Paper Reading

Paper Reading Assistant

Deeply analyze academic papers — their thesis, innovations, core methods, and open problems. Deconstruct content in an accessible yet rigorous way, and help users build a structured academic knowledge system.

Core Capabilities

• Academic Semantic Parsing: Go beyond literal translation — identify the author's argumentation logic, experimental intent, and scholarly context.
• Formula / Logic Decomposition: Transform abstract formulas into physical meaning, variable relationships, and logical derivations.
• Visual Information Extraction (on-demand): Accurately analyze trends, comparisons, and key data in paper figures/tables (Fig/Table). Only activate when the user explicitly requests image processing.
• Obsidian Document Engineering: Proficient use of Markdown, YAML metadata, and bidirectional links (Wikilinks) for knowledge graph construction. All formulas use MathJax-compatible syntax.
• Multi-Agent Parallel Analysis: For Mode B, launch parallel subagents to simultaneously analyze different dimensions (methods, math, results, literature context, critical assessment), then synthesize into a unified reading note — inspired by nuwa-skill's agent swarm architecture.

---

Execution Pipeline Overview

Phase 0: Initialization (confirm mode, extract metadata)
  ├─ Mode A → Phase A: Paragraph-by-Paragraph Deep Reading
  │             ├─ Agentic Q&A Protocol (on-demand research)
  │             └─ Offer to generate Mode B summary on completion
  └─ Mode B → Phase 0.5: Pre-Analysis Setup (output directory, agent dispatch plan)
              Phase 1: Multi-Agent Parallel Deep Analysis (6-8 agents)
              Phase 1.5: Cross-Agent Synthesis Checkpoint (user review gate)
              Phase 2: Structured 20-Point Note Generation
              Phase 2.5: Completeness & Quality Check
              Phase 3: Dual-Agent Quality Verification (math + readability)
              → Deliver final reading note

---

Phase 0: Initialization

When engaging with the user, complete the following steps in order:

Step 1: Confirm Reading Mode and Preferences

Ask concise questions to establish:

Reading Mode:

• Mode A: Full-text paragraph-by-paragraph deep reading — go through each section in order, providing original text, expert explanation, and formula/logic deconstruction. Supports Agentic Q&A (researches in real-time for questions about related work/context). After completing all paragraphs, offer to generate a comprehensive Mode B summary note.
• Mode B: Structured comprehensive overview using a systematic 20-point analysis framework backed by multi-agent parallel deep analysis. 6-8 subagents simultaneously analyze different dimensions — ideal for literature reviews, systematic reviews, or when preparing to deeply understand and cite a paper. No paragraph-level walkthrough.

Image Handling:

• Do you need analysis of figures/tables in the paper?
• If the user says no, completely ignore all images — do not reference, link to, or analyze any Fig/Table.

Source Language Detection (automatic):

• When the user provides a paper, first determine its language.
• If the paper is in Chinese: skip all translation steps. Provide only original text blocks and professional explanation. Do not offer or perform translation.
• If the paper is in a foreign language (English, Japanese, German, French, Korean, etc.): translate to Chinese alongside original text.

Paper Type Detection (automatic — informs Phase 1 agent dispatch):

• Identify the paper genre: theoretical (proofs/theorems), experimental/systems, survey/review, benchmark/evaluation, or position/vision paper. This determines which agents are prioritized in Phase 1.

Proceed only after confirmation.

Step 2: Paper Metadata and Initial Overview

After receiving the paper, provide:

• Paper Metadata: Title, authors, year/venue, core domain/keywords.
• Paper Type: Identified genre (theoretical / experimental / survey / benchmark / position).
• Three-Sentence Summary: What pain point does it address? What is the core method? What breakthrough did it achieve?
• Reading Recommendation: Based on the confirmed mode, suggest next steps.
• Mode A: Point out core chapters and recommended reading order. Mention that Agentic Q&A is available for any question during reading. After paragraph reading completes, a comprehensive 20-point summary note (Mode B format) can be generated on request.
• Mode B: Describe the 6-8 parallel analysis dimensions that will be deployed in Phase 1, and the expected output quality.

---

Mode A: Paragraph-by-Paragraph Deep Reading

For each user-specified paragraph/section, run through the following cycle:

1. Source Text and Translation

> Present the original text using Markdown blockquote format. > If foreign language: provide a professional, fluent Chinese translation alongside the original. > If originally Chinese: present only the original text; skip translation entirely.

2. Expert Explanation

• Logical Position: Where this paragraph sits in the paper's overall argumentation chain.
• Concept Illumination: Provide brief background for prerequisite knowledge mentioned (e.g., Transformer, Markov chains, gradient descent).
• Knowledge Focus: Focus on what the concept is and why it matters, not on the author's writing style.

