arxiv-digest

分类: 调研与分析 | 上传者: zhangzzk | 下载: 0 | 版本: v1.0（最新）

Personalized daily arxiv paper digest with preference learning. Use this skill whenever the user asks about new arxiv papers, wants a daily paper digest, asks "what's new on arxiv", wants paper recommendations, or mentions reading/browsing arxiv. Also trigger when the user says things like "papers today", "new preprints", "arxiv update", or "what should I read". This skill fetches, ranks, and presents papers tailored to the user's research interests, and iteratively learns their preferences over time. --- # Arxiv Daily Digest Fetch, rank, and present arxiv papers using two persistent user files: - `researcher_profile.json` (research identity + network) - `arxiv_preferences.json` (reading interests + feedback history) ## Workflow 1. Load or bootstrap profile/preferences. 2. Fetch papers for requested period via web search. 3. Rank papers with preference + network signals. 4. Present period-scaled digest in 3 tiers. 5. Collect feedback and profile updates from interaction. 6. When user asks to dive into a specific paper, hand off to relevant companion skills. ## Step 1: Profile and Preferences ### Storage

更新日志: Source: GitHub https://github.com/zhangzzk/arxiv-digest

目录结构

当前层级: 根目录

📁 references/
- 📄 preference-schema.md 4.2 KB
- 📄 profile-schema.md 5.7 KB
- 📄 scoring-guide.md 2.7 KB
- 📄 storage-guide.md 5.6 KB
📁 scripts/
- 📄 arxiv_fetch.py 37.3 KB
- 📄 build_profile.py 21.9 KB
- 📄 storage_manager.py 19.6 KB
- 📄 storage_paths.py 5.8 KB
📄 .DS_Store 6.0 KB
📄 README.md 1.1 KB
📄 SKILL.md 13.3 KB

SKILL.md

---
name: arxiv-digest
description: >
  Personalized daily arxiv paper digest with preference learning. Use this skill whenever
  the user asks about new arxiv papers, wants a daily paper digest, asks "what's new on arxiv",
  wants paper recommendations, or mentions reading/browsing arxiv. Also trigger when the user
  says things like "papers today", "new preprints", "arxiv update", or "what should I read".
  This skill fetches, ranks, and presents papers tailored to the user's research interests,
  and iteratively learns their preferences over time.
---

# Arxiv Daily Digest

Fetch, rank, and present arxiv papers using two persistent user files:
- `researcher_profile.json` (research identity + network)
- `arxiv_preferences.json` (reading interests + feedback history)

## Workflow

1. Load or bootstrap profile/preferences.
2. Fetch papers for requested period via web search.
3. Rank papers with preference + network signals.
4. Present period-scaled digest in 3 tiers.
5. Collect feedback and profile updates from interaction.
6. When user asks to dive into a specific paper, hand off to relevant companion skills.

## Step 1: Profile and Preferences

### Storage

Use this storage root resolution order:
1. `--storage-dir /path/to/arxiv-digest`
2. `ARXIV_DIGEST_HOME`
3. `XDG_DATA_HOME/arxiv-digest`
4. `~/.local/share/arxiv-digest` (fallback)

Durability rule (required):
- never store profile/preferences under a skill install directory (for example `~/.codex/skills/...` or `~/.claude/skills/...`), because removing the skill can delete that directory.
- if a configured storage root resolves under a skills directory, scripts will auto-redirect to `~/.local/share/arxiv-digest`.

Expected files:
- `researcher_profile.json`
- `arxiv_preferences.json`
- `user_record.json` (auto-updating file index/summary)

### If profile is missing

Ask:
```
What's your name as it appears on your papers? Optionally affiliation, ORCID, or a few arxiv IDs.
```

Build:
```bash
python3 scripts/build_profile.py \
  --name "Jane M. Doe" \
  --affiliation "University of Somewhere" \
  --categories astro-ph.CO astro-ph.IM
```

Refresh existing profile when user asks, monthly, or after new publications:
```bash
python3 scripts/build_profile.py \
  --update ~/.local/share/arxiv-digest/researcher_profile.json \
  --output ~/.local/share/arxiv-digest/researcher_profile.json
```

### If preferences are missing

Bootstrap from profile:
- `core_interests` from `research_fingerprint.topic_keywords`
- `arxiv_categories` from `publications.primary_categories`
- `favorite_authors` from `network.active_coauthors`
- start `negative_signals` empty

If no profile and no name available, ask for interest keywords and create a basic preferences file.

## Step 2: Fetch Papers

Default fetch mode (required):
- use the platform's `web_search`/browse capability to collect papers.
- do not use `scripts/arxiv_fetch.py` for normal digest generation.
- collect from authoritative pages first: arXiv category `/list/{category}/new`, `/recent`, and paper `abs` pages.

Period handling:
- `today`: target papers announced on the current UTC arXiv cycle.
- `week` / `7d`: last 7 days.
- `month` / `30d`: last 30 days.
- `Nd`: last `N` days.
- `YYYY-MM-DD`, `YYYY-MM`, `YYYY-MM-DD:YYYY-MM-DD`: exact day/month/range.
- `recent`: latest available postings if exact date filtering is unavailable.

