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 7 academic research skills and multiple CLI-backed AI providers 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 CLI-backed AI assistants as a sidebar panel in Obsidian — specifically designed for researchers who write in their vault and want AI assistance without leaving the editor. It supports Claude Code, Gemini, Codex, and Antigravity tabs, plus streaming, tool calls, inline diff editing, and academic research skills.

┌─────────────────────────────────────────────────────┐
│  Obsidian Vault                                     │
│  ┌──────────────────┐  ┌────────────────────────┐  │
│  │                  │  │  KatmerCode Sidebar     │  │
│  │  Your Manuscript │  │  ┌──────────────────┐  │  │
│  │  (Markdown)      │◄─┤  │ 7 Research Skills│  │  │
│  │                  │  │  │ /lit-search      │  │  │
│  │  Inline Diffs:   │  │  │ /citation-network│  │  │
│  │  ~~old~~ new     │  │  │ /research-gap    │  │  │
│  │                  │  │  │ /abstract        │  │  │
│  └──────────────────┘  │  │ /journal-match   │  │  │
│                        │  │ /peer-review     │  │  │
│                        │  │ /cite-verify     │  │  │
│                        │  └──────────────────┘  │  │
│                        │         │              │  │
│                        │         ▼              │  │
│                        │  Academic databases:   │  │
│                        │  Semantic Scholar,     │  │
│                        │  CrossRef, OpenAlex,   │  │
│                        │  arXiv, PubMed,        │  │
│                        │  Unpaywall             │  │
│                        └────────────────────────┘  │
└─────────────────────────────────────────────────────┘

The 7 Academic Skills

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

Setup & Requirements

# Requires at least one supported provider 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 reads each provider’s own CLI and MCP configuration — so Claude, Codex, Gemini, and Antigravity can each keep their native auth, context, and tooling.

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