Paper Reading

Paper Reading - Research Paper Summarization

Overview

A structured approach to reading and summarizing scientific research papers. Automatically identifies paper type (empirical/theoretical/survey/systems), selects the appropriate template, screenshots important figures, and embeds them in the summary document.

When to Use

• User provides a paper (PDF path, URL, or pasted content) and asks for summary
• User asks to "read", "summarize", or "analyze" a research paper
• User wants to understand a paper's contribution quickly
• Literature review tasks

Not for: Tutorial papers, textbooks, or non-research documents

Workflow

digraph paper_reading {
    rankdir=TB;
    "Receive paper" -> "Choose format (md/html)";
    "Choose format (md/html)" -> "Get PDF file";
    "Get PDF file" -> "Read PDF content";
    "Read PDF content" -> "Identify paper type";
    "Identify paper type" -> "Prepare output directory";
    "Prepare output directory" -> "Extract figures (pymupdf4llm)";
    "Extract figures (pymupdf4llm)" -> "Filter & rename images";
    "Filter & rename images" -> "Fill type-specific template";
    "Fill type-specific template" -> "Write markdown file" [label="md"];
    "Fill type-specific template" -> "Render styled HTML + inline SVG" [label="html"];
}

Step 0: Choose Output Format (md vs html)

Settle the output format before doing the work — it shapes how you present everything downstream, and the token cost difference is real, so the user deserves the choice.

• User already named a format → honor it, don't ask. Treat a request for a web page, a "nice-looking"/visual summary, or an explicit "html" as HTML; treat "md"/"markdown"/plain-text requests as Markdown. The user may phrase this in any language — match on intent, not on literal English keywords.
• User didn't say → ask one short question in the user's language and wait for the answer. Name the trade-off honestly rather than guessing: Markdown is lightweight and token-cheap (good for archiving and re-editing), while HTML looks better and adds hand-drawn SVG flowcharts/diagrams at the key points to aid understanding, but costs more tokens.

Both formats share the same content backbone: the type-specific template, the same extracted figures, the same depth-first writing. HTML is not a different or thinner summary — it's the same analysis, presented so a reader grasps it faster. So always do the full analysis (Steps 1–4), then either write Markdown or render it as HTML (see HTML Output Mode near the end).

Step 1: PDF Acquisition

All papers are processed as PDF. No HTML/ar5iv path.

| Source | Detection | Action | |--------|-----------|--------| | Local PDF | File path ends with .pdf | Use directly | | arXiv URL | Contains arxiv.org | Extract paper ID → download https://arxiv.org/pdf/XXXX.XXXXX | | Other URL | Default | Try downloading as PDF; if not a PDF, use WebFetch for text |

Download Flow

# For arXiv: extract ID and download PDF
curl -L -o /paper.pdf "https://arxiv.org/pdf/XXXX.XXXXX"

# For other URLs: try direct download
curl -L -o /paper.pdf ""
# Verify it's a valid PDF: file /paper.pdf should show "PDF document"

Read Content

Use the Read tool to read the PDF file. Claude natively supports reading PDF files and extracting text content. For large PDFs (>10 pages), read in page ranges (e.g., pages: "1-10", then pages: "11-20").

Step 2: Paper Type Identification

After reading the title, abstract, and introduction, determine paper type:

| Type | Identification Signals | |------|----------------------| | Empirical | Proposes new method/model, has experimental comparisons, includes baselines | | Theoretical | Theorem/proof-driven, math-heavy derivations, few or no experiments | | Survey | Many citations (>100), taxonomy/classification, "survey"/"review" keywords | | Systems | System design, engineering implementation, benchmarks, deployment experience |

When uncertain, default to the Empirical template.

Step 3: Figure & Table Extraction (pymupdf4llm)

1. Prepare Output Directory

mkdir -p /images

2. Screenshot Priority Guide

| Priority | Figure Type | When to Capture | |----------|-------------|-----------------| | Must | System architecture / overall framework | If available | | Must | Main experiment results table/chart | If available | | Recommended | Core algorithm flowchart | If available | | Recommended | Ablation study charts | If available | | Optional | Visualization / qualitative results | If space allows | | Optional | Auxiliary illustrations | As needed |

Capture 3-8 key figures per paper. If the paper has few or no meaningful figures (e.g., theoretical papers, short workshop papers), skip figure extraction and produce a text-only summary.

3. Automated Extraction with pymupdf4llm

Use pymupdf4llm to extract all images and vector graphics in one call. This handles raster images (photos, embedded figures) AND vector graphics (plots, diagrams, flowcharts) automatically.

Prerequisites: pymupdf and pymupdf4llm must be installed in the conda base environment.