Web-search query strategy:
- run focused queries per category and time window instead of one broad query.
- prefer queries that return arXiv-native pages:
  - `site:arxiv.org/list/{category}/new`
  - `site:arxiv.org/list/{category}/recent`
  - `site:arxiv.org abs {topic keyword} {date window}`
- for each candidate paper, resolve and verify:
  - arXiv ID
  - full title
  - authors
  - abstract snippet/source page
  - posting/update date
  - category

Exclusion policy (required):
- exclude cross-listed entries from digest candidates.
- exclude replacement entries (`[replacement for ...]`) from digest candidates.
- keep canonical new submissions only.
- if the same arXiv ID appears multiple times across sources/categories, keep only one canonical record.

Latency and quality guardrail:
- do one search pass, one normalization pass, and one ranking pass.
- deduplicate by arXiv ID before ranking.
- if search returns sparse results, widen categories or move from `/new` to `/recent` before giving up.

Unread-day tracking workflow:
- mark read days after delivering a digest:
```bash
python3 scripts/storage_manager.py mark-read --date 2026-02-10
```
- for "digest of last unread days", compute the unread window first:
```bash
python3 scripts/storage_manager.py unread-range
```
- then use that returned date range as the web-search fetch window.

Collection targets by period:
- `today` or single day: 30-80 papers
- `week` (7 days): 80-180 papers
- `month` (30 days): 180-400 papers

Legacy/local fallback (optional only):
- keep `scripts/arxiv_fetch.py` available for environments where web search is not available.
- only use script mode when explicitly requested by the user.

Fetch integrity checks before ranking:
- record fetch window (`date_from`, `date_to`) and categories used.
- verify non-zero candidate pool for multi-category windows unless user explicitly requested a sparse niche query.
- if any category/source query failed, report it and mark digest as partial.
- verify each selected entry has a valid arXiv ID and source page before ranking output.
- verify no selected entry is marked cross-list or replacement.

## Step 3: Score and Rank

Combine preference and network signals.

Preference signals:
- strong: `core_interests`, `methods_interests`, `positive_signals`, favorite authors
- medium: adjacent topic, major survey relevance
- negative: `negative_signals`, clearly off-focus niche

Network boosts from profile:
- very strong: `network.active_coauthors`
- strong: `network.coauthor_rank`
- mild: `network.second_degree_rank`
- moderate: overlap with `research_fingerprint.topic_keywords`

Co-author policy:
- always surface active co-author papers
- annotate them clearly

Deterministic ranking output:
- build a single ranked list with explicit `score` and `reasons` fields.
- persist ranked data to JSON before writing the human digest.
- never hand-pick Top Picks directly from an unsorted pool.

## Step 4: Present Digest

Use 3 sections in this order:
1. Top Picks
2. Solid Matches
3. Boundary Expanders

### Period-scaled counts

| Period window | Top Picks | Solid Matches | Boundary Expanders | Total shown |
|---|---:|---:|---:|---:|
| `today` / 1 day | 3-8 | 6-12 | 3-5 | 12-25 |
| `week` / ~7 days | 8-14 | 10-16 | 5-8 | 23-38 |
| `month` / ~30 days | 24-36 | 14-24 | 8-12 | 46-72 |
| custom N days | interpolate | interpolate | interpolate | capped |

7-day default:
- 10-12 top picks
- 12-14 solid matches
- 6-8 boundary expanders

Selection policy (required):
- `Top Picks` = first `N_top` items from the ranked list.
- `Solid Matches` = next `N_solid` items from the ranked list.
- `Boundary Expanders` = next `N_boundary` items that have lower direct relevance but clear adjacent value.
- only reorder within a tier for readability if relevance order is preserved.
- if a top-ranked paper is excluded, include an explicit exclusion reason (`duplicate topic`, `superseded`, `outside user scope`, etc.).

Preflight omission check (required):
- compute `top_guard = max(N_top + N_solid, 12)` for daily/weekly windows and scale up for longer windows.
- verify every paper in ranked positions `1..top_guard` appears in `Top Picks` or `Solid Matches`, or has an explicit exclusion reason.
- if this check fails, regenerate the digest instead of replying.

Per paper include:
- unique reference index (global within digest): `#01`, `#02`, ...
- title + arxiv ID
- 1-3 authors (mark recognized co-authors)
- concise 2-4 sentence summary
- why-it-matches tag
- fixed link line: `Link: arXiv`
- for boundary papers: `Extends toward: ...`

Title rendering rule (required):
- always print the full original paper title from metadata; do not truncate, abbreviate, or replace with shorthand.
- preserve title punctuation/capitalization as given by source metadata.
- if line length is long, wrap onto the next line instead of shortening the title.

