huangkiki/dailypaper-skills
Agent skills for daily paper discovery, critical reviews, and connected research notes. Supports English and Chinese.
dailypaper‑skills – AI‑powered daily paper assistant
What it does
- Connects a code‑capable LLM agent (Claude, Codex, Gemini, etc.) to a workflow that finds, filters, rates, and summarises the latest AI/robotics papers.
- Each day the agent can produce a short list (≤10) of papers that match your research interests, with a concise “sharp review” that explains what the paper does, why it matters to you, and what questions remain.
- For the papers marked 必读 (must‑read) the system automatically generates a structured Markdown note containing:
- problem statement, method, key formulas, figures, experimental results
- critical commentary and links to related concepts (
[[concept]]style)
- The notes are stored locally (e.g., in an Obsidian vault) and are fully cross‑linked, turning your reading list into a growing knowledge graph.
How it works
- Fetching – pulls new titles from HuggingFace Daily/Trending and arXiv.
- Filtering & scoring – uses the Jev model (a cheap, fast LLM) to evaluate relevance to your configured topics.
- Prioritisation – classifies each paper as 🔥 必读, 👀 值得看, or 💤 可跳过, with evidence‑based reasoning.
- Note generation – the host agent (Claude, Codex, etc.) reads the full PDF (via Poppler) and creates a markdown note following the provided template.
- Linking – concepts extracted from the note are turned into bidirectional links, building a personal concept library.
Key components (called Skills)
| Skill | Trigger phrase | Role |
|---|---|---|
daily-papers |
今日论文推荐 |
End‑to‑end daily recommendation + note creation |
paper-reader |
读一下这篇论文 <url> |
Reads a specific PDF/arXiv link, produces a focused note |
github-trending |
GitHub 周榜 |
Finds relevant open‑source projects (simulators, RL infra, etc.) |
generate-mocs |
更新索引 |
Re‑generates the concept‑link index |
Supporting skills (daily-papers-fetch, daily-papers-review, daily-papers-notes) |
– | Low‑level steps for fetching, reviewing, and note‑writing |
Installation & use
- Clone the repo, run
python3 install.py --agent <your‑agent>(supports Claude, Codex, Cursor, Copilot, Gemini, OpenCode, OpenClaw). - Initialise a config with
python3 skills/_shared/user_config.py --initand tell the agent your research topics and note directory. - In the agent chat, simply type
今日论文推荐(or other trigger phrases) and the whole pipeline runs, returning paths to the recommendation page and the generated notes.
Why Jev?
- Jev 1.13.0 costs ~0.0006 USD per 10‑paper scoring run and finishes in ~0.5 s, compared with $0.11 and 10 s for GPT‑6 Astra – a >99 % cost reduction while selecting almost the same papers.
Who it’s for
- AI/robotics researchers, graduate students, and engineers who need to stay up‑to‑date without manually sifting through dozens of new arXiv submissions.
- Works with any LLM‑enabled agent that can run Python, access the filesystem, and make HTTP requests.
Outputs
| Output | What you see |
|---|---|
| Daily recommendation page | List of up to 10 papers with sharp reviews, priority tags, and links to notes |
| Full paper note (Markdown) | Method description, formulas, figures, critical analysis, [[concept]] links |
| Concept library & index | A browsable network of related concepts inside your Obsidian vault |
Dependencies
- Python 3.10+, Git, curl
- Poppler (for PDF text/image extraction)
- Optional: Obsidian, Zotero, a web viewer for markdown
License – Apache‑2.0.
All details are taken directly from the repository’s README.
Related
- Project
- Project
- Project
- Project
- Project