Forcing structure on AI research assistants

What it does

Claude Scholar is a configuration and prompt framework that wraps AI coding assistants—Claude Code, Codex CLI, Kimi Code CLI, and OpenCode—into a semi-automated research workflow. It routes work through a rigid path from question to evidence to experiment to analysis to claim to writing, while integrating with Zotero for paper management and Obsidian for project knowledge. The system uses skills, agents, and hooks to enforce structure without pretending to replace the researcher.

The interesting bit

Instead of letting the AI freestyle, it imposes a conclusion-first discipline through an expression-skill and gates claims behind an evidence contract before they can enter the knowledge base or manuscript. It is essentially a layer of research bureaucracy that keeps the assistant from hallucinating its way through your paper.

Key highlights

• Supports four major AI coding CLIs via dedicated branches: Claude Code, Codex CLI, Kimi Code CLI, and OpenCode.
• Integrates Zotero for paper import, full-text reading, and citation export, plus an Obsidian vault workflow for durable project memory.
• Enforces an evidence-gated workflow where paper notes live under Sources/Papers and claims must pass promotion gates before entering Knowledge or Writing.
• Includes specialized publication tooling through integrated pubfig and pubtab packages for figure and table generation.
• Provides hooks for security guards, session summaries, and skill evaluation to prevent the assistant from wandering off-task.

Caveats

• The workflow is designed primarily for computer science and AI researchers; other fields may find the structure mismatched.
• Because it is tightly coupled to specific CLI tools and their branching strategies, switching between assistants or mixing workflows requires careful branch management.
• The project is young enough that its changelog reads like a live refactoring log, with skills and agents being pruned, renamed, and consolidated regularly.

Verdict

Computer science and AI researchers who already live in Claude Code or Codex CLI and want their assistant to stop generating loose prose will find this a useful harness. If you are looking for a fully autonomous scientist or a simple chatbot, this is explicitly not it.