An AI-native toolkit for academic research — a collection of composable agentic skills covering the full paper workflow from ingestion, reading, translation, summarization to peer review.
In academic research, acquiring papers, deep reading, managing knowledge, and accumulating insights are high-frequency, repetitive tasks. Traditional workflows rely on manual operations and fragmented tools, making it hard to form a continuous knowledge pipeline. This project packages these steps into composable agentic skills driven by Claude Code, so researchers can focus on thinking itself.
PDF / URL
│
▼
paper-ingestion ──→ full_text.md + assets/
│
├──→ summary ──→ notes.md (structured summary + keyword backfill)
└──→ paper-translate ──→ full_text_ch.md (Chinese translation)
| Skill | Purpose | Engine / Backend |
|---|---|---|
| paper-ingestion | PDF → Markdown with image extraction and YAML frontmatter | GLM-OCR (cloud) / MinerU (GPU) / Docling |
| paper-translate | Translate paper markdown, preserving LaTeX, code, and image refs | DeepSeek / TensorBlock |
| paper-summary | Structured summary + keyword extraction + tag backfill | Claude (built-in skill) |
| paper-validator | Peer-review style critique targeting top-tier conferences | Claude (built-in skill) |
| review-research-plan | Critical review & iterative polishing of research plans/roadmaps | Claude (built-in skill) |
| subtitle-translate | SRT subtitle translation, preserving timestamps | DeepSeek / TensorBlock |
| claude-sync | Cross-device ~/.claude sync with interactive conflict resolution | claude-sync CLI (built-in skill) |
| atom-commit | Atomic git commits with professional Conventional Commits messages | Claude (built-in skill) |
| afk | Autonomous task execution — complete all planned work unattended | Claude (built-in skill) |
DeepSeek is the recommended backend for translation. For OCR, the default is Zhipu GLM-OCR cloud API. Both are chosen for their extremely low cost (processing 100 PDFs costs less than $1), strong accuracy on academic text and technical terminology, and low latency for concurrent long-document processing.
cp .env.template .env
# Edit .env and fill in your keys.env supports two-level loading: root-level takes priority, with per-skill .env files as fallback. If you only use a few skills, you can configure keys in the corresponding subdirectory only.
cd paper-ingestion
uv run scripts/ingest_paper.py "https://arxiv.org/pdf/2401.12345.pdf"Invoke the corresponding skill directly in Claude Code:
/summary path/to/paper_folder
/paper-translate path/to/paper_folder/full_text.md
/paper-validator path/to/paper.tex
- Python >= 3.10
- uv — each skill self-manages dependencies via
uv run - Claude Code — drives the agentic skills
Each skill has independent dependencies. Install only what you need — see SKILL.md in each subdirectory for details.
research-utils/
├── .env.template # Root-level API key template (shared across all skills)
├── paper-ingestion/ # PDF → Markdown conversion
│ ├── scripts/
│ ├── mineru-fork/ # Local MinerU fork (optional)
│ └── SKILL.md
├── paper-translate/ # Paper translation
│ ├── scripts/
│ └── SKILL.md
├── paper-summary/ # Structured summarization
│ └── SKILL.md
├── paper-validator/ # Paper review
│ └── SKILL.md
├── review-research-plan/ # Research plan review & iteration
│ ├── commands/ # quick-review, full-review, re-review
│ ├── references/ # adversarial, novelty, feasibility, report-format
│ └── SKILL.md
├── subtitle-translate/ # Subtitle translation
│ ├── scripts/
│ └── SKILL.md
├── claude-sync/ # Cross-device sync with conflict resolution
│ └── SKILL.md
├── atom-commit/ # Atomic git commits with professional messages
│ └── SKILL.md
└── afk/ # Autonomous task execution until all work is done
└── SKILL.md