where an ordinary search optimises for a fast first answer, this skill optimises for an answer that holds up when you check the footnotes.
Why it earns the Skill slot today. Most of the skills in this rotation so far have been about producing a format — a spreadsheet, a PDF, a Word file. This one is about producing confidence. The value is not that Claude can search; it is that the skill builds in the steps a careful researcher does by hand and a hurried one skips: casting a wide net, reading past the first result, and actively looking for the source that contradicts the tidy story. That adversarial step is the part people most often leave out, and it is the part that catches the wrong answer before it reaches you.
Fan-out, not one-shot. The skill decomposes your question into many sub-queries and runs them in parallel, so a question like “which managed Postgres offering fits a small team on a tight budget” pulls pricing pages, quota docs, migration guides, and third-party comparisons at once rather than one blog post. Breadth is what surfaces the disagreement between sources — and the disagreement is usually where the real answer lives.
Adversarial verification. Gathered claims are not taken at face value. The harness cross-checks them against other sources and discards or flags the ones that do not survive, which is how it avoids laundering a single confident-but-wrong page into a confident-but-wrong conclusion. You get fewer claims, but the ones that remain have been pressure-tested rather than merely collected.
Cited synthesis. The final step assembles what survived into a report that keeps the sources attached, so every non-obvious claim points back at where it came from. That is what makes the output auditable: you, or whoever you forward it to, can follow a footnote and check it instead of trusting the summary. A report you can verify is worth far more than a paragraph you have to take on faith.
The trap to avoid: deep research is only as good as the question you give it, and a vague question wastes its breadth. “What car should I buy” has no budget, no use case, and no region, so the fan-out scatters. The skill is built to push back here — if the question is underspecified it will ask two or three clarifying questions before it spends the effort — so answer them rather than skipping past, and scope the question yourself when you can. The other trap is reaching for it on lookups: if one search would settle it, the full harness just adds latency for no gain. Save it for questions where being wrong would actually cost you something.
The try-it block runs the skill on one tightly scoped question so you can see the difference between a search and an investigation.
Invoke the skill with a question that is narrow enough to research but real enough to matter. The scope — budget, use case, constraints — is what lets the fan-out aim instead of scatter.
/deep-research Compare managed Postgres options (AWS RDS,
Neon, Supabase) for a two-person side project under $25/month:
free-tier limits, what triggers the first real bill, and cold-start
behaviour. Cite pricing pages from the last 6 months.Watch what it does that a single search does not: it pulls each provider's own pricing and quota docs alongside independent write-ups, flags where they disagree, and returns a report with the sources attached so you can check the bill-triggering claims yourself. If your question is too broad, expect it to ask a couple of clarifying questions first — answer them; that is the skill aiming the search rather than guessing. For a one-fact lookup, skip it and just ask Claude directly; the harness is for questions where the citations are the point.