deep-research

This is a plugin for Claude Code, a command-line AI coding tool, that runs research questions through three separate AI systems at once instead of just one. The idea is that any single AI can confidently state wrong things, especially when its training data overlaps with other AIs and they all learned the same incorrect information from the same sources. By running three different AI families in parallel (Claude, Google's Gemini, and OpenAI's GPT-5) and comparing what they say, the plugin has a better chance of catching errors.

When you type /research followed by a question, the plugin spawns a researcher agent that queries all three AI systems using their own separate web search tools. Each system searches the web independently. After all three return answers, the researcher sorts the claims into categories: things all three agreed on, things only one said, and things where they disagreed. Before including any claim in the final report, the plugin checks whether a real URL supports it, fetching the page and verifying that the quoted passage actually appears there word for word.

The plugin is designed to defend against several specific failure modes. Fabricated citations, where an AI confidently cites a paper or URL that does not exist or does not say what was claimed, are caught by the mechanical URL check. Claims about command-line tool flags or API syntax are verified by actually running the tool's help output rather than trusting the AI's memory. Content fetched from web pages is treated as untrusted data to prevent malicious instructions on a page from hijacking the research session. The entire researcher process runs in an isolated copy of the project so it cannot touch your actual files.

The final report includes a summary, numbered findings with source URLs, a confidence rating per topic, claims that were rejected after verification, and a flag for claims where all three AIs agreed but no primary source could be found, which the plugin treats as a higher hallucination risk rather than a confirmation.

Installation involves cloning the repository into a local Claude Code plugins folder and restarting the tool. The two optional cross-checker CLIs must be installed separately.