Claude Deep Research Agent

June 9, 2026 · View on GitHub

See example outputs here: https://claude-code-deep-research.vercel.app/

Table of Contents

  1. Why This Exists
  2. Repo Structure
  3. Version 4.0
  4. Quick Start
  5. How It Works
  6. Customization
  7. Credits & Acknowledgements
  8. License

UPDATE: Added Claude2.md — updated for deeper research and closer Graph-of-Thought patterns.

Version 4.0 (June 2026)

V4 replaces thinking scaffolds with verification scaffolds: deterministic stop rules and citation gates (scripts), cross-model verification, consequence-tier confidence floors, run cards, and signed residue statements. See Version4/README.md. V3 remains unchanged below for existing users.

Recommended default for new research: start with V4. Use V3 only when you need the original public Graph-of-Thoughts workflow or are comparing V3 against V4.

Why This Exists

Large Language Models (LLMs) excel at single queries but struggle with complex, multi-step research requiring iterative querying, source verification, and citations—what OpenAI and Google call "Deep Research." Anthropic’s Claude Code can achieve the same results, provided the right instructions. This repo supplies those instructions, streamlined into an easy-to-use workflow.

Repo Structure

File/FolderPurpose
CLAUDE.mdMaster instructions for Claude Code. Includes Graph-of-Thoughts integration and deep-research methodology.
Deep Research Question Generator System Prompt.mdChatGPT system prompt (o3/o3-pro) refining raw questions into structured prompts (OpenAI format recommended).
deepresearchprocess.mdComprehensive 7-phase deep research playbook inspired by OpenAI & Google Gemini, foundational to CLAUDE.md.
.template_mcp.jsonOptional MCP server configuration for local filesystem and browser automation with Claude.
examples/Sample refined questions and completed Claude reports compared to other outputs.
Version4/Current verification-gated workflow: control file, installable agents, scripts, schemas, prompt refiner, skill wrapper, benchmark artifacts, and deployment reports.
AGENTS.md / GEMINI.mdPointers for Codex/Gemini-style surfaces to read Version4/skills/deep-research/SKILL.md and follow the V4 control logic.

Quick Start

  1. Read Version4/README.md for the V3 → V4 changes.
  2. Use Version4/CLAUDE.md as the controller and Version4/agents/*.md for the six research roles.
  3. Keep Version4/scripts/ and Version4/schemas/ with the run so stop_rule.py, citation_audit.py, run cards, ledgers, and residue statements are available.
  4. Start with: Deep research [your question].
  5. Before finalizing, run the deterministic gates and record any unresolved claims in the residue statement.

Legacy V3 workflow

Example output and comparisons from the examples folder: https://claude-code-deep-research.vercel.app/

Step 1: Refine Your Question with ChatGPT (or your favorite LLM)

  1. Open ChatGPT (model o3/o3-pro or other thinking models work best).
  2. Set the system prompt to the contents of Deep Research Question Generator System Prompt.md.
  3. Paste your raw research question into the user prompt.
  4. Respond to any clarifying questions from ChatGPT.
  5. Copy the generated OpenAI-formatted prompt.

Step 2: Prepare Claude Code

  1. Launch a new Claude Code session.

  2. Set the model using /model opus.

  3. Type:

    Please read the CLAUDE.md file and confirm when ready for my deep research question.

  4. Wait for Claude’s confirmation.

Step 3: Launch Deep Research

  1. Paste your refined question.

  2. Claude autonomously performs:

    • Research planning with Graph-of-Thoughts.
    • Spins up multiple subagents to do the work faster
    • Iterative search and data scraping.
    • Fact verification and cross-referencing.
    • Markdown report generation with citations and bibliography.

  3. Review and refine as needed.

Tip ⚡: Include directory instructions, such as:

"Save all outputs in the /RESEARCH/[topic] folder."

Bonus Step (Optional)

After obtaining the report, instruct Claude to convert it into a user-friendly website format for enhanced accessibility and readability.

How It Works

Workflow Overview

[ Question Refinement ] → [ Claude Code / agent surface ] → [ V4 verification-gated research pipeline ] → [ Cited Markdown Report + Evidence Ledger + Run Card + Residue ]
  • Version4: Current workflow. It uses branch quotas, disjointness checks, deterministic stop rules, citation audits, cross-model verification, consequence-tier confidence floors, run cards, and signed residue statements.
  • DeepResearchProcess: Implements a 7-phase pipeline—Scope → Plan → Retrieve → Triangulate → Draft → Critique → Package.
  • Graph-of-Thoughts: Allows Claude to branch and merge multiple reasoning paths rather than relying on linear chains.
  • CLAUDE.md: Integrates instructions, enabling Claude to autonomously select tools, verify information, and embed citations systematically.

Build Notes

  • Research Methodology: Derived from OpenAI and Gemini’s deep-research playbooks.
  • Graph-of-Thoughts Integration: Adapted from Graph-of-Thoughts to support dynamic research pathways.
  • Prompt Generation: ChatGPT-based structured prompt ensures clarity, reducing confusion during Claude’s research by over 50% in tests.
  • Automation Hooks: The .template_mcp.json demonstrates local automation options via MCP servers, enabling advanced Claude operations.

Customization

  • Output Styles: Adjust formatting and citation preferences directly within the CLAUDE.md file.
  • Model Flexibility: Alternative Gemini-specific prompts provided by the ChatGPT system prompt generator if preferred.
  • Tool Integration: Expand automation via MCP by updating .template_mcp.json and referencing additional tools within CLAUDE.md.

Credits & Acknowledgements

License

MIT License. See LICENSE file for full details.