Claude Scholar: The Research Co-Pilot Built for Scientists Who Want to Stay in Control

A few weeks ago we wrote about AutoResearchClaw — a fully autonomous 23-stage pipeline that takes a research topic and returns a conference-ready paper with minimal human involvement. We ran it. The results were interesting, the artifacts were real, and the experience surfaced some honest limitations of fully hands-off research automation.

Claude Scholar takes a different position. Same goal — accelerating academic research with AI — but a fundamentally different philosophy: keep the human in the loop at every critical decision point.

What Claude Scholar Is

Claude Scholar is a research assistant framework built on top of coding agents — specifically Claude Code, Codex CLI, and OpenCode. It provides a structured set of skills, agents, and slash commands that guide you through the full research lifecycle:

• Literature review — search, import via Zotero MCP, annotate, synthesize
• Ideation — hypothesis generation with literature grounding
• Coding & experiments — write code, run experiments, version results
• Results analysis — strict statistical analysis, real scientific figures, appendices
• Writing — ML paper writing with a shared paper-miner writing memory
• Review response — structured replies to peer reviewer comments
• Project knowledge management — Obsidian-based knowledge base, auto-synced to repo

It’s model-agnostic in practice: the main branch uses Claude Code, there’s a codex branch for OpenAI Codex CLI (40 skills, 14 agents), and an opencode branch for the OpenCode alternative.

The Architecture: Skills, Agents, Commands

Claude Scholar’s structure mirrors how a well-organized research group actually works — specialized roles, not a single generalist agent doing everything.

Skills are focused capabilities injected into the coding agent’s context:

• literature-reviewer — systematic search, Zotero integration, citation export
• research-ideation — hypothesis generation grounded in retrieved literature
• ml-paper-writing — structured writing with shared writing memory
• results-analysis — strict statistics, figure generation, appendix scaffolding
• review-response — point-by-point reviewer response drafting

Slash commands are one-shot workflows:

• /zotero-review — import papers from Zotero, extract key claims
• /zotero-notes — structured annotation from Zotero library
• /analyze-results — single command for analysis + report generation (replaces the old data-analyst entry point)
• /mine-writing-patterns — scans existing papers to build a paper-miner writing memory that improves future writing outputs

The Obsidian knowledge base (added March 17) is particularly well-designed. Rather than relying on MCP or external services, it’s filesystem-first: project knowledge is organized into Papers / Knowledge / Experiments / Results / Writing folders, bound to the repo, and auto-synced. Round-level experiment reports land in Results/Reports/. Everything is durable and version-controlled without cloud dependencies.

How It Compares to AutoResearchClaw

The contrast with AutoResearchClaw is instructive.

	AutoResearchClaw	Claude Scholar
Philosophy	Fully autonomous	Semi-automated, human-in-loop
Input	Research topic → paper	Human drives each stage
Stages	23 automated stages	Modular skills on demand
Human role	Approve or reject output	Active co-pilot at each step
Strength	Speed, breadth	Quality, rigor, control
Best for	Rapid prototyping, idea validation	Publication-track research

AutoResearchClaw is impressive for generating a complete paper artifact quickly — useful for exploring whether an idea is worth pursuing. Claude Scholar is for when you’ve decided the idea is worth pursuing and you want AI assistance that amplifies your judgment rather than replacing it.

The distinction matters because research quality lives in the decisions that only experts can make: which papers are actually relevant, whether a statistical approach is appropriate for the data, whether the framing of a contribution is honest about its limitations. Claude Scholar keeps those decisions with the researcher.

The Writing Memory System

One of the most interesting recent additions (March 18) is the paper-miner writing memory.

The /mine-writing-patterns command scans a corpus of existing papers — your own previous work, top-cited papers in your field, or any set of PDFs — and extracts writing patterns: how abstracts are structured, how contributions are framed, how results are hedged appropriately, how related work is organized.

This writing memory is then shared across ml-paper-writing and review-response skills, so every writing task benefits from the accumulated patterns. It’s a form of style grounding that reduces the generic, floaty prose that LLMs tend to produce when writing academic papers cold.

For researchers who have a defined writing style or work in a field with specific conventions (NeurIPS vs. Nature vs. clinical journals), this is a practical way to make the AI output feel less foreign.

The Zotero Integration

For researchers already using Zotero as their reference manager, the MCP integration is genuinely useful. You can:

• Import papers directly from DOI, arXiv ID, or URL into your Zotero library
• Access your full Zotero library from within the agent workflow
• Export formatted citations in any style
• Annotate and tag papers without leaving the research workflow

The March 26 update added Web API mode for remote Zotero access — meaning you can work from a server or Railway deployment without needing Zotero running locally.

Who It’s For

Claude Scholar is explicitly built for CS and AI researchers — the examples, commands, and defaults are oriented toward ML papers, experiment-heavy research, and the NeurIPS/ICLR/ICML conference circuit. The ml-paper-writing skill reflects this.

That said, the architecture is general enough that researchers in adjacent fields (computational biology, quantitative finance, clinical AI) could adapt it with minimal effort. The Obsidian knowledge base and Zotero integration are domain-agnostic.

If you’re a researcher who:

• Writes papers regularly and feels the friction of switching between literature, code, experiments, and writing
• Uses Claude Code or Codex CLI for coding work already
• Wants AI assistance that stays in your workflow rather than replacing it

— Claude Scholar is worth installing.

Getting Started

# Clone the repo
git clone https://github.com/Galaxy-Dawn/claude-scholar
cd claude-scholar

# Run the setup script (backup-aware, won't clobber existing ~/.claude config)
bash setup.sh

# For Codex CLI users
git checkout codex

# For OpenCode users
git checkout opencode

The setup.sh installer does an incremental merge into ~/.claude — it won’t overwrite your existing Claude Code settings.

---

How This Fits Into the Broader Picture

We’re in an interesting moment for research automation. The fully autonomous end of the spectrum (AutoResearchClaw, AI Scientist from Sakana AI) can generate paper-shaped artifacts quickly but struggles with the judgment calls that determine whether research is actually good. The pure coding assistant end (vanilla Claude Code) gives you AI capability without research-specific scaffolding.

Claude Scholar occupies the middle deliberately — semi-automated, modular, human-centered. For researchers who take publication seriously, that’s probably the right place to be.

Links:

• GitHub: Galaxy-Dawn/claude-scholar
• Our AutoResearchClaw experiment — the fully autonomous alternative
• AutoResearch pipeline writeup — what happened when we ran it