GitHub - Galaxy-Dawn/claude-scholar: Semi-automated research assistant for academic research and software development. Supports Claude Code, OpenCode, and Codex CLI across ideation, coding, experiments, writing, and publication.

Semi-automated research assistant for academic research and software development, especially for computer science and AI researchers. Supports Claude Code, Codex CLI, and OpenCode across literature review, coding, experiments, reporting, writing, and project knowledge management.

Branch note: the main branch is the Claude Code workflow. If you use Codex CLI, please see the codex branch. If you use OpenCode, please see the opencode branch.

Recent News

• 2026-05-14: expression-skill made the communication core, planning-with-files restored as the default persistence layer, and the Nature writing stack expanded — made expression-skill the explicit conclusion-first discipline for reporting, planning, file operations, and multi-step technical work; reintroduced planning-with-files as the default on-disk planning and progress-tracking workflow for complex tasks; introduced nature-writing for section drafting and argument construction; refreshed nature-polishing to the latest upstream article-pattern release; and kept nature-response plus nature-data in the journal-writing stack.
• 2026-05-13: Evidence-gated research workflow and Sources/Papers routing tightened — added a shared research-contract.md for Evidence Records, claim strength, and Claim Promotion Gates; connected research ideation, Zotero ingestion, literature synthesis, results reporting, writing, and rebuttal workflows to that contract; and clarified that project paper notes live under Sources/Papers before promoted claims move into Knowledge or Writing.
• 2026-04-24: Project-scoped Obsidian KB workflow consolidated — rebuilt Obsidian project knowledge management into a vault-first workflow, consolidated the older overlapping memory skills into four focused skills, kept repo-local project binding metadata as a runtime layer, and made project navigation human-first instead of a machine registry dump.
• 2026-04-22: Lean core, pruned default agents, safer install lifecycle, and cleaner paper discovery — replaced large always-on CLAUDE.md / AGENTS.md files with compact core instructions, pruned the default agent set to the retained core agents, added safe install-state based uninstall support, generalized daily-paper-generator to broader topics with arXiv / bioRxiv support and a fixed Top 10 -> Top 3 -> Top 1 selection flow.
• 2026-04-15: pubfig and pubtab introduced — introduced pubfig, a Python package for publication-grade scientific figures, and pubtab, a Python package for publication-ready tables and Excel↔LaTeX workflows. Together they provide a cleaner production stack for paper figures, benchmark tables, export control, and final artifact QA.

Quick Navigation

Claude Scholar is not an end-to-end autonomous research system that tries to replace the researcher.

Its core idea is simple:

human decision-making stays at the center; the assistant accelerates the workflow around it.

That means Claude Scholar is designed to help with the heavy, repetitive, and structure-sensitive parts of research — literature organization, note-taking, experiment analysis, reporting, and writing support — while still keeping the key judgments in human hands:

• which problem is worth pursuing,
• which papers actually matter,
• which hypotheses are worth testing,
• which results are convincing,
• and what should be written, submitted, or abandoned.

In other words, Claude Scholar is a semi-automated research assistant, not a “fully automated scientist.”

Who This Is For

Claude Scholar is especially well-suited to:

• computer science researchers who move between literature review, coding, experiments, and paper writing,
• AI / ML researchers who need one assistant workflow spanning ideation, implementation, analysis, reporting, and rebuttal,
• research engineers and graduate students who want stronger workflow structure without giving up human judgment,
• and software-heavy academic projects that benefit from Zotero, Obsidian, CLI automation, and reproducible project memory.

It can still help in other research settings, but its current workflow design is most aligned with computer science, AI, and adjacent computational research.

Core Workflow

Claude Scholar routes research work through a traceable path: question -> evidence -> experiment -> analysis -> claim -> writing. Each stage should preserve what is known, what is uncertain, and what decision should happen next.

• Ideation: turn a vague topic into concrete questions, research gaps, and an initial plan.
• Literature: search, import, organize, and read papers through Zotero collections.
• Paper notes: convert papers into structured reading notes and reusable claims.
• Knowledge base: route durable knowledge into Obsidian across Sources / Knowledge / Experiments / Results / Results/Reports / Writing / Daily / Maps.
• Experiments: track hypotheses, experiment lines, run history, findings, and next actions.
• Analysis: generate strict statistics, real scientific figures, and analysis artifacts with results-analysis.
• Reporting: produce a complete post-experiment report with results-report, then write it back into Obsidian.
• Writing and publication: carry stable findings into literature reviews, papers, rebuttals, slides, posters, and promotion.