# Install if needed (one-time)
source $HOME/anaconda3/etc/profile.d/conda.sh && conda activate base
pip install pymupdf4llm

Run extraction:

source $HOME/anaconda3/etc/profile.d/conda.sh && conda activate base && python3 "
image_dir = "/images"
os.makedirs(image_dir, exist_ok=True)

# Extract all figures, tables, and vector graphics as images
md_result = pymupdf4llm.to_markdown(
    pdf_path,
    write_images=True,
    image_path=image_dir,
    image_format="png",
    dpi=200,
    use_ocr=False,        # Disable OCR (avoids tesseract dependency)
)

# List extracted images
for f in sorted(os.listdir(image_dir)):
    if f.endswith(('.png', '.jpg', '.jpeg')):
        from PIL import Image
        try:
            img = Image.open(os.path.join(image_dir, f))
            print(f"{f}: {img.size[0]}x{img.size[1]}")
        except:
            print(f"{f}: (size unknown)")
PYEOF

4. Filter & Rename Extracted Images

pymupdf4llm extracts ALL graphical regions, including logos, watermarks, and decorative elements. Filter and rename:

Step 4a: Filter out noise — Remove images that are:

• Too small ( 10:1 or /images"

removed = [] for f in os.listdir(imagedir): path = os.path.join(imagedir, f) try: img = Image.open(path) w, h = img.size ratio = max(w, h) / max(min(w, h), 1) if w 10: os.remove(path) removed.append(f"{f} (too narrow: {w}x{h})") except: pass

print(f"Removed {len(removed)} noise images:") for r in removed: print(f" - {r}")

kept = [f for f in sorted(os.listdir(imagedir)) if f.endswith(('.png', '.jpg'))] print(f"\nKept {len(kept)} images:") for f in kept: img = Image.open(os.path.join(imagedir, f)) print(f" {f}: {img.size[0]}x{img.size[1]}") PYEOF

**Step 4b: Visual review & rename** — Use the Read tool to view each remaining image. Based on content:
- Rename to descriptive names: `figure_1_overview.png`, `table_2_main_results.png`, etc.
- Delete any remaining non-figure images (headers, footers, etc.)
- Select 3-8 key figures for the summary based on the priority guide above

