Deep Research

Deep Research is an official Anthropic agent skill that transforms Claude Code into an autonomous research assistant capable of thoroughly investigating topics across codebases, documentation, and project structures. Unlike standard Claude interactions where research context accumulates in the main conversation window, Deep Research operates in a forked Explore agent context -- isolating its investigation in a dedicated subagent that can freely navigate files, search patterns, and analyze code without polluting the primary conversation's context window. The result is comprehensive, source-grounded research summaries delivered back to the main session with specific file references and structured findings. The skill leverages Claude Code's context fork architecture, which creates an entirely separate execution environment with its own context window. When invoked, the Explore agent subagent receives read-only tools optimized for codebase exploration: Glob for pattern-based file discovery, Grep for content searching with full regex support, and Read for examining file contents. This constrained toolset is intentional -- by limiting the agent to observation-only operations, Deep Research ensures it cannot accidentally modify source code, configuration files, or project state during its investigation. The agent systematically works through its research task, building understanding incrementally across multiple files and directories before synthesizing its findings. What makes Deep Research particularly effective is its multi-step investigation methodology. Rather than performing a single search and returning surface-level results, the skill instructs Claude to follow a disciplined research workflow: first discovering relevant files through targeted glob and grep patterns, then reading and analyzing the most pertinent code sections, and finally summarizing findings with precise file path references so the developer can verify and act on the research. This mirrors how an experienced engineer approaches unfamiliar code -- methodically narrowing scope from broad directory structure down to specific implementations, cross-referencing patterns across modules, and building a coherent mental model before drawing conclusions. The skill is invocable both automatically (when Claude detects a research-oriented question) and manually via the /deep-research slash command followed by the topic to investigate. Arguments passed after the command are injected directly into the research prompt through the $ARGUMENTS substitution mechanism, making it straightforward to target specific areas of investigation. With the anthropics/skills repository holding over 83,000 stars on GitHub, Deep Research has become one of the most widely adopted skills in the Claude Code ecosystem, valued by developers who need thorough codebase analysis without manual file-by-file navigation.