read-paper-online

name

read-paper-online

description

Search, browse, and read academic papers online using HuggingFace Papers API and arxiv2md. TRIGGER when: user asks to search for papers online, find recent papers, look up papers on a topic, search arXiv/HuggingFace for papers, check what's new in a research area, browse HuggingFace daily papers, get detailed info about a specific paper (authors, abstract, keywords, github), read a paper's full text by arXiv ID online, or download a paper as markdown. DO NOT TRIGGER when: user wants to read a local paper folder or paper already downloaded locally (use read-paper-local skill), import papers into Zotero (use zotero-connector skill), or convert a local PDF (use pdf-to-markdown skill).

Online Paper Search

Search for academic papers via the HuggingFace Papers hub using the hf CLI. This skill covers online discovery — finding and previewing papers from the web.

Prerequisites

• huggingface_hub CLI installed (pip install -U huggingface_hub[cli])
• scripts/arxiv2md.sh (bundled with this skill, requires curl + python3)
• Internet connection

1. Searching Papers

hf papers search "<query>" [--limit N] [--format table|json] [-q]

Options:

Flag	Default	Description
--limit N	20	Max number of results
--format	table	Output format (table for display, json for parsing)
-q, --quiet	—	Print only paper IDs (one per line)

Examples:

# Basic search
hf papers search "chain of thought reasoning"

# Search with limited results for quick overview
hf papers search "diffusion models" --limit 5

Presenting Search Results

Always use --format table (the default). Table output is compact (ID, title, upvotes, date — one line per paper) and token-efficient. JSON output contains bloated author/avatar/org metadata that wastes context.

If the user needs the abstract of a specific paper, use hf papers read <ID> to fetch it — don't switch to JSON format for summaries.

After showing results, offer next steps: "Want me to read the full text of any of these?" or "Want me to import any into Zotero?"

---

2. Browsing Daily Papers

hf papers list [--date YYYY-MM-DD] [--sort publishedAt|trending] [--limit N]

Only use when the user explicitly wants to browse HuggingFace daily papers (e.g. "看看 hf papers", "今天有什么新论文", "show me today's papers"). This is a browsing command, not a search — do NOT use it as a search substitute.

Options:

Flag	Default	Description
--date	today	Specific date (YYYY-MM-DD or today)
--week	—	ISO week, e.g. 2025-W12
--month	—	Month, e.g. 2025-03
--sort	—	publishedAt or trending
--submitter	—	Filter by HF username
--limit	50	Max results

Examples:

# Today's papers (default)
hf papers list --limit 10

# Trending papers
hf papers list --sort trending --limit 10

# Papers from a specific date
hf papers list --date 2025-03-20 --limit 10

---

3. Getting Paper Details

hf papers info <PAPER_ID>

Returns rich JSON metadata for a single paper: title, full author list, complete abstract, AI-generated summary, AI keywords, publication date, upvotes, github repo/stars, project page, and discussion info.

Only use when detailed metadata is needed — e.g. the user asks about a paper's authors, keywords, github repo, or wants the full abstract without reading the entire paper. Do NOT use as a routine step; search + read already cover most needs.

Example:

hf papers info 2509.22624

Key fields in the output:

Field	What it provides
title	Full title (not truncated)
summary	Complete abstract
ai_summary	One-sentence AI summary
ai_keywords	Topic keywords list
authors[].name	Author names (ignore avatar/user fields)
github_repo	GitHub URL (if available)
github_stars	Star count (if available)
project_page	Project homepage (if available)
upvotes	Community upvote count

---

4. Reading Paper Full Text

Two tools are available for fetching paper full text. Always try arxiv2md.sh first — it produces higher-quality markdown (with frontmatter, ToC, proper formatting). Fall back to hf papers read only if arxiv2md.sh fails.

Primary: arxiv2md.sh (preferred)

bash scripts/arxiv2md.sh <PAPER_ID> -o /tmp/papers/<ID>.md

Uses the arxiv2md.org API to convert arXiv papers to well-structured markdown with YAML frontmatter, table of contents, and clean formatting. Accepts either a bare arXiv ID (e.g. 2301.07041) or a full URL.

Fallback: hf papers read

hf papers read <PAPER_ID> > /tmp/papers/<ID>.md

If arxiv2md.sh fails (network error, API timeout, unsupported paper format), use hf papers read as a fallback. It fetches the paper's full text from arXiv's HTML rendering — functional but less polished than arxiv2md output.

Reading Strategy for Full Papers

Since paper outputs can be very long, follow this strategy:

1. Save to a temp file so you can read selectively:

   mkdir -p /tmp/papers
   bash scripts/arxiv2md.sh <ID> -o /tmp/papers/<ID>.md
   

2. Read the beginning (title, abstract, introduction) first using the Read tool with limit
3. Jump to specific sections based on user interest using the Read tool with offset
4. For quick overview: read just the first 100-200 lines (abstract + intro), then conclusion (last ~100 lines)

---

Workflow Examples

"Search for papers on topic X"

hf papers search "topic X" --limit 50

Show results in a clean format. Offer to read any paper or import into Zotero.

"看看 hf papers" / "今天有什么新论文"

hf papers list --date <yyyy-mm-dd>

Browse mode — just show the table, let the user pick what interests them.

"这篇论文的作者是谁？有没有代码？"

hf papers info <ID>

Extract authors, github_repo, project_page from the JSON output.

"Search for X and read the most relevant one"

1. Search: hf papers search "X" --limit <number>
2. Pick the most relevant (by title match + upvotes)
3. Read (try arxiv2md first, fallback to hf):

   mkdir -p /tmp/papers
   bash scripts/arxiv2md.sh <best_id> -o /tmp/papers/<best_id>.md \
     || hf papers read <best_id> > /tmp/papers/<best_id>.md
   

4. Summarize key findings

When to Use This Skill

Use when the user says things like:

• "Search for papers on reinforcement learning"
• "Find recent papers about LLM agents"
• "Look up papers on vision transformers"
• "Search arXiv for chain-of-thought"
• "Find papers about reward models, limit to 5"
• "Read paper 2301.07041 online"
• "看看 hf papers" / "今天有什么新论文"
• "这篇论文的作者是谁" / "有没有代码仓库"
• "搜索关于多模态的论文"
• "帮我找一下最近的强化学习论文"
• "在线看一下这篇论文"