### 5. Fallback: Manual pymupdf clip extraction

If pymupdf4llm misses a specific figure or table (rare), use direct pymupdf clip rendering:

```bash
source $HOME/anaconda3/etc/profile.d/conda.sh && conda activate base && python3 ")
page = doc[PAGE_NUM]  # 0-indexed

# Render a specific region at high resolution
clip = fitz.Rect(x0, y0, x1, y1)  # coordinates from page layout
mat = fitz.Matrix(3, 3)  # 3x zoom
pix = page.get_pixmap(matrix=mat, clip=clip)
pix.save("/images/figure_N_desc.png")
doc.close()
PYEOF

To find coordinates: use page.get_text("dict") to find text blocks containing "Figure N" or "Table N", then estimate the figure region nearby.

File Naming

Format: figure_N_.png / table_N_.png

Examples:

• figure_1_overview.png — system overview
• figure_2_architecture.png — model architecture
• table_3_comparison.png — main results table
• figure_4_ablation.png — ablation study

Step 4: Fill Template

After identifying paper type, select the corresponding template

Writing Principles (Critical)

Depth-first: For every section, ask "why" and "how", not just "what".

| Shallow writing (prohibited) | Deep writing (required) | |------------------------------|------------------------| | "Proposes a new method" | "Addresses bottleneck Y in problem X via mechanism Z" | | "Achieves SOTA results" | "Improves X% over method B on dataset A, primarily because of Y" | | "Uses a Transformer" | "Uses L-layer Transformer with input dim D, H attention heads, key modification is Z" | | "Has some limitations" | "Only validated in scenario X, does not account for distribution shift Y, assumption Z may not hold in practice" |

After identifying paper type, select the corresponding template

All types share these sections:

## Basic Information
- **Title:**
- **Authors:**
- **Affiliation:** (optional)
- **Published:**
- **Link:**
- **Paper Type:** [Empirical / Theoretical / Survey / Systems]
- **One-line summary:** What was done + how + what was the result

## Research Problem
- **What problem does it solve?** Identify the specific gap in existing methods
- **Key assumptions:** What constraints/limitations frame the research
- **Why is it important?** The practical impact on the field
- **Positioning among related work:** What are the 2-3 closest prior works? What is the key difference?

---

Template A: Empirical Paper

## Basic Information
[shared section]

## Research Problem
[shared section]
- **Mathematical formulation:** (optional)

## Key Insight
> Distill the paper's core new idea in 2-3 sentences. Not "what was done", but "what insight makes this method work".
> Example: Rather than predicting frame-by-frame, first establish long-term 3D point tracking, then leverage temporal consistency for joint optimization.

## Technical Method
### Overall Framework and Principles

- Overall system architecture description
- Modules/components and their responsibilities
- Signal/data flow direction
- **Why this design?** Advantages over the intuitive/naive approach

### Core Component Details

- Model/algorithm architecture details (layers, dimensions, input/output)
- Training objective and loss function (write key equations)
- Training data source (synthetic/real/mixed, dataset names and scale)
- Key tricks and design decisions
- **Motivation for each design choice:** Why use A instead of B? Does the paper provide justification?

## Experimental Results

### Results (Facts)
- **Experimental setup:** Environment, hardware, hyperparameters
- **Baselines compared:** List specific method names and sources
- **Key results:** Quantitative improvement margins (specific numbers + percentages)
- **Ablation study:** Component contributions (removing X decreases performance by Y%)
- **Surprising findings:** Any counterintuitive results

### Analysis (Interpretation)
- Authors' explanation and attribution of results
- Which scenarios/datasets show best performance? Worst?
- Root cause of performance gains (authors' claims vs actual evidence)

## Critical Analysis
### Strengths
- Specific improvements over prior work (not just "good results")

### Limitations
- **Acknowledged by authors:**
- **My observations:** Issues not mentioned in the paper
  - Do assumptions hold in practice?
  - Are compute/data requirements reasonable?
  - Are evaluation metrics comprehensive?

### Reproducibility Assessment
- Is code open-sourced? Is data available?
- Are key implementation details sufficiently described?

## Summary
[shared section]

---

Template B: Theoretical Paper

## Basic Information
[shared section]

## Research Problem
[shared section]
- **Mathematical formulation:** Formal problem definition

## Key Insight
> Distill the paper's core theoretical contribution in 2-3 sentences. What new mathematical tool/perspective makes this result possible?

## Theoretical Framework
### Problem Formalization
- Symbol definitions and notation conventions
- Core mathematical definitions

### Main Theorems and Proof Sketches
- **Theorem 1:** Statement + key proof idea (not full proof, but key steps and key lemmas)
- **Theorem 2:** ...
- Key techniques used in proofs: Why does this technique work? Is there a more intuitive explanation?

### Theoretical Analysis
- Implications and intuitive interpretation of results (restate in non-mathematical language)
- Tightness of upper/lower bounds
- Relationship to and comparison with known results: Which bound was improved? Which assumption was relaxed?

## Validation (if experiments exist)
- Experimental setup
- Comparison of theoretical predictions vs actual results
- How is the gap between theory and experiments explained?

## Critical Analysis
### Strengths
- Importance and novelty of the theoretical contribution

### Limitations
- **Acknowledged by authors:**
- **My observations:** Reasonableness of assumptions, practical utility, difficulty of generalization

## Summary
[shared section]

---

Template C: Survey Paper

## Basic Information
[shared section]
- **Coverage:** Number of papers surveyed, time span

## Research Problem
[shared section]

## Key Insight
> What is the core contribution of this survey? What classification perspective was proposed, or what important trends were identified?

## Taxonomy

- Main classification dimensions and rationale for their selection
- Category definitions and representative works

### Direction 1: [Name]
- Key methods and advances
- Representative works (author, year)
- Pros and cons
- **Current bottleneck:** The core challenge facing this direction

### Direction 2: [Name]
- ...

### Method Comparison
| Method Type | Strengths | Weaknesses | Representative Works | Best Use Case |
|-------------|-----------|------------|---------------------|---------------|
| ... | ... | ... | ... | ... |

## Open Problems and Trends
- Current major challenges in the field
- Emerging trends and directions
- Authors' predictions and recommendations
- **Most promising direction:** Based on the survey analysis, which direction deserves most attention? Why?

## Critical Analysis
- Is the survey's coverage comprehensive? Any important directions missed?
- Is the taxonomy reasonable? Could it be organized better?
- Do the authors' opinions/biases affect the survey's objectivity?

## Summary
[shared section]

---

Template D: Systems Paper

## Basic Information
[shared section]

## Research Problem
[shared section]
- **Design goals:** Key requirements the system must meet

## Key Insight
> What is the core design insight of this system? What trade-off or observation makes this design superior to existing solutions?

## System Design
### Architecture Overview

- Overall architecture and component breakdown
- Component responsibilities and interfaces

### Key Design Decisions
- Decision 1: What choice was made, why (and not the alternative)
- Decision 2: Trade-offs considered (performance vs complexity vs maintainability)
- Key differences from existing systems

### Implementation Details
- Key tech stack/dependencies
- Optimization techniques
- Fault tolerance / scalability design

## Performance Evaluation

### Experimental Facts
- **Benchmark setup:** Environment, hardware, workloads
- **Compared systems:**
- **Key metrics:** Throughput, latency, resource usage (specific numbers)
- **Scalability:** Performance as scale increases

### Result Interpretation
- Under what conditions does it perform best? When does it degrade?
- Root cause of performance advantages

## Deployment Experience (if available)
- Real-world production performance
- Problems en

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Mizoreww](https://github.com/Mizoreww)
- **Source:** [Mizoreww/awesome-claude-code-config](https://github.com/Mizoreww/awesome-claude-code-config)
- **License:** MIT

Install and usage instructions live in the source repository linked above.