🌐 English · 中文

paper-wiki-starter

A Claude-Code-driven scaffold for turning a pile of research paper PDFs into a hyperlinked Obsidian vault — concept pages, reading cards, cross-paper reports, research ideas, and experiment plans.

This repository contains the framework: workflow rules, prompt templates, Claude skills, and a bootstrap script. It contains zero paper content, zero personal notes, and zero model weights. You bring those.

Note: the paper-write skill (end-to-end manuscript drafting) is not included in this initial release. A similar workflow will be added in a future update. The three shipped skills — paper-report, paper-idea, and paper-experiment — are fully functional standalone.

---

What you get

• 3 Claude Code skills that turn natural-language questions into structured wiki entries:
  • paper-report — cross-paper synthesis, intent-aware (gap-finding / landscape / method-compare / …)
  • paper-idea — research-idea scouting with mandatory novelty check
  • paper-experiment — transferable experiment plans with "steal-able" baselines, datasets, metrics
• Prompt templates (schema/prompts/) covering NotebookLM paper review, Claude concept update, and wikilink conventions
• Bootstrap scripts that install the surrounding CLI stack (MinerU, qmd, NotebookLM CLI, endnote-cli, kepano's obsidian-skills bundle) with a single command
• Deterministic audit helpers (scripts/concept-audit.py) for periodic wiki health checks

Architecture at a glance

raw/ (PDFs + MinerU output)
    │
    ▼
CLI tools: mineru → qmd → obsidian-cli → notebooklm → endnote-cli
    │
    ▼
Claude skills: paper-report ← paper-idea, paper-experiment
    │
    ▼
wiki/ (concepts, papers, reports, ideas, experiments)

Full picture: architecture.md · Workflow rules: CLAUDE.md

Quick start

On a fresh machine with Claude Code installed:

git clone https://github.com/TCMzhoutong/paper-wiki-starter.git
cd paper-wiki-starter
./scripts/bootstrap.sh

Then open Claude Code in the repo directory and follow the prompts. For the full deployment walk-through (including the steps that can't be automated — NotebookLM OAuth, Obsidian install, EndNote), see BOOTSTRAP.md.

Prerequisites

Required: git, conda, node ≥ 22, npm, python ≥ 3.10, Claude Code.

Installed automatically by the bootstrap: MinerU, qmd, notebooklm-py, endnote-cli, kepano/obsidian-skills bundle.

Manual install (bootstrap tells you when): Obsidian desktop app, NotebookLM OAuth login, EndNote (optional).

Full dependency matrix with upstream URLs and licenses: DEPENDENCIES.md.

Directory layout

paper-wiki-starter/
├── CLAUDE.md                  ← session-wide rules for Claude Code
├── architecture.md            ← how the pieces fit
├── BOOTSTRAP.md               ← Claude-Code-facing deploy prompt
├── DEPENDENCIES.md            ← third-party inventory + licenses
├── skills.lock.yaml           ← machine-readable dep spec
├── schema/prompts/            ← templates (paper cards, concept updates, NotebookLM prompts)
├── scripts/
│   ├── bootstrap.sh           ← one-shot deploy
│   ├── install-skills.sh      ← Claude skills + CLI wrappers
│   ├── check-env.sh           ← status report
│   ├── batch-read.sh          ← headless bulk paper-card generation
│   └── concept-audit.py       ← wiki health check
├── skills/                    ← first-party Claude skills (shipped)
│   ├── paper-report/
│   ├── paper-idea/
│   └── paper-experiment/
├── setup.sh                   ← environment installer (conda + MinerU + qmd + Ollama)
├── add-paper.sh               ← PDF → Markdown → qmd-indexed workflow
├── wiki/                      ← YOUR curated notes (gitignored)
└── raw/                       ← YOUR PDFs + MinerU output (gitignored)

Canonical workflows

Workflow	Trigger phrase in Claude	Output
Add paper	Drop PDF, run ./add-paper.sh	wiki/papers/<title>.md
Research report	"write a report on X", "compare A vs B"	wiki/reports/<topic>_<date>.md
Idea scouting	"find ideas in X", "/idea X"	wiki/ideas/<slug>_<date>.md
Experiment plan	"baseline for X", "/experiment X"	wiki/experiments/<slug>_<date>.md

Details for each workflow — intent detection, block assembly, validation gates — live inside the respective skill's SKILL.md.

Bulk paper reading

For ingesting tens or hundreds of papers at once without bloating a single Claude Code session's context:

./add-paper.sh                 # MinerU-convert all PDFs in raw/pdf/
./scripts/batch-read.sh        # headless loop: one `claude -p` per paper

Run this from a regular terminal (PowerShell / Git Bash), not from inside an interactive Claude Code session. Each paper runs in its own fresh Claude context — no carry-over, no bloat. The script is idempotent (skips papers whose cards already exist) and safe to Ctrl-C and resume. Progress is streamed to batch-read.log.

After the batch finishes, ask Claude Code once to run the 校验维护流程 (concept health audit). This dedupes concepts and resolves alias conflicts across the whole batch in a single pass — the designed-in "cleanup stage" that keeps bulk mode fast.

License

• Code (.sh, .py, .yaml, .json) — MIT
• Content (prompt templates, skill instructions, documentation prose) — CC-BY-NC-4.0

Commercial use of the prompt templates requires separate licensing. Non-commercial academic, personal, and research use is explicitly permitted.

Contributing

See CONTRIBUTING.md. Short version: framework improvements welcome, personal notes aren't. Don't PR PDFs.

Acknowledgements

Stands on the shoulders of:

• MinerU — PDF → Markdown conversion
• MinerU-Document-Explorer (qmd) — local markdown index + semantic search
• notebooklm-py — Google NotebookLM automation
• kepano/obsidian-skills — obsidian-cli, defuddle, and friends
• endnote-cli — EndNote .enl library access
• Obsidian — the vault browser the whole thing orbits around