paper-to-insight

name	paper-to-insight
description	Extract structured insights from a research paper and write them into the research knowledge layer. Use when the user provides a paper (PDF, DOI, arXiv link, or title) and wants it archived as reusable knowledge.

Paper → Insight

Read a research paper and extract structured insights into the research knowledge layer (D:\Playground\research-knowledge\). Every claim must be source-grounded.

Context Budget

MUST load:

• The target paper (PDF/abstract)
• research-knowledge/INDEX.md (to identify target line)
• research-knowledge/lines/<line>/STATE.md — hypotheses summary section only
• research-knowledge/lines/<line>/hypotheses.md — only hypotheses relevant to the paper's topic

MUST NOT load:

• Full experiment data files
• Other papers in the line's papers/ directory (unless cross-referencing)
• methodology files (unless the paper is methods-focused)
• Other research lines' STATE files

Rule: If context exceeds ~500 lines of input, stop loading and proceed with what's loaded.

When to Use

• User gives you a paper PDF, DOI, arXiv link, or title+author
• User says "读这篇", "归档这篇", "提取这篇"
• After a literature search session when papers need to be archived
• When you encounter a paper during research and want to preserve the key findings

Workflow

1. Identify the research line

Ask or determine which research line(s) the paper belongs to:

• lsco-single-layer
• superconducting-diode
• sc-magnetic
• topological-order

A paper can cross lines — note cross-line in the frontmatter.

2. Read the paper

• If given a PDF path, read it directly
• If given a DOI or arXiv ID, search for it and fetch the abstract + key content
• If the paper cannot be accessed, note this and extract what you can from the metadata

3. Load current state

Read the target line's current files before writing:

1. research-knowledge/lines/<line>/STATE.md
2. research-knowledge/lines/<line>/hypotheses.md
3. research-knowledge/lines/<line>/open-questions.md

This ensures you know what's already established and can connect the new paper to existing knowledge.

4. Extract structured insight

Write to research-knowledge/lines/<line>/papers/<slug>.md using this template:

---
type: paper-insight
status: extracted
line: <primary-line>
confidence: high | medium | low
cross-line: [<other-lines>]  # optional
source_grounding:
  doi: <DOI>
  venue: <Journal Volume, Page (Year)>
  authors: <First Author et al.>
  key_figure: <most important figure>
  key_equation: <most important equation>  # optional
  arxiv: <arXiv link>  # if applicable
  publication_status: published | preprint
---

# <Title>

## 一句话结论
- <1 sentence that captures the core finding — this is what the agent reads first>

## 关键发现
- <Finding 1>
- <Finding 2>
- <Finding 3>

## 与我们的关系
- **涉及线**: <line>
- **支撑/挑战假说**: <which hypothesis does this affect?>
- **适用边界**: <what conditions/limits apply?>

## 可行动的下一步
- <concrete action this paper enables>

## 方法标签
- <method 1>
- <method 2>

5. Update knowledge state

After writing the paper insight, update the line's knowledge files:

Update STATE.md:

• Add new established facts (only if confidence is high and from published work)
• Add to "最近更新"
• Update "下一步" if the paper enables new actions

Update hypotheses.md:

• If the paper supports or challenges an existing hypothesis, add evidence
• If the paper suggests a new hypothesis, create it

Update open-questions.md:

• If the paper answers an existing open question, mark it resolved
• If the paper raises new questions, add them

Update cross-cutting/connections.md:

• If the paper connects multiple research lines, add the connection

6. Report to user

Summarize:

• What paper was processed
• Which line(s) it was written to
• Which hypotheses were affected
• Any new next actions discovered

Source Grounding Rules

• Every factual claim must link to a source: DOI, specific figure, equation number, or experiment sample ID
• Distinguish established facts (confidence: high, from published work) from tentative findings (confidence: medium/low, from preprints or single studies)
• Preprints must be labeled publication_status: preprint and confidence capped at medium
• In-house experimental data are valid sources — cite sample ID (e.g., #156 XRD)