dull-bird/do.note-summary

Note Summary Skill

Fetch, read, and summarize any content into structured Markdown notes with guaranteed completeness. Uses a phased workflow with depth control and a post-generation quality checklist.

🚫 CRITICAL ANTI-PATTERNS

• DO NOT summarize based on abstracts, titles, or metadata alone. You MUST obtain the full original content.
• DO NOT produce the entire summary in one monolithic pass for long content. Break it into sections.
• DO NOT invent or hallucinate information not present in the source.

---

Depth Modes

The skill supports three depth levels, auto-detected or user-specified:

| Mode | Trigger | What You Do | |------|---------|-------------| | Quick | User says "quick", "skim", "glance" | Phase 1 only → 3-line summary | | Standard | User says "standard", "brief" | Phase 1 + Phase 2 → full structured summary | | Deep | Default. Also triggered by "deep", "critique", "精读" | Phase 1 + Phase 2 + Phase 3 → full summary + critical analysis |

If unsure, default to Deep.

---

Workflow

Phase 0: Input Detection & Content Fetching

Identify the input type and fetch the complete source content.

Input types:

| Type | Detection | Action | |------|-----------|--------| | URL (web page) | Starts with http | Fetch full page content | | URL (arXiv) | Contains arxiv.org | Visit /abs/ for metadata, then read the PDF via read_file | | URL (audio/video) | YouTube, podcast, etc. | Extract subtitles via yt-dlp or transcription tools | | Local PDF | .pdf extension | Use read_file to visually read (multimodal) | | Local Markdown | .md extension | Read directly | | Paper title / topic | No URL, looks like a title | Search for it, find the source, fetch full text |

Error handling:

• If only an abstract is obtainable after retries, prepend a bold warning:

  ⚠️ Warning: Full text not found. This summary is based solely on the abstract.

• If audio/subtitle extraction fails after all fallback attempts, report error and stop.

---

Phase 1: Screening (海选) — ~5-10 minutes

Goal: Understand what the content is about. Determine category and depth.

1. Skim the content: read title, introduction/abstract, conclusion, and key figures.
2. Categorize the content:

| Category | Examples | |----------|---------| | papers | Academic papers, research reports | | podcasts | Interview transcripts, long-form audio | | videos | YouTube summaries, tutorials | | articles | Blogs, news, technical tutorials |

3. Build the Section Manifest — create a mental (or written) outline of all major sections:

Manifest for: <title>
- [ ] Section 1: <heading> (approx scope)
- [ ] Section 2: <heading>
- [ ] ...
- [ ] Section N: <heading>
- Tables: T
- Figures: F

4. Write the preliminary output:
   • Frontmatter (complete)
   • > Core Summary (1-2 sentences)
   • ## 1. Key Insights (preliminary, 2-3 bullet points)

Decision point:

• If user requested Quick mode → finalize and stop here.
• Otherwise → proceed to Phase 2.

---

Phase 2: Structured Summarization

Goal: Produce a complete, section-by-section summary.

For each section in the manifest:

1. READ the section carefully.
2. SUMMARIZE with:
   • Key points with source references (timestamps for audio/video, section numbers for papers)
   • Preserved technical terms and proper nouns
   • Mermaid diagram if the section contains a process flow or concept hierarchy (use do.mermaid skill for syntax)
3. CHECK OFF the section in the manifest.
4. Link concepts — identify core concepts and create [[wikilinks]] to 40_Wiki/.

After all sections:

• Update ## 1. Key Insights with refined understanding.
• Write ## 3. Socratic Questions (≥2 challenging questions).
• Write ## 4. Related Concepts with wikilinks.

Decision point:

• If Standard mode → proceed to Phase 4 (Quality Checklist).
• If Deep mode → proceed to Phase 3.

---

Phase 3: Deep Analysis (精读) — Papers & Deep Mode Only

Goal: Critical analysis with mental reimplementation (Li Mu's Three-Pass Method, Pass 3).

For each major claim in the content, ask:

• What exactly is being claimed?
• What is the evidence?
• What are the hidden assumptions?
• Could this be done better?

"Mental Reimplementation" test:

• Can you close the paper and explain the core method from memory?
• Can you identify at least one weakness or improvement?

Output — append ## 5. Critical Analysis:

## 5. Critical Analysis

### Strengths
* ...

### Weaknesses / Limitations
* ...

### Potential Improvements
* ...

### Future Directions
* ...

For papers — Citation Double-Linking:

Discover citation relationships. If related articles exist in 60_Notes, use precise double links:

🚨 Format: [[<Paper Name>_summary|<Paper Name>]] — never bare [[Paper Name]].

**References:**
- [[Attention Is All You Need_summary|Attention Is All You Need]]

**Cited By:**
- DepthAnything-v2 (Title only, not in vault)

---

Phase 4: Quality Checklist (MANDATORY)

Before finalizing, you MUST verify every item. If any item fails, go back and fix it.

## Output Quality Checklist
- [ ] Frontmatter complete (type, title, category, date, source, tags)
- [ ] Core summary present (1-2 sentences, first-principles decomposition)
- [ ] Key insights listed (≥2 points with concise explanations)
- [ ] Detailed structure covers ALL major sections from source
- [ ] Source references included (timestamps, section numbers, page numbers)
- [ ] Mermaid diagram generated for key process/concept (if applicable)
- [ ] Socratic questions posed (≥1, genuinely challenging)
- [ ] Wikilinks added for key concepts → 40_Wiki
- [ ] Source reference in frontmatter is accurate and linked
- [ ] (Papers only) Citation links use [[Name_summary|Name]] format
- [ ] (Papers only) Critical analysis section present (Deep mode)
- [ ] (Audio/Video only) Subtitle file saved as source, audio deleted

Do NOT finalize the note until all applicable items are checked.

---

Phase 5: Archive & Link

1. Storage path: 60_Notes/<Category>/<Title>/

   • Summary: <Title>_summary.md
   • Original file (if local): copy to same directory
   • Example: 60_Notes/papers/Attention Is All You Need/Attention Is All You Need_summary.md

2. For audio/video: Save the complete subtitle file as source, delete audio (.m4a files are too large).

3. Auto-Linking via RAG:

   • MANDATORY: Execute python scripts/rag/query_vault.py . "<Core Topic or Title>"
   • Read the context. If highly relevant notes exist in the vault, extract the top 2-3 notes.
   • Append a ## Related Links section to the VERY BOTTOM of your generated summary output. Briefly explain why each past note is relevant and use proper [[Wikilinks]].

---

Output Template

---
type: note
title: "<Title>"
category: "<Category>"
date: YYYY-MM-DD
source: "[[<Original Filename or URL>]]"
depth: quick | standard | deep
tags:
  - <Tag1>
---

# <Title> Summary

> Core Summary: <1-2 sentences, first-principles decomposition of core value>

## 1. Key Insights

* <Insