AI Autopilot for PaperBell

[EdgerHao]

Motivation

PaperBell already provides a well-structured academic management workflow — literature reading, project management, note writing, and document export. However, these workflows still require manual driving by the user. Introducing an AI intelligence layer that understands PaperBell's vault structure and rules would enable an upgrade from "manual driving" to "AI autopilot."

I have built an AI academic assistant system in my own Obsidian vault using Claudian (an Obsidian integration client for Claude Code), which has been running stably for months. The architecture is highly compatible with PaperBell, and I'm proposing it here for community discussion.

Architecture: Three-Layer AI Autopilot System

┌─────────────────────────────────────────────┐
│  Layer 3: Skill (On-demand workflow modules) │
│  - Literature review, manuscript triage,     │
│    export checks, etc.                       │
│  - User-extensible                           │
├─────────────────────────────────────────────┤
│  Layer 2: MCP (External tool connections)    │
│  - Zotero MCP → Literature search & fulltext │
│  - AI4Scholar MCP → Academic search engines  │
│  - Obsidian CLI → Vault operations           │
├─────────────────────────────────────────────┤
│  Layer 1: CLAUDE.md + AGENTS.md (Rule layer) │
│  - AI role definition, output format,         │
│    boundary constraints                       │
│  - Domain terminology, citation norms,        │
│    academic writing rules                     │
└─────────────────────────────────────────────┘

Layer 1 — Rule Layer: Define AI assistant behavior through CLAUDE.md and AGENTS.md. Zero-code, configuration-driven — users simply install Claudian or Claude Code CLI to get an AI assistant experience. This is the lightest integration approach and the recommended first step.

Layer 2 — Tool Layer: Connect external tools like Zotero and academic search engines via MCP (Model Context Protocol), enabling the AI to search literature, read PDFs, and manipulate notes directly. PaperBell already has a Zotero integration foundation that MCP can further enhance.

Layer 3 — Skill Layer: Create on-demand workflow modules (Skills) under .claude/skills/, each defined by a SKILL.md specifying trigger conditions and execution logic. Users can enable or customize Skills as needed.

Multi-Model Collaboration

Different models can serve different roles, with task routing defined in rule files:

Model	Role	Use Cases
Claude (Opus)	Pilot	Complex academic writing, logic diagnosis
Codex / Claude (Sonnet)	Co-pilot	Format fixes, batch operations
OpenCode	Cruiser	Code generation, automation scripts

Suggested Implementation Path

1. Phase 1: Add CLAUDE.md + AGENTS.md to the PaperBell repository, defining basic AI assistant rules (zero-code, ready to use out of the box)
2. Phase 2: Add on-demand workflow modules (Skills)
3. Phase 3: Integrate MCP Servers (Zotero MCP first)
4. Phase 4: Write user documentation and quick-start guide

Alignment with PaperBell

• PaperBell's philosophy of "Obsidian open ecosystem" aligns naturally with the configuration-driven AI assistant approach
• PaperBell already has Zotero integration and Pandoc export — the AI layer can directly enhance these existing capabilities
• Users can customize terminology and rules by domain, consistent with PaperBell's customizability

Looking forward to community discussion and feedback!