Command reference ยท Claude Code
/deep-research
Fans out web searches for a question, fetches and cross-checks sources, and produces a cited report.
/deep-research is a bundled dynamic workflow: a script written by Claude orchestrates multiple subagents in the background that fan out web searches on a question from different angles, retrieve the sources found, and cross-check them against each other. Every claim is voted on; claims that don't survive the cross-check are dropped from the report. The result is a cited report instead of a turn-by-turn transcript.
The workflow runs in the background while the session stays free, and it requires the WebSearch tool. Since version 2.1.218, /deep-research only starts on manual invocation โ before that, Claude could also start the workflow on its own when it considered the research useful.
/deep-research differs from a simple web search through the bundling of many subagents with mutual verification, and from /code-review or /security-review in that it analyzes open research questions instead of code.
โ WHEN TO USE IT?
Comprehensive, multi-source research with evidence for an open question
/deep-research what are current best practices for LLM prompt caching in production?
Let research run in the background while working on something else
/deep-research compare pricing models of major vector database providers
โ WHEN NOT TO?
Quick single search for a concrete fact
the multi-agent fan-out across many subagents is unnecessary overhead for a single quick search
Better: use /web-search for the quick single search
Expecting Claude to start the research on its own as soon as an open question comes up
Since version 2.1.218, /deep-research only starts on manual invocation; Claude no longer triggers the workflow on its own.
Better: Call /deep-research explicitly with the question each time.
SOURCES
- Claude Code docs (official) โ code.claude.com
MORE COMMANDS ยท Claude Code