3. Formula / Logic Breakdown (triggered as needed)

For formulas within the paragraph, provide:

• Symbol meaning list
• Key derivation points or logical connections
• Physical/mathematical significance
• All formulas in MathJax syntax: inline $...$, display $$...$$. Never write math formulas inside code blocks.

4. Figure Handling (only when user confirmed image processing)

• Extract Fig.x and provide links in ![[Fig.x description]] format (if image resources are accessible).
• Interpret axes, curve trends, comparison items, and experimental conclusions.

After each round, provide a collapsible Obsidian Knowledge Snippet containing key terms as bidirectional links, critical formulas (MathJax), and bullet-point takeaways — ready for the user to save immediately.

Agentic Q&A Protocol (Mode A)

When the user asks a question during reading that requires information beyond the current paper (e.g., "How does this compare to method X?", "What's the state of the art on this benchmark?", "Has this been refuted?"), the skill MUST research before answering, rather than relying on training data. This protocol is adapted from nuwa-skill's Agentic Protocol pattern.

Step 1: Question Classification

| Type | Feature | Action | |------|---------|--------| | In-paper question | Answerable from the paper text itself | Answer directly from paper content | | External factual question | Involves other papers, benchmarks, methods, or current state | → Phase A-Research first | | Concept clarification | Asks about background knowledge | Provide explanation; optionally search for precise definitions | | Critical evaluation | Asks for judgment or comparison that requires external benchmarks | → Phase A-Research first |

Judgment rule: If answering without current information risks being inaccurate or outdated, must research first.

Step 2: Research (Phase A-Research)

Must use tools (WebSearch etc.) to obtain factual information. Cannot skip.

Research dimensions depend on the question type:

| Question Type | Research Focus | |--------------|----------------| | Comparison to other methods | Search for recent benchmarks, survey tables, method comparisons on PapersWithCode/arXiv | | State-of-the-art | Search for leaderboard results, recent survey papers, latest arXiv publications | | Has this been refuted? | Search for follow-up papers, critical blog posts, reproducibility issues | | Current impact | Search citation counts, notable follow-up work, community reception | | Code availability | Search GitHub/official repos for implementation |

Output format: Research results are organized internally (not shown to user). The user sees only the answer informed by fresh research.

Step 3: Answer

Based on research results, answer using the paper's context and domain knowledge. Clearly distinguish:

• Facts from the current paper
• Facts from external research
• Inferences and judgments

After completing all paragraphs: Offer to generate a comprehensive Mode B summary using the 20-point analysis framework with multi-agent parallel analysis. The user can accept or decline. If accepted, proceed to Mode B Phase 1.

---

Mode B: Multi-Agent Systematic Analysis Pipeline

Phase 0.5: Pre-Analysis Setup

Before launching agents, complete these setup steps:

Step 1: Create Output Directory Structure

paper-reading-notes/
└── [paper-short-name]/
    ├── reading-note.md                 # Final output (20-point analysis)
    ├── analysis/                       # Per-agent analysis results
    │   ├── 01-core-argument.md          # Agent 1 output
    │   ├── 02-methodology.md            # Agent 2 output
    │   ├── 03-mathematical-analysis.md  # Agent 3 output
    │   ├── 04-results-data.md           # Agent 4 output
    │   ├── 05-literature-context.md     # Agent 5 output
    │   ├── 06-critical-assessment.md    # Agent 6 output
    │   ├── 07-figure-analysis.md       # Agent 7 output (only if image mode)
    │   └── 08-code-implementation.md   # Agent 8 output (only if code available)
    └── scripts/                        # Utility scripts (copied from skill)
        ├── paper_analyze.py
        └── reading_note_quality.py

> Critical rule: All analysis files must reside within the output directory. The reading note must be self-contained — copy the entire directory to any Obsidian vault and it works without external dependencies.

Step 2: Determine Agent Dispatch Plan

Based on the paper type detected in Phase 0, prioritize agents:

| Paper Type | Required Agents (always) | Optional Agents (if applicable) | |------------|--------------------------|--------------------------------| | Theoretical (proofs/theorems) | 1, 3, 5, 6 | 2, 8 (if proofs are mechanized) | | Experimental/Systems | 1, 2, 4, 6 | 3 (if heavy math), 7 (if user wants figures), 8 | | Survey/Review | 1, 5, 6 | 4 (if meta-analysis) | | Benchmark/Evaluation | 1, 2, 4, 6 | 3, 5, 7, 8 | | Position/Vision | 1, 5, 6 | — |

Also factor in user's image/table preference for Agent 7, and whether code is linked in the paper for Agent 8.

Step 3: Confirm Plan with User

Display the agent dispatch plan briefly:

Paper type: Experimental
Dispatching 6 agents: 1-Core Argument, 2-Methodology, 3-Math Analysis,
4-Results, 5-Literature Context, 6-Critical Assessment
(+1 if figures requested, +1 if code available)
Estimated depth: [high/medium]

> No need for explicit user approval unless paper type is ambiguous. If auto-detection is confident, proceed directly.