Quick Start

Requirements

• Claude Code
• Git
• (Optional) Python + uv for Python development
• (Optional) Zotero + Galaxy-Dawn/zotero-mcp for literature workflows
• (Optional) Obsidian for project knowledge-base workflows

Option 1: Full Installation (Recommended)

git clone https://github.com/Galaxy-Dawn/claude-scholar.git /tmp/claude-scholar
bash /tmp/claude-scholar/scripts/setup.sh

Windows: please use Git Bash or WSL to run the installer.

The installer is backup-aware and incremental-update friendly:

• updates repo-managed skills/commands/agents/rules/hooks/scripts/CLAUDE*.md,
• backs up overwritten files to ~/.claude/.claude-scholar-backups/<timestamp>/,
• backs up settings.json to settings.json.bak,
• preserves an existing ~/.claude/CLAUDE.md and installs the repo-managed version as ~/.claude/CLAUDE.scholar.md,
• preserves an existing ~/.claude/CLAUDE.zh-CN.md and installs the repo-managed version as ~/.claude/CLAUDE.zh-CN.scholar.md,
• preserves your existing env, model/provider settings, API keys, permissions, and current mcpServers values,
• adds missing hook entries instead of replacing your entire hook set.

Important CLAUDE note: if you already maintain your own ~/.claude/CLAUDE.md or ~/.claude/CLAUDE.zh-CN.md, review ~/.claude/CLAUDE.scholar.md and ~/.claude/CLAUDE.zh-CN.scholar.md after installation and manually merge the Claude Scholar sections you want into your own files. Do not assume the sidecar files are applied automatically.

To update later:

cd /tmp/claude-scholar
git pull --ff-only
bash scripts/setup.sh

To uninstall later:

cd /tmp/claude-scholar
bash scripts/uninstall.sh

The installer now writes:

• ~/.claude/.claude-scholar-manifest.txt for the exact files managed by Claude Scholar
• ~/.claude/.claude-scholar-install-state for install ownership metadata used by safe uninstall

The uninstaller removes only files and settings entries recorded in that install state. It does not guess ownership from the current repo checkout.

Option 2: Minimal Installation

Install only a small research-focused subset:

git clone https://github.com/Galaxy-Dawn/claude-scholar.git /tmp/claude-scholar
mkdir -p ~/.claude/hooks ~/.claude/skills
cp /tmp/claude-scholar/hooks/*.js ~/.claude/hooks/
cp -r /tmp/claude-scholar/skills/ml-paper-writing ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/research-ideation ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/results-analysis ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/results-report ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/review-response ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/writing-anti-ai ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/git-workflow ~/.claude/skills/
cp -r /tmp/claude-scholar/skills/bug-detective ~/.claude/skills/

Post-install: minimal/manual install does not auto-merge settings.json; copy only the hooks or MCP entries you want from settings.json.template. If you already have your own ~/.claude/CLAUDE.md or ~/.claude/CLAUDE.zh-CN.md, also merge the relevant sections from this repo's Claude files into yours instead of blindly overwriting them.

Option 3: Selective Installation

Copy only the parts you need:

git clone https://github.com/Galaxy-Dawn/claude-scholar.git /tmp/claude-scholar
cd /tmp/claude-scholar

cp hooks/*.js ~/.claude/hooks/
cp -r skills/latex-conference-template-organizer ~/.claude/skills/
cp -r skills/architecture-design ~/.claude/skills/
cp agents/paper-miner.md ~/.claude/agents/
cp rules/coding-style.md ~/.claude/rules/
cp rules/agents.md ~/.claude/rules/

Post-install: selective/manual install does not auto-merge settings.json; copy only the hooks or MCP entries you actually want from settings.json.template. If you already have your own ~/.claude/CLAUDE.md or ~/.claude/CLAUDE.zh-CN.md, merge the relevant sections from this repo's Claude files into yours instead of blindly overwriting them.

