AI Agents for Academic Research & Writing — From KatmerCode to the Nature Playbook

The gap between “ChatGPT can help me write” and a full AI-powered research pipeline is enormous. Two recent resources bridge it: KatmerCode, an Obsidian plugin that puts 8 research-specific AI skills in your writing sidebar, and a Nature career column by Dritjon Gruda outlining 3 responsible ways to use LLMs for academic writing. Together, they show both the tooling and the mindset for AI-assisted research in 2026.

KatmerCode: Full Research Pipeline Inside Obsidian

KatmerCode integrates Claude Code as a sidebar panel in Obsidian — specifically designed for researchers who write in their vault and want AI assistance without leaving the editor. It runs Claude Code CLI as a subprocess, supports streaming, tool calls, and inline diff editing.

┌─────────────────────────────────────────────────────┐
│  Obsidian Vault                                     │
│  ┌──────────────────┐  ┌────────────────────────┐  │
│  │                  │  │  KatmerCode Sidebar     │  │
│  │  Your Manuscript │  │  ┌──────────────────┐  │  │
│  │  (Markdown)      │◄─┤  │ 8 Research Skills│  │  │
│  │                  │  │  │ /lit-search      │  │  │
│  │  Inline Diffs:   │  │  │ /citation-network│  │  │
│  │  ~~old~~ new     │  │  │ /research-gap    │  │  │
│  │                  │  │  │ /abstract        │  │  │
│  └──────────────────┘  │  │ /journal-match   │  │  │
│                        │  │ /peer-review     │  │  │
│                        │  │ /cite-verify     │  │  │
│                        │  │ /report-template │  │  │
│                        │  └──────────────────┘  │  │
│                        │         │              │  │
│                        │         ▼              │  │
│                        │  Academic databases:   │  │
│                        │  Semantic Scholar,     │  │
│                        │  CrossRef, OpenAlex,   │  │
│                        │  arXiv, PubMed,        │  │
│                        │  Unpaywall             │  │
│                        └────────────────────────┘  │
└─────────────────────────────────────────────────────┘

The 8 Research Skills

Each skill is triggered via slash command and produces structured HTML reports with charts, tables, and interactive elements.

Setup & Requirements

# Requires Claude Code CLI installed globally
npm install -g @anthropic-ai/claude-code

# Clone and build the plugin
git clone https://github.com/hkcanan/katmer-code.git
cd katmer-code && npm install && npm run build

# Copy to your Obsidian vault
cp main.js manifest.json styles.css <vault>/.obsidian/plugins/katmer-code/

Key detail: KatmerCode inherits MCP servers from ~/.claude.json — so if you’ve configured paper-search-mcp, arxiv-mcp-server, or openalex-research-mcp for your terminal Claude Code, they automatically work in the sidebar too.

Important Caveat

The developers emphasize: these are research aids, not oracles. Skills query real databases and apply structured analysis, but the outputs are starting points. They surface overlooked patterns — they don’t replace expert review.

Nature’s 3-Way Playbook for AI in Academic Writing

Dritjon Gruda, a professor of organizational behavior at Universidade Católica Portuguesa, published a widely-shared Nature career column on responsible AI use in academic writing. His three use cases:

1. Polishing Drafts for Clarity and Coherence

Use AI to refine phrasing in papers you’ve already written. The key: you write the content and ideas first, then use the LLM as an editor — like a native-speaker colleague who smooths your prose. This is especially valuable for non-native English speakers.

2. Elevating Peer Review

When reviewing manuscripts, use AI to help organize and articulate feedback — not to generate opinions, but to structure your existing assessment into clear, actionable points. The AI helps you be a better reviewer, not a replacement reviewer.

3. Optimizing Editorial Feedback

As an editor, use AI to make feedback more precise, actionable, and empathetic. The goal: communication quality, not content generation.

The Common Thread

All three use cases share a principle: AI refines your thinking, it doesn’t replace it. You bring domain expertise, original ideas, and judgment. The AI brings language polish, structural consistency, and coverage checks.

Full Research Workflow: Combining Both

Here’s how the two approaches complement each other across the research lifecycle:

Agentic-Researcher: Lightweight Lit Workflow for IS Researchers

A newer, more focused tool by SheeanBen — designed specifically for Information Systems grad students who use Zotero + Obsidian. Less feature-rich than KatmerCode but more opinionated about the daily reading workflow.

Keywords → Auto-expand to IS terms → Search top venues
    → Score & filter papers → Confirm selections
    → Sync PDFs from Zotero → Generate Chinese research notes
    → Daily reading report

5 Key Features

Setup

git clone https://github.com/SheeanBen/Agentic-Researcher.git
cd Agentic-Researcher
cp .env.example .env.local  # Add API keys if using LLM scoring

Requires Python 3.9+, pdftotext (Poppler), Zotero with Attanger plugin for PDF sync. Works offline with heuristic scoring or with OpenAI API for LLM-powered evaluation.

Who It’s For

Best fit: IS/management PhD students who read papers daily in Chinese and use Obsidian for notes. Less general-purpose than KatmerCode but more streamlined for the daily literature grind.

DeepScientist: Local AI Research Workstation

DeepScientist fills a different gap than KatmerCode or Agentic-Researcher — it manages the experiment side of research, not just literature and writing. Set up in 15 minutes, it gives you a persistent, local-first workspace where code, experiments, notes, and paper drafts all live together.

Research question or paper
        │
        ▼
┌──────────────────────────────────┐
│  DeepScientist Workspace         │
│                                  │
│  Git repo per project            │
│  ├── Baseline reproduction       │
│  ├── Branch per experiment       │
│  ├── Ablation studies            │
│  ├── Failed paths preserved      │  ← keeps ALL attempts, not just successes
│  ├── Metrics & traces            │
│  └── LaTeX draft + figures       │
│                                  │
│  Access: Web UI (:20999) / TUI   │
│  IM: WeChat, Telegram, Feishu    │
└──────────────────────────────────┘

Key Features

Setup

npm install -g @researai/deepscientist
codex --login
ds --here
# Access at http://127.0.0.1:20999

Philosophy

“Persistent repository-based projects over ephemeral chats. Human control over black-box automation. Preserved learning from failed paths.”

DeepScientist’s core insight: failed experiments are knowledge too. Most tools discard failed paths. DeepScientist preserves them with full traces, so you (or your advisor) can see why approach A failed and why approach B worked.

How LearnAI Team Could Use This

• Build a research-writing workflow template for LearnAI docs: literature search, citation verification, draft polishing, and peer-review checks before publication.
• Use KatmerCode-style slash commands as a model for internal AI documentation agents that produce structured, source-linked reports.
• Turn the Nature playbook into team guidance: AI may improve clarity, structure, and feedback quality, but authors remain responsible for claims, citations, and judgment.
• Use DeepScientist-style experiment preservation for AI tool evaluations so failed trials, prompts, and benchmark notes remain auditable.

Real-World Use Cases

Other Tools in the Ecosystem

Links

• KatmerCode: github.com/hkcanan/katmer-code
• Agentic-Researcher: github.com/SheeanBen/Agentic-Researcher
• DeepScientist: github.com/ResearAI/DeepScientist
• Nature article: Three ways ChatGPT helps me in my academic writing
• HN discussion: Show HN: KatmerCode
• XDA coverage: Claude Code inside Obsidian <!– REVIEW-TODO: [source_links] Xiaohongshu source link is generic (https://www.xiaohongshu.com/), not a specific post URL — find actual 哈泰利 post or remove –>