Summary construction rule (required):
- use only information supported by the paper's original abstract metadata (plus explicit ranking signals).
- write exactly 2-4 sentences; do not output one-line blurbs.
- sentence 1: core problem + method/task from the abstract.
- sentence 2: key result, contribution, or dataset/scope from the abstract.
- sentence 3 (optional): caveat, limitation, or context from abstract wording.
- final sentence: explicit `Why it matches` linkage to ranking reasons (topic/method/author/network), not generic praise.
- avoid vague fillers like "interesting work" or "relevant paper" without concrete abstract-grounded details.

Output style:
- use the exact entry template below for every paper
- most relevant first within each section
- concise summaries unless user requests detail

Standard entry template (required):
```text
#01 Full Paper Title (2501.12345) -- Author A, Author B, Author C
2-4 sentence summary grounded in abstract metadata.
Why it matches: specific overlap with user interests/signals.
Link: arXiv
```

Formatting rules (required):
- first line must be exactly: `#{index} {full title} ({arxiv_id}) -- {1-3 authors}`.
- always include `Why it matches:` as a separate line.
- always include `Link: arXiv` as the final line.
- do not output extra bullets/sub-lines inside one paper entry.

Digest provenance block (required):
- report fetch period, fetch timestamp, categories, total fetched papers, total excluded cross-lists, total excluded replacements, and whether any source queries failed.
- if partial, label the digest as partial at the top.

## Step 5: Collect Feedback

Ask:
1. Which paper IDs were useful? (e.g., `#03`, `#11`)
2. Which topics should be filtered next time?
3. Which authors/groups should be watched?
4. Any profile changes? (new paper IDs, new collaborators, affiliation/ORCID updates)

## Step 6: Deep Dive on Selected Papers

If the user asks to dive into a paper (by digest index like `#03` or by arxiv ID):
1. Resolve digest index -> `arxiv_id`.
2. Load the full paper, not just abstract metadata.
3. Use full-paper content as the basis for detailed explanation and Q&A.
4. Route to the minimal relevant companion skill(s) for the requested deep-dive outcome.

Companion-skill routing:
- `pdf`: full-paper extraction, section-by-section reading, table/figure checks, and claim verification.
- `jupyter-notebook`: reproduction attempts, exploratory experiments, derivations, and executable walkthroughs.
- `doc`: polished `.docx` reading notes, review memos, and formatted writeups.
- `notion-knowledge-capture`: structured notes, decision logs, and linked Notion knowledge capture.

Coordination policy:
- keep `arxiv-digest` as the orchestrator for paper-selection context.
- if multiple skills apply, invoke them in a minimal sequence and state the sequence briefly.
- preserve traceability in downstream outputs by including paper title and arXiv ID.

Recommended source order:
1. arxiv PDF: `https://arxiv.org/pdf/{arxiv_id}.pdf`
2. arxiv abs page for metadata cross-check: `https://arxiv.org/abs/{arxiv_id}`

Deep-dive response should include:
- problem setup and assumptions
- method/algorithm details
- key equations or model components (in words unless user asks for derivation)
- experiment/data setup
- main results and limitations
- practical takeaways and follow-up reading pointers

When uncertain, quote section names/figure numbers from the full paper and clearly distinguish paper claims from assistant interpretation.

## Step 7: Dynamically Update Profile

During each interaction, update `researcher_profile.json` when the user provides durable profile facts:
- new publication arxiv IDs
- new/recent collaborators
- affiliation, homepage, ORCID, or preferred name updates
- explicit request: "update my profile"

Update policy:
1. Lightweight update immediately after interaction:
   - patch known fields in profile JSON directly
   - append newly provided paper IDs if missing
2. Full refresh when needed:
   - run `build_profile.py --update ...` when user reports new publications,
     collaborator network likely changed, or profile is stale.

Refresh command:
```bash
python3 scripts/build_profile.py \
  --update ~/.local/share/arxiv-digest/researcher_profile.json \
  --output ~/.local/share/arxiv-digest/researcher_profile.json
```

Save updated profile:
- `~/.local/share/arxiv-digest/researcher_profile.json` (or resolved storage root)

## Step 8: Update Preferences

Apply feedback:
- liked papers -> add keywords to `positive_signals`
- disliked papers -> add to `negative_signals`
- mentioned authors -> add to `favorite_authors`
- liked expanders -> promote topic toward `core_interests` when appropriate

Save updated preferences:
- `~/.local/share/arxiv-digest/arxiv_preferences.json` (or resolved storage root)

## Edge Cases

- Weekend/holiday: if `/new` is empty, use `/pastweek` or `/recent`.
- Too many papers: enforce period caps; do not dump unbounded lists.
- Too few papers: broaden categories or use `recent`.
- Web-search/network failures: report failure; do not fabricate listings.
- Missing interest data: ask user directly instead of guessing.

## References

- `references/scoring-guide.md`
- `references/storage-guide.md`
- `references/preference-schema.md`
- `references/profile-schema.md`

登录后下载/点赞/收藏 ❤ 2 | ★ 0

arxiv-digest

目录结构

SKILL.md

举报内容

提示