Option 4: Plugin Marketplace Installation

Step 1: Install the Plugin

/plugin marketplace add Galaxy-Dawn/claude-scholar
/plugin install claude-scholar@claude-scholar

This auto-loads all skills, commands, agents, and hooks. During installation, you can choose the scope: user (all projects) or project (single project).

Step 2: Install Rules (Required)

Claude Code plugins cannot distribute rules automatically. Install them manually:

git clone https://github.com/Galaxy-Dawn/claude-scholar.git /tmp/claude-scholar

# User-level (all projects)
mkdir -p ~/.claude/rules
cp /tmp/claude-scholar/rules/*.md ~/.claude/rules/

# Or project-level (current project only)
mkdir -p .claude/rules
cp /tmp/claude-scholar/rules/*.md .claude/rules/

Post-install: plugin installation does not auto-load CLAUDE.md or configure settings.json; if you already have your own ~/.claude/CLAUDE.md or ~/.claude/CLAUDE.zh-CN.md, merge the relevant Claude Scholar sections into yours instead of assuming the plugin applies them automatically. If you need Zotero MCP or other integrations, see the Integrations section for manual setup.

Getting Started Scenarios

After installation, the simplest way to begin is to describe your task in natural language. You do not need to memorize the whole system first. Below are a few realistic starting points.

1. Start a New Research Topic

You can say:

Help me start research on [your topic]. I want a literature-grounded plan, the key open questions, and the next concrete steps.

What Claude Scholar will typically help with:

• clarify the topic and narrow the research question,
• identify promising literature directions,
• suggest an initial plan or hypothesis list,
• optionally route the work into Zotero or Obsidian if you use them.

2. Review a Zotero Collection

You can say:

Review my Zotero collection on brain foundation models and summarize the main directions, gaps, and promising next steps.

Typical output:

• paper grouping by theme,
• a short literature synthesis,
• gap analysis,
• candidate research directions worth pursuing next.

3. Analyze Finished Experiment Results

You can say:

Analyze the results in this experiment folder, check what changed across runs, and write a decision-oriented summary.

Typical output:

• metric comparison,
• ablation or error-analysis suggestions,
• a result summary that highlights what is solid, what is weak, and what to run next.

4. Draft a Paper or Rebuttal Section

You can say:

Help me draft the related work section for this project based on the current findings and paper notes.

or:

Help me write a rebuttal draft for these reviewer comments.

Typical output:

• a structured section draft,
• improved argument flow,
• clearer claims and evidence mapping,
• follow-up points that still need support or verification.

Practical Notes

• Start with one concrete task, not a vague request for "everything."
• If you already maintain your own local CLAUDE.md files, merge the Claude Scholar sections you want into them instead of assuming sidecar files apply automatically.
• Zotero and Obsidian are optional, but they become much more useful when you want durable literature notes or project memory rather than one-off chat output.

Platform Support

Claude Scholar is maintained for:

• Claude Code — the primary installation target.
• Codex CLI — supported workflow and documentation are available in this repo ecosystem.
• OpenCode — supported as an alternative CLI workflow.

The top-level workflow is the same: research, coding, experiments, reporting, and project knowledge management.

Integrations

Zotero

Use Zotero when you want Claude Scholar to help with:

• paper import via DOI / arXiv / URL,
• collection-based reading workflows,
• full-text access through Zotero MCP,
• detailed paper notes and literature synthesis.

See MCP_SETUP.md.

Obsidian

Use Obsidian when you want Claude Scholar to maintain a filesystem-first research knowledge base:

• Sources/
• Knowledge/
• Experiments/
• Results/
• Results/Reports/
• Writing/
• Daily/
• Maps/

See OBSIDIAN_SETUP.md.

Primary Workflows

Complete academic research lifecycle — 7 stages from idea to publication.

1. Research Ideation (Zotero-Integrated)

End-to-end research startup from idea generation to literature management.

How it works

• 5W1H Brainstorming: turn a vague topic into structured questions (What / Why / Who / When / Where / How).
• Literature Search & Import: search papers, extract DOI/arXiv/URLs, import them into Zotero, and organize them into themed collections.
• PDF & Full Text: attach PDFs when available, read full text when possible, and fall back to abstract-level analysis when necessary.
• Gap Analysis: identify literature, methodological, application, interdisciplinary, or temporal gaps.
• Research Question & Planning: convert the review into concrete questions, initial hypotheses, and next-step planning.
• Evidence Gate: keep weak sources, project hypotheses, and missing evidence explicit before promoting a claim into Knowledge, Writing, or a proposal.

