skill-find-papers

---
name: skill-find-papers
description: Find, rank, summarize, and track high-quality research papers with Scholar MCP, accept user-submitted PDFs, DOIs, or links through a persistent local inbox as special-interest signals, then create evidence-rich bilingual Feishu deep-read documents with detailed abstracts, contributions, methods, figures, and verified links using lark-cli. Use for recurring paper recommendations, weekly research digests, receiving papers directly from the user, processing paper feedback, changing research interests, or creating detailed paper notes in Feishu.
---

# Skill Find Papers

Discover five recent high-quality papers plus one classic paper, learn from user feedback, and archive selected deep reads in Feishu.

## Required Tools

- Use the configured `scholar_mcp` MCP server for literature discovery, metadata, citations, PDF retrieval, and paper analysis.
- Use `lark-cli` for Feishu folders and documents.
- Use `scripts/paper_state.py` for persistent preferences and recommendation history.
- Before any Feishu operation, run `lark-cli doctor`. Before choosing Feishu flags, read the version-matched embedded skills with `lark-cli skills read lark-shared`, `lark-cli skills read lark-drive`, and `lark-cli skills read lark-doc`.

## Start Every Run

1. Run:

   ```powershell
   python "$env:USERPROFILE\.codex\skills\skill-find-papers\scripts\paper_state.py" init
   python "$env:USERPROFILE\.codex\skills\skill-find-papers\scripts\paper_state.py" show
   python "$env:USERPROFILE\.codex\skills\skill-find-papers\scripts\paper_state.py" inbox-list --status pending
   ```

2. Inspect `needs_initial_interests` from `init`. On first use, stop before discovery and ask the user to set initial research interests. Defaults are `transformers`, `LLM`, and `agent`.
   - If the user accepts the defaults, run `paper_state.py interests --confirm`.
   - If the user supplies topics, run `paper_state.py interests --replace "<topic1,topic2>"`.
3. Treat saved preferences and feedback as authoritative after onboarding.
4. If the user directly changes interests, apply the change immediately with `interests --add`, `--remove`, or `--replace`, then confirm the resulting list.

## Local Paper Inbox

Treat any paper PDF, DOI, publisher/arXiv link, or explicit paper submission sent by the user as an inbox submission.

1. Add it immediately with `paper_state.py inbox-add`. For a local PDF, pass `--file`; for a remote paper, pass `--url` or `--doi`. Omit unavailable metadata.
2. The script copies local PDFs into the managed inbox at `$HOME/.codex/skill-data/skill-find-papers/inbox/files`. Use that managed path for later Scholar MCP ingestion.
3. Mark every inbox submission as `特别感兴趣（用户主动发送）`. Submission is explicit consent to analyze it and include it in the next deep-read document; do not ask the user to select it again.
4. At the start of every weekly run, process all pending inbox papers before ordinary discovery:
   - prefer the managed local PDF and use Scholar MCP `local_pdf_path`;
   - otherwise resolve the DOI or URL to the fullest lawful source;
   - apply the complete deep-read contract with greater depth than an ordinary recommendation;
   - infer focused topics from the full paper, not only broad parent categories;
   - after successful analysis and document inclusion, run `paper_state.py inbox-process --paper-id "<id>" --topics "<topic1,topic2>" --document-url "<url>"`.