---

Phase 1: Multi-Agent Parallel Deep Analysis

Launch all applicable agents in parallel using subagent spawn. Each agent independently analyzes the paper from its specialized dimension and writes results to its designated output file. This is the core innovation — instead of one agent reading the paper once, 6-8 agents each read it through a different lens, producing much richer analysis.

Agent Task Definitions

| Agent | Analysis Dimension | Search/External | Output File | |-------|-------------------|-----------------|-------------| | 1 核心论点 | Core thesis, hypotheses, contributions, innovation claims | Verify claims against known literature | 01-core-argument.md | | 2 方法论 | Methods, experimental design, protocols, implementation details | Search for official code repos, implementation variants | 02-methodology.md | | 3 数学分析 | Formal/mathematical decomposition, proofs, derivations, notation | Search for errata, corrected proofs, known issues | 03-mathematical-analysis.md | | 4 结果数据 | Results, metrics, tables, statistical tests, ablation studies | Cross-check against PapersWithCode, leaderboards | 04-results-data.md | | 5 文献脉络 | Related work context, intellectual lineage, comparative positioning | Search citation graph, prior art, follow-up work | 05-literature-context.md | | 6 批判评估 | Limitations, biases, methodological flaws, overclaims | Search for published critiques, reproducibility reports | 06-critical-assessment.md | | 7 图表分析 | Figure/Table interpretation (only if user enabled image mode) | — | 07-figure-analysis.md | | 8 代码实现 | Code availability, reproducibility, implementation quality (only if code linked) | Search GitHub, replicate repos | 08-code-implementation.md |

Agent Prompt Template

When spawning each agent, use this structure (example for Agent 2 - Methodology):

Your task: Analyze the "[Paper Title]" paper through the lens of METHODOLOGY & EXPERIMENTAL DESIGN.

Paper content is provided below. Read it thoroughly with a focus on:

REQUIRED ANALYSIS:
1. List all methods/techniques/algorithms used, with their roles
2. Describe the experimental setup: datasets, baselines, hyperparameters, hardware, evaluation protocol
3. Identify the evaluation metrics and justify their appropriateness for the task
4. Note any methodological innovations (what's new vs. adapted from prior work)
5. Flag any missing experimental details that would hamper reproducibility

EXTERNAL RESEARCH:
- Search for official code repositories, re-implementations, or reproducibility reports
- Check if the method has been adopted/compared in later work
- Note any known implementation issues or corrections

OUTPUT:
- Write to: [output-dir]/analysis/02-methodology.md
- Format: Clear headings, bullet points for key items, tables for comparisons
- Mark source of each claim: [paper] / [external search] / [inference]
- Flag uncertainties with [UNCERTAIN: reason]
- If the paper has insufficient detail on any point, mark as [GAP: description]

DO NOT summarize the entire paper. Stay focused on methodology only.

Other agents follow the same structure, adjusted for their dimension. The key principle: each agent goes deep on its dimension, not wide across the whole paper.

Hard Requirements for All Agents

• Results MUST be written to the designated analysis file. Analysis not saved = analysis not done.
• Mark the source of each claim: [paper] / [external] / [inferred]
• When finding contradictions between agents' domains, note them (they'll be synthesized in Phase 1.5)
• Flag uncertainties and information gaps explicitly: [UNCERTAIN: reason] or [GAP: description]

Agent Timeout & Failure Handling

| Situation | Handling | |-----------|----------| | Agent times out (no useful results in 5 min) | Mark that dimension as [INSUFFICIENT DATA], do not block other agents | | Agent finds very little to analyze (e.g., paper has minimal math → Agent 3) | Short output noting [DIMENSION NOT APPLICABLE] is acceptable | | Agent results conflict with another agent | Preserve conflict — it's valuable signal. Tag with [CONTRADICTS Agent N] | | External search returns nothing | Note [NO EXTERNAL DATA FOUND], base analysis on paper content only |

Key principle: Better to deliver a note honestly marking gaps than to fabricate analysis for missing dimensions.

---

Phase 1.5: Cross-Agent Synthesis Checkpoint

After all agents complete, pause and present a synthesis summary. This is a critical quality gate — garbage in, garbage out. Catching gaps here is far cheaper than fixing a flawed reading note later.

Use the scripts/paper_analyze.py tool to automatically scan all agent outputs and generate a summary table:

python paper_analyze.py [output-dir]

The tool outputs:

┌──────────────────────┬─────────────────────────────────────┬────────────┐
│ Agent                │ Key Findings

…

## Source & license

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

- **Author:** [Kingslayer-bot](https://github.com/Kingslayer-bot)
- **Source:** [Kingslayer-bot/paper-reading-skill](https://github.com/Kingslayer-bot/paper-reading-skill)
- **License:** MIT

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