Deep Science Writer is an industrial-grade, end-to-end scientific research pipeline and AI agent skill. Designed for the Hermes/ECC framework, it completely automates the literature review process: from background full-text database querying down to anti-hallucination verification, Nature/Science-level peer review, .docx compilation, and Obsidian/NotebookLM knowledge base ingestion.

• Precision Background Sourcing (Two-Stage Pipeline): Subagents first fetch and screen a large pool of abstracts (e.g., 100+). For the highly relevant subset, the agent downloads and deeply reads the full text (Methodology, Results) to prevent abstract-induced hallucination. FULL-TEXT verification of final claims is absolutely mandatory.
• Strict Quality Control: Explicitly targets Q1-Q2 journals. Marks Q3 when strictly necessary. Bans all Q4 and MDPI publications.
• Strict Compliance: The agent is hard-coded to strictly follow every step in order. Skipping phases or taking shortcuts (like drafting without full-text verification) is strictly forbidden.
• Zero-Hallucination Guarantee (Phase 4.5): Automatically runs live HTTP requests tests against every generated DOI to ensure 100% validity. Cross-references generated claims against raw full texts to prevent AI overstatement.
• Automated Peer Review: Integrates the remi peer-review skill (named in tribute to my academic advisor, Remi Chauvy) to aggressively strip "AI fluff" (e.g., "delve", "tapestry") and enforce rigorous academic tone.
• English Output Only: Enforces strict English language generation for all academic reports and drafts, regardless of conversational language.
• Hands-Free Output: Programmatically builds a fully formatted Microsoft Word document (.docx) with APA 7th hanging indents and auto-generated data visualizations.
• Knowledge Management Loop: Automatically saves research summaries to your local Obsidian Vault and explicitly uploads every individually cited reference as a separate source into Google NotebookLM for precise audio overviews and cross-referencing.

To run this skill successfully, your host environment must be configured with the following dependencies and Model Context Protocol (MCP) servers:

• Hermes Agent (or a compatible ECC/Claude Code runner).
• Node.js (v18+ recommended) and npx for running MCP servers.
• Python 3.10+ (in your system PATH or agent's virtual environment).

The agent relies on Python to generate documents and verify links.

pip install python-docx PyMuPDF requests matplotlib seaborn pandas

Add these to your config.yaml or claude_desktop_config.json:

• Scopus MCP (scopus-mcp)
  • Required for premium literature retrieval.
  • API Key: Requires a free Elsevier Scopus API Key (SCOPUS_API_KEY). Apply at the Elsevier Developer Portal.
• NotebookLM MCP (notebooklm-mcp-server)
  • Required for Phase 7 knowledge ingestion.
  • Auth: Must run npx notebooklm-mcp-server auth once in your terminal to authenticate your Google Account locally.
• Exa Search MCP (Highly Recommended)
  • Used for neural search fallback and broad open-access discovery.
• GitHub MCP & Playwright MCP (Optional but recommended for broader functionality).

• Obsidian: The skill looks for %OBSIDIAN_VAULT_PATH%\Hermes\ (fallback: %USERPROFILE%\Documents\Obsidian Vault\Hermes\). Update SKILL.md Phase 7 or set the environment variable if your vault is located elsewhere.
• Output Drive: All outputs strictly save to D:\Tommy (or your configured environment).

Clone this repository into your agent's skills directory:

cd <AGENT_SKILLS_DIR>
git clone https://github.com/CYC2002tommy/deep-research-agent.git

(Replace <AGENT_SKILLS_DIR> with your agent's skills path, e.g., ~/.hermes/skills/ or .agents/skills/ for the ECC framework).

1. Phase 0 & 0.5 (Plan & Background Execution): The agent formulates a search plan and halts for your explicit approval. Once approved, it launches a local background process (terminal(background=true)) to scrape the full text of exactly 30 high-impact papers via APIs.
2. Phase 1 (Discovery): Enforces journal quality limits (Q1-Q2 only, bans MDPI).
3. Phase 2 (Deep Extraction): Consolidates metadata, full texts, and key findings.
4. Phase 3 (Structural Drafting): Outlines the article with evidence-backed claims and APA 7th citations.
5. Phase 4 & 4.5 (Anti-Hallucination): Strips AI vocabulary. Pings all DOIs to ensure they resolve (404 = citation deleted).
6. Phase 5 (Remi Review): An internal peer-review loop that critiques and rewrites the draft until academic standards are met. (Named in tribute to my academic advisor, Remi Chauvy).
7. Phase 6 (Compilation): Python scripts draw Mermaid/Matplotlib charts and compile the final .docx.
8. Phase 7 (Knowledge Graph): Updates Obsidian and pushes papers to NotebookLM.

Simply trigger the agent with a research prompt:

"Please use the deep-science-writer skill to research the sociological and psychological acceptance of sustainability policies, segmented by age (public) and firm size (SMEs vs Large Enterprises)."

The agent will take over, present a blueprint, ask for your approval, and then execute the entire pipeline autonomously.

License: MIT