Agent 原生文献管理库 Skill

---
name: agent-native-paper-library
description: Build and maintain a literature library designed for coding agents. Use when the user wants to ingest papers, compare papers, write related work, or keep research context queryable across a local workspace.
---

# Agent Native Paper Library

## Goal
Turn a folder of papers into an agent-readable research workspace where every paper has clean source text, a compact note, stable metadata, and explicit instructions for how agents should read, compare, and write from the library.

## When To Use
- You are starting a new research topic from an empty folder.
- You need cross-paper comparison, research trend mapping, or related work drafting.
- You want the agent to read only the information needed for the current question.
- You want your own reading criteria, annotations, and experimental concerns to stay visible to the agent.

## Folder Contract
```text
paper-library/
  AGENTS.md                 # how agents must behave in this library
  papers.md                 # top-level index: title, id, tags, 3-line contribution
  papers/
    1512.03385-resnet/
      source/               # arXiv LaTeX, Markdown, or parsed text
      paper.pdf             # optional, for human reading
      notes.md              # agent/human note for this paper
      metadata.json         # id, title, authors, venue, year, links, tags
  prompts/
    ingest-paper.md
    compare-papers.md
    related-work.md
```

## Core Principles
1. Prefer LaTeX or Markdown over PDF. PDF is for humans; source text is for agents.
2. Use progressive disclosure. The agent reads `papers.md` first, `notes.md` second, and source sections only when needed.
3. Keep every paper structurally isomorphic. Same fields, same headings, same metadata keys.
4. Put your thinking in `notes.md`, not only in PDF highlights.
5. For many-paper jobs, use one subagent per paper, then merge summaries in a clean parent context.
6. Related work must cite from local `.bib`/source paths, not model memory.

## Ingest Workflow
1. Resolve the paper id or URL.
2. Download arXiv source when available; otherwise parse PDF into Markdown/LaTeX with MinerU or another OCR/parser.
3. Create the paper folder using a stable id and short slug.
4. Create `metadata.json`.
5. Ask the agent to read source text and write `notes.md` using the library note schema.
6. Update `papers.md` with tags, location, and a three-line contribution card.
7. Verify that links, paths, and citations resolve locally.

## `notes.md` Schema
```markdown
# <Paper Title>

## Problem
What problem does the paper solve, and why did it matter at publication time?

## Method
What is the core method, architecture, dataset, loss, or experimental design?

## Results
What changed compared with prior work? Include metrics only when they affect decisions.

## Useful For
When should this paper be retrieved for future research questions?

## My Notes
Personal judgment, objections, follow-up ideas, and connections to current projects.

## Evidence Pointers
- source/<file>.tex:<section or line if available>
- paper.bib:<citation key>
```

## `papers.md` Entry Schema
```markdown
## <citation-key> — <title>
- Path: `papers/<id-slug>/`
- Tags: <topic>, <method>, <dataset>, <task>
- One-liner: <what it does>
- Solves: <problem>
- How: <method>
- Result: <effect or limitation>
```

## Agent Rules
- Start every task by reading `AGENTS.md` and `papers.md`.
- Do not open source files for every paper unless the task requires it.
- When comparing papers, first build a shortlist from `papers.md`, then open relevant `notes.md`, then open source sections.
- For more than five papers, spawn subagents with isolated context and require each to return the same schema.
- Never invent citations. Use local citation keys or explicit source paths.
- If parsing quality is uncertain, state the uncertainty and point to the original PDF.

## Verification
- `papers.md` mentions every folder under `papers/`.
- Every `notes.md` has Problem, Method, Results, Useful For, My Notes, Evidence Pointers.
- Every related-work draft includes citation keys that exist in local `.bib` or metadata.
- Large comparison tasks do not paste all papers into one context.