HuggingFace Daily Papers Digest

Track the most impactful AI papers every day — filtered by community upvotes and GitHub stars.

What This Skill Does

When invoked (/hf or asking about today's papers), this skill:

1. Fetches today's HuggingFace daily papers via their JSON API
2. Filters to papers with upvotes >= 20 OR GitHub stars >= 100 (configurable)
3. Generates a concise structured summary for each qualifying paper
4. Delivers via your configured method (stdout, Telegram, or email)

---

Onboarding (First Run)

Check for ~/.hf-papers/config.json with onboardingComplete: true. If absent, guide the user through these steps sequentially:

Step 1 — Introduction

Explain: "This skill monitors HuggingFace's daily paper submissions and surfaces the ones getting real traction — based on community upvotes and GitHub stars. You'll get a daily digest of only the papers worth reading."

Step 2 — Upvote Threshold

Ask: "How many upvotes should a paper have to be included? Default is 20. Papers with fewer upvotes but popular GitHub repos will also be included."
Save to config.upvoteThreshold.

Step 3 — GitHub Stars Threshold

Ask: "What's the minimum GitHub stars for a paper's code repo to qualify? Default is 100."
Save to config.githubStarsThreshold.

Step 4 — Language

Ask: "Would you like the digest in English, Chinese (中文), or bilingual (both, interleaved paragraph by paragraph)?"
Options: en, zh, bilingual. Save to config.language.

Step 5 — Delivery Method

Ask: "How should I deliver the digest?"

• Here in the terminal (stdout) — works right now, no setup needed
• Telegram — sends to a Telegram chat/channel automatically
• Email — sends via Resend API

Save to config.deliveryMethod.

If Telegram: explain how to create a bot via @BotFather, get the token, and find chat ID. Ask for chatId, save to config.telegram.chatId. Tell them to add TELEGRAM_BOT_TOKEN=xxx to ~/.hf-papers/.env.

If Email: ask for email address, save to config.email.address. Tell them to add RESEND_API_KEY=xxx to ~/.hf-papers/.env.

Step 6 — Schedule (optional)

Ask: "Would you like a daily or weekly digest? At what time and timezone?"
Save to config.schedule and config.timezone.

For automated delivery (Telegram/Email), suggest setting up a cron job:

# Add to crontab (crontab -e):
# Daily at 9:00 AM Beijing time (UTC+8 = 1:00 AM UTC):
0 1 * * * cd /path/to/auto-hf-papers && node scripts/fetch-papers.js | claude -p "$(cat prompts/summarize-paper.md)" | node scripts/deliver.js

Step 7 — Mark Complete & Run First Digest

Set config.onboardingComplete = true. Write the complete config to ~/.hf-papers/config.json (create directory if needed).

Immediately run the first digest (see Content Delivery Workflow below) and ask for feedback.

---

Content Delivery Workflow

When the user invokes /hf, asks for "today's papers", or the scheduled task fires:

Step 1 — Load config

Read ~/.hf-papers/config.json. Use defaults if missing (upvotes=20, stars=100, language=en, delivery=stdout).

Step 2 — Fetch and filter papers

Run: node scripts/fetch-papers.js

Optional overrides:

• node scripts/fetch-papers.js --date=2026-03-20 for a specific date
• node scripts/fetch-papers.js --upvotes=10 --stars=50 to override thresholds

The script outputs a JSON blob to stdout containing:

• date — the date fetched
• stats — total papers, filtered count, thresholds used
• papers — array of filtered papers with title, upvotes, githubStars, abstract, authors, links
• prompts — summarizePaper, digestIntro, translate prompt texts
• errors — non-fatal errors (empty array if all went well)

Step 3 — Check for content

If papers array is empty:

• Tell the user: "No papers today exceeded the thresholds (upvotes >= X, stars >= Y). Total papers submitted: N."
• Suggest: "You can lower thresholds with node scripts/fetch-papers.js --upvotes=5"
• Stop here.

Step 4 — Generate the digest

Using the JSON output:

Header: Apply the digestIntro prompt — fill in date, stats (filtered, total, upvoteThreshold, starsThreshold).

Per paper: For each paper in the papers array (sorted by upvotes desc), apply the summarizePaper prompt:

• Include title, upvotes, githubStars, TL;DR from abstract, what they did, why it matters
• Always include hfUrl and arxivUrl links
• Only include githubRepo link if the field is non-null
• Use aiSummary and abstract fields — do NOT fabricate content or visit external URLs
• Keep each paper under 150 words

CRITICAL RULES:

• NEVER invent or fabricate content. Only use what's in the JSON.
• Every paper must have its hfUrl and arxivUrl included.
• Job titles / affiliations come from the authors field only.

Step 5 — Apply language setting

• en → keep English only
• zh → translate full digest to Chinese following translate prompt
• bilingual → interleave English and Chinese paragraph by paragraph following translate prompt

Step 6 — Deliver

Pipe the digest text through node scripts/deliver.js or display inline based on config.deliveryMethod.

---

Configuration Management

All preferences can be changed via conversation at any time:

• "Change language to Chinese" → update config.language = "zh"
• "Lower the upvote threshold to 10" → update config.upvoteThreshold = 10
• "Switch delivery to Telegram" → update config.deliveryMethod = "telegram", collect chatId and token

Config file location: ~/.hf-papers/config.json
API keys location: ~/.hf-papers/.env

Prompt Customization

Users can override any prompt by copying files to ~/.hf-papers/prompts/:

• ~/.hf-papers/prompts/summarize-paper.md — customize paper summary format
• ~/.hf-papers/prompts/digest-intro.md — customize digest header
• ~/.hf-papers/prompts/translate.md — customize translation style

These take priority over the project defaults. Changes take effect on the next run.

---

Data Source

Papers are fetched from the official HuggingFace API:
https://huggingface.co/api/daily_papers?date=YYYY-MM-DD

The API returns upvote counts and GitHub star counts directly — no GitHub API calls needed.
Papers are submitted by the community to https://huggingface.co/papers each day.