study

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: study
description: Use this skill when the user wants to read, study, analyze, or deeply understand a research paper (PDF).
disable-model-invocation: false
allowed-tools: Bash, Write, Edit, Read
---

# Paper Study Workflow

Invoke this skill with a paper PDF path.

**Language Detection**: Detect the user's language from their input and generate ALL materials in that language.
- Example: User says "我们学习一下这篇论文吧" → Generate materials in Chinese
- Example: User says "Let's study this paper" → Generate materials in English

---

# Core Philosophy

Primary Objective:
Facilitate deep conceptual understanding and research-level thinking.

Secondary Objective:
Create a structured, reusable paper knowledge system.

This workflow is not just for summarizing — it builds a learning environment around the paper.

---

# Step 0: Check Dependencies (First Run Only)

```bash
if [ ! -f "${CLAUDE_PLUGIN_ROOT}/.installed" ]; then
  echo "First run - installing dependencies..."
  cd "${CLAUDE_PLUGIN_ROOT}"
  npm install || exit 1

  # Install Python dependencies for image extraction
  python3 -m pip install pymupdf --user 2>/dev/null || pip3 install pymupdf --user 2>/dev/null || echo "Warning: Failed to install pymupdf"

  touch "${CLAUDE_PLUGIN_ROOT}/.installed"
  echo "Dependencies installed!"
fi
```

Recommended:

* Node >= 18
* Python 3 with pip (for image extraction)

---

# Step 1: Download and Parse PDF

Supports multiple input formats:

* **Local path**: `~/Downloads/paper.pdf`
* **Direct PDF URL**: `https://arxiv.org/pdf/1706.03762.pdf`
* **arXiv URL**: `https://arxiv.org/abs/1706.03762`

## Step 1a: Check input type and download if URL

```bash
USER_INPUT="<user-input>"

# Check if input is a URL (starts with http:// or https://)
if [[ "$USER_INPUT" =~ ^https?:// ]]; then
  # Download PDF from URL
  INPUT_PATH=$(node ${CLAUDE_PLUGIN_ROOT}/skills/study/scripts/download-pdf.cjs "$USER_INPUT")
else
  # Use local path directly
  INPUT_PATH="$USER_INPUT"
fi
```

For URLs, the download script will:
* Download PDFs to `/tmp/claude-paper-downloads/`
* Convert arXiv `/abs/` URLs to PDF URLs automatically
* Validate that URLs point to PDF files
* Return the local file path for processing

For local paths, use the path directly without downloading.

## Step 1b: Parse PDF

Extract structured information:

```bash
node ${CLAUDE_PLUGIN_ROOT}/skills/study/scripts/parse-pdf.js "$INPUT_PATH"
```

Output includes:

* title
* authors
* abstract
* full content
* githubLinks
* codeLinks
* tags (generated in Step 2.5)

Save to:

```
~/claude-papers/papers/{paper-slug}/meta.json
```

Copy original PDF:

```bash
cp <pdf-path> ~/claude-papers/papers/{paper-slug}/paper.pdf
```

Fallback:
If structured parsing fails, extract raw text and continue with degraded structure.

---

# Step 2: Assess Paper Before Generating Materials

Before generating any files, evaluate:

1. Difficulty Level

   * Beginner
   * Intermediate
   * Advanced
   * Highly Theoretical

2. Paper Nature

   * Theoretical
   * Architecture-based
   * Empirical-heavy
   * System design
   * Survey

3. Methodological Complexity

   * Simple pipeline
   * Multi-stage training
   * Novel architecture
   * Heavy mathematical derivation

This assessment determines:

* Whether to create method.md
* Whether to create .ipynb
* Explanation depth
* Code demo complexity

---

# Step 2.5: Generate Exactly 2 Semantic Tags (Mandatory)

Before generating files, infer exactly 2 tags from semantic understanding of the paper.

Rules:

* Generate exactly 2 tags, no more and no less
* Tags must be distinct
* Each tag should be short (1-3 words)
* Avoid generic tags: `paper`, `research`, `ai`, `ml`
* Prefer one tag for problem/domain and one for method/core idea

Examples:

* `machine translation`, `self-attention`
* `3d detection`, `bev transformer`
* `protein folding`, `structure prediction`

Persist these 2 tags in both locations:

* `~/claude-papers/papers/{paper-slug}/meta.json` as `tags`
* `~/claude-p