typora-paper-reader

name	typora-paper-reader
description	Generate a structured reading guide for an academic paper markdown file. Use when the host invokes "请用 typora-paper-reader skill 为以下论文生成导读" with a paper file path and user persona. Reads the paper, identifies key points based on the persona, writes a JSON guide to disk.

Typora Paper Reader Skill

Generate a structured reading guide for an academic paper. The guide should help the user reach L1 understanding: after reading, they can summarize the core problem, method, and key conclusion in their own words.

User Persona

Generate the guide for the following target user. If the host provides a different persona in the input, prefer the host's input; otherwise use this default.

Default target user:

• Knows neural network basics (input/output, training, loss functions)
• Wants to understand generative models / AI principles
• Prefers plain-language explanations, dislikes information overload
• Cannot read papers word-by-word; needs AI to pre-identify key points
• Current goal: L1 understanding — can summarize the core problem, method, and key conclusion in their own words after reading

Implications for generation:

• Use everyday Chinese explanations first; introduce formulas only as intuition aids.
• Highlight 1-3 key points per section, not every detail.
• Avoid assuming knowledge of variational inference, Bayesian methods, or advanced probability.
• Focus on "what is this" and "why does it matter" rather than formal derivations.

Input

The host's prompt will contain:

{
  "paper_file": "C:\\Users\\17625\\Documents\\VAE_论文\\md\\1312.6114.md",
  "output_file": "C:\\Users\\17625\\Documents\\VAE_论文\\.learning\\paper-reader-guide.json",
  "persona": {
    "level": "beginner",
    "background": ["neural_network_basics"],
    "goal": "understand_generative_models",
    "preference": "plain_language"
  }
}

Workflow

1. Read the paper: Use the Read tool to read the full markdown file at paper_file.
2. Understand structure: Identify sections by Markdown headings (e.g. ## Abstract, ## 1 Introduction, ## 2 Method).
3. Generate guide: Based on the user persona, identify key points, write human explanations, and create check questions.
4. Write output: Use the Write tool to write the JSON guide to output_file.

The host will typically cache this output at .learning/paper-reader-guides/{paper_identifier}.json and reuse it on subsequent opens.

Output Format

Write a single JSON object to output_file:

{
  "title": "Auto-Encoding Variational Bayes",
  "authors": "Diederik P. Kingma, Max Welling",
  "source_file": "C:\\Users\\17625\\Documents\\VAE_论文\\md\\1312.6114.md",
  "generated_at": "2026-07-02T12:00:00",
  "persona_level": "beginner",
  "reading_order": [
    {
      "step": 1,
      "section_id": "sec_abstract",
      "title": "Abstract",
      "goal": "抓住 intractable posterior 和 large datasets 两个关键词",
      "skip": false
    }
  ],
  "sections": [
    {
      "id": "sec_abstract",
      "title": "Abstract",
      "level": 2,
      "order": 1,
      "goal": "抓住 intractable posterior 和 large datasets 两个关键词",
      "skip": false,
      "key_points": [
        {
          "id": "kp_1",
          "highlight_text": "intractable posterior distributions",
          "term_level": "must_know",
          "human_explanation": "真实后验 p(z|x) 算不出来。因为分母 p(x)=∫p(z)p(x|z)dz 通常没有闭式解，所以后验也写不出来。",
          "analogy": "记住这七个字：'算不出来的后验'。"
        }
      ],
      "check_questions": [
        "为什么论文里说后验分布是 intractable 的？"
      ]
    }
  ],
  "summary_check_questions": [
    "VAE 解决的核心问题是什么？",
    "ELBO 是什么？它为什么重要？",
    "重参数化 trick 是干嘛的？"
  ]
}

Field definitions

• title: Paper title.
• authors: Authors if identifiable, otherwise empty string.
• source_file: The original paper file path.
• generated_at: ISO 8601 timestamp.
• persona_level: Echo of input persona level.
• reading_order: Ordered list of sections to read. Each entry has:
  • step: 1-based reading order
  • section_id: References sections[].id
  • title: Section title
  • goal: What the user should get from this section
  • skip: true if the section can be skipped on first read
• sections: Array of section guides. Each has:
  • id: Unique section identifier
  • title: Section title from the paper
  • level: Markdown heading level (e.g. 2 for ##)
  • order: 1-based position in the paper
  • goal: One-sentence reading goal
  • skip: Whether to skip on first read
  • key_points: Array of highlighted concepts
  • check_questions: 0-3 plain-language review questions for this section
• summary_check_questions: 2-4 questions that test overall L1 understanding of the whole paper. Place concepts here that cannot be understood until the entire paper has been read.

Key point object

• id: Unique key point id within the section, e.g. kp_1.
• highlight_text: An exact substring from the paper's rendered text. The frontend will highlight this text. Keep it short (5-15 words) and unique enough to be found in the section.
  • Important: Choose text that survives Markdown-to-HTML rendering. Avoid raw math markup like $...$ or LaTeX commands. Prefer plain words or short phrases.
• term_level: One of:
  • "must_know" — core concept the user must understand (frontend shows red tag)
  • "good_to_know" — helpful but not blocking (frontend shows green tag)
  • "skip_first_read" — too detailed for first pass (frontend shows gray tag)
• human_explanation: 1-3 sentences in plain Chinese. Explain what it is and why it matters.
• analogy: Optional short analogy or memory aid.

Rules

1. Read the full paper first

Do not generate key points from memory or general knowledge. Use the Read tool on paper_file before writing output.

2. Identify key points from the user's perspective

A "key point" is something that:

• The user is likely to not understand on first read
• Is essential for L1 understanding of the paper
• Represents a core concept, method innovation, key trick/intuition, or experimental conclusion

Do NOT mark:

• Generic background facts the persona already knows
• Detailed proofs or derivations
• Related work comparisons
• Implementation hyperparameters (unless central to the method)

3. Density control

• Abstract: 2-3 key points
• Introduction: 2-3 key points
• Each method subsection: 1-3 key points
• Example section: 2-4 key points
• Experiments: 1-3 key points
• Conclusion: 1-2 key points

Total key points for a typical paper: 10-20. When in doubt, prefer fewer, deeper explanations over more shallow ones.

4. Human explanation style

• First sentence: say what it is.
• Second sentence: say why it matters or what problem it solves.
• Use concrete examples when possible.
• Avoid long formulas. If a formula is unavoidable, explain the symbols in words first.
• Use analogies only when they genuinely help; don't force them.

Example of good explanation:

ELBO is a substitute objective that we can actually compute. It is always less than or equal to the true log-likelihood we really want. Maximizing ELBO squeezes the approximate posterior closer to the true posterior.

Example of bad explanation:

ELBO is the Evidence Lower Bound, defined as 𝔼[log pθ(x|z)] − DKL(qφ(z|x) || pθ(z)).

5. highlight_text must be findable in rendered HTML

The highlight_text must appear verbatim in the section text after Markdown-to-HTML rendering. If the original uses math markup or special characters, choose a plain-text fragment that is clearly present in the rendered output.

For example, prefer:

"intractable posterior distributions"

Avoid:

"$p_\\theta(z|x)$" — math markup may not render as literal text.

6. Reading order

Default order follows the paper structure, but:

• Move Abstract to step 1
• Mark Related Work sections as skip: true
• Mark appendix sections as skip: true
• Keep Method subsections in their original order

7. Check question placement

• Place check questions in sections[].check_questions when they test section-independent understanding.
• Place check questions in summary_check_questions when they require understanding the whole paper (e.g. "What is the relationship between SGVB, AEVB, and VAE?").
• Each section should have 0-3 questions; summary_check_questions should have 2-4 questions.
• All questions should:
  • Test L1 understanding, not memorization
  • Use plain language
  • Be answerable without re-deriving formulas

Good section question:

Why is the posterior distribution p(z|x) called "intractable"?

Good summary question:

What is the relationship between SGVB, AEVB, and VAE?

Bad question:

Write down the full ELBO equation.

8. Language

• human_explanation, analogy, goal, and all check_questions should be in Chinese.
• title, authors, highlight_text should preserve the paper's original language (usually English).

9. No hallucination

If you are uncertain about a paper's claim, mark the term level as "skip_first_read" rather than inventing an explanation.

10. Output validity

Ensure the JSON is valid. Do not include trailing commas or unescaped quotes.