Typical output

• research question cards with hypotheses, evidence needs, falsification criteria, and next actions
• literature review notes
• structured Zotero collection
• project proposal only when the selected question has enough verified evidence; otherwise a research direction / intake draft

2. ML Project Development

Maintainable ML project structure for experiment code and iteration.

How it works

• Structure: use Factory / Registry patterns for new ML components when appropriate.
• Code Quality: keep files maintainable, typed, and config-driven.
• Debugging: inspect stack traces, shell failures, and code-path issues systematically.
• Git Discipline: use branch hygiene, conventional commits, and safer merge/rebase workflows.

3. Experiment Analysis

Strict analysis of experimental results with scientific figures and report-ready artifacts.

How it works

• Data Processing: read experiment logs, metrics files, and result directories.
• Blocker-First Gate: lock unit of analysis, primary metric, seeds/folds/runs, provenance, and comparison family before producing claims.
• Statistical Testing: run strict statistical checks such as t-test / ANOVA / Wilcoxon where appropriate.
• Visualization: generate real scientific figures with interpretation guidance, not just vague plotting suggestions.
• Ablation & Comparison: analyze component contribution, performance tradeoffs, and stability.
• Post-Experiment Reporting: turn the analysis bundle into a full retrospective report with conclusions, limitations, and next actions.

Typical output

• analysis-report.md
• stats-appendix.md
• figure-catalog.md
• figures/
• post-experiment summary report in Obsidian Results/Reports/
• blocker summary / audit note when evidence is incomplete

4. Paper Writing

Systematic academic writing from structure setup to draft refinement.

How it works

• Template Preparation: clean conference templates into an Overleaf-ready structure.
• Journal-Style Polishing: tighten paragraph logic, hedging, and section moves for Nature-leaning prose when needed.
• Reviewer Response: structure major/minor revision comments into an auditable point-by-point response package.
• Data Availability: prepare Nature-ready repository plans, dataset citations, and availability statements.
• Citation Verification: verify references, metadata, and claim-citation alignment.
• Systematic Writing: draft sections from repo context, experiment evidence, and literature notes, while keeping unsupported claims marked instead of polished.
• Claim Ledger: every contribution, result, and contrast should trace to evidence or remain explicitly speculative.
• Style Refinement: reduce robotic phrasing and improve rhythm, clarity, and tone.

5. Paper Self-Review

Quality assurance before submission.

How it works

• Structure Check: logical flow, section balance, and narrative coherence.
• Logic Validation: claim-evidence alignment, assumption clarity, and argument consistency.
• Claim Audit: verify that main claims are supported by evidence, weaken over-strong language, and preserve uncertainty when needed.
• Citation Audit: reference correctness and completeness.
• Figure Quality: caption completeness, readability, and accessibility.
• Compliance: page limits, formatting, and disclosure requirements.

6. Submission & Rebuttal

Submission preparation and review response workflow.

How it works

• Pre-submission Checks: venue-specific formatting, anonymization, and checklist requirements.
• Review Analysis: classify reviewer comments into actionable categories.
• Response Strategy: decide whether to accept, defend, clarify, or propose new experiments.
• Rebuttal Writing: generate structured responses with professional tone, evidence anchors, and explicit unresolved items.

7. Post-Acceptance Processing

Conference preparation and research promotion after acceptance.

How it works

• Presentation: prepare talk structure and slide guidance.
• Poster: organize content into poster-ready layout and hierarchy.
• Promotion: generate social media, blog, or summary material for broader communication.

Supporting Workflows

These workflows run in the background to strengthen the primary workflows.

Obsidian Project Knowledge Base

Use Obsidian as the project-scoped durable knowledge surface, not just as a note dump.

How it works

• bind an existing repo to an Obsidian vault,
• route stable knowledge into Sources / Knowledge / Experiments / Results / Results/Reports / Writing / Daily / Maps,
• keep Daily/ and repo-local binding metadata updated conservatively,
• ingest new source material into the correct canonical destination,
• keep abstract-only and webpage-placeholder sources from supporting durable claims,
• only generate extra Bases or canvases on explicit request.
• use /kb-sync for deterministic resyncs and /kb-links for standalone link repair.

Note language configuration

Generated and synced Obsidian notes resolve their language with this priority:

1. project config: .claude/project-memory/registry.yaml -> note_language
2. environment variable: OBSIDIAN_NOTE_LANGUAGE
3. default: en

Note: the file is currently named registry.yaml for historical reasons, but its on-disk format is JSON.

Per-project example:

{
  "projects": {
    "my-project": {
      "project_id": "my-project",
      "vault_root": "/path/to/vault/Research/my-project",
      "note_language": "zh-CN"
    }
  }
}

English and Chinese section headings remain mutually compatible during sync, so older notes in either language can still be updated safely after switching configuration.

Automated Enforcement Workflow

Cross-platform hooks automate routine workflow checks and reminders.

Hooks

• skill-forced-eval.js
• session-start.js
• session-summary.js
• stop-summary.js
• security-guard.js

How it works

• Before prompts: evaluate applicable skills and surface relevant workflow hints.
• At session start: show Git state, available commands, and project-memory context.
• At session end/stop: summarize work and remind the user about minimum maintenance tasks.
• Security: block catastrophic commands and require confirmation for dangerous but legitimate ones.

Communication and Reporting Discipline

Use a reusable communication layer when the task needs conclusion-first reporting, concrete evidence, visible risk, or compact next-step guidance.

How it works

• lead with the conclusion instead of narration,
• prefer commands, paths, counts, checks, and observable behavior over abstract process language,
• ask clarifying questions only when ambiguity changes the outcome,
• surface risk, uncertainty, and destructive boundaries early,
• keep long-running work visible with step / checkpoint style roadmarks,
• persist multi-step work to disk with task_plan.md and notes.md instead of relying only on transient context.

Knowledge Extraction Workflow

Specialized agents can mine reusable knowledge from papers and competitions.

How it works

• extract writing patterns, venue expectations, and rebuttal strategies from papers,
• extract engineering patterns and solution structure from Kaggle workflows,
• feed those insights back into skills and reference material.

Skill Evolution System

Claude Scholar also contains a self-improvement loop for its own skills.

How it works

• create new skills with clear trigger descriptions,
• review them across quality dimensions,
• apply structured improvements and iterate.

Documentation

• MCP_SETUP.md — Zotero/browser MCP setup
• OBSIDIAN_SETUP.md — Obsidian knowledge base workflow
• CLAUDE.md — lightweight Claude Code core instructions
• CLAUDE.zh-CN.md — Chinese companion for the lightweight core instructions
• settings.json.template — optional settings template for hooks/plugins/MCP

Project Rules

Claude Scholar includes project rules for:

• coding style,
• agent orchestration,
• security,
• experiment reproducibility.

These are reflected in the shipped rules and in CLAUDE.md.

Contributing

Issues, PRs, and workflow improvements are welcome.

If you propose changes to installer behavior, Zotero workflows, or Obsidian routing, please include:

• the user scenario,
• the current limitation,
• the expected behavior,
• and any compatibility concerns.

License

MIT License.

Citation

If Claude Scholar helps your research or engineering workflow, you can cite the repository as:

@misc{claude_scholar_2026,
  title        = {Claude Scholar: Semi-automated research assistant for academic research and software development},
  author       = {Gaorui Zhang},
  year         = {2026},
  howpublished = {\url{https://github.com/Galaxy-Dawn/claude-scholar}},
  note         = {GitHub repository}
}

Acknowledgments

Built with Claude Code CLI and enhanced by the open-source community.

References

This project is inspired by and builds upon excellent work from the community:

• everything-claude-code - Comprehensive resource for Claude Code CLI
• AI-research-SKILLs - Research-focused skills and configurations
• expression-skill - Public conclusion-first communication skill reused here for reporting and response discipline
• nature-skills - Nature-oriented writing, polishing, reviewer-response, and data-availability skills reused here with attribution

These projects provided valuable insights and foundations for the research-oriented features in Claude Scholar.

For data science, AI research, and academic writing.

Repository: https://github.com/Galaxy-Dawn/claude-scholar