Deep research workflows

How it worksCopy Anchor Link

A deep research system plans sub-questions, runs searches in parallel, fetches and reads pages, extracts claims with their sources, checks claims against each other and writes a cited report. In Claude Code, the built-in /deep-research workflow follows this shape and runs verifier agents that vote on each claim, filtering ones that fail cross-checking. The structure matters more than the brand: parallel gathering, independent verification, citations.

Using it wellCopy Anchor Link

Ask a specific question with a scope and time frame. Check the citations of any claim you will rely on, since a citation proves a source exists, not that it says what the report claims. Prefer primary sources. Expect cost in tokens to be many times a single chat answer.

A worked example: reading a research reportCopy Anchor Link

Claim: "Filings fell 8 percent in 2024 [3]." Click source 3. Does the page really say 8 percent, and for which group? Often it says something close but narrower. Verify the three claims your decision depends on. A citation shows where the model got it, not that it understood it.

Writing a good questionCopy Anchor Link

Specific scope, time frame and what counts as evidence. "Compare 2023 and 2024 filing counts using government statistics" is better than "is the tax rule working?"

How verification voting worksCopy Anchor Link

A verifier is a separate agent that checks one claim. Take the claim "Filings fell 8 percent in 2024." Three verifiers each search for support on their own. One rule, an invented policy, keeps a claim only if at least 2 of 3 find a source that states it. Here two find it and one finds a figure of 6 percent for a smaller group. The claim stays, perhaps with the dispute noted.

The weakness: votes are only independent if the sources are. If all three verifiers land on the same blog post, three votes are really one.

Step by step: running a question you can trustCopy Anchor Link

1. Write the question with a scope, a period and the kind of evidence you accept.
2. Ask for the plan of sub-questions first. Add any angle that is missing.
3. Run the research.
4. Mark the three claims your decision depends on.
5. Open each cited page and search it for the exact figure and group.
6. For any that fail, ask for a re-run that uses primary sources only.

What goes wrong: citations that trace to one originCopy Anchor Link

Sources 2, 3 and 5 all state the same figure in similar words. They look like three confirmations. Often two of them quote the third, and that one quotes an old press release. To confirm, open each source and follow its own citation backward. If the trail ends at one page, you have one source. Also check dates: a page from 2022 cannot support a claim about 2024.

When deep research is the wrong toolCopy Anchor Link

A single fact, such as a city's population, needs one search. A question whose answer lives in your own private files needs retrieval over those files, because the web has nothing on it. A decision due in five minutes cannot wait for a run that costs many times a normal answer. Use it for open questions where several sources must agree.

What the cost looks likeCopy Anchor Link

Parallel gathering reads a lot. Invented numbers: 5 sub-questions, 10 pages each, about 3,000 tokens a page. That is 5 x 10 x 3,000 = 150,000 tokens of reading, before the verifiers fetch pages of their own. A normal chat answer might read 5,000 tokens and write 1,000. You pay for breadth, so a narrower question is both cheaper and easier to check. Cutting the scope from five sub-questions to two cuts the reading to 60,000 tokens.

Common mistakeCopy Anchor Link

Treating a long cited report as proof. Check the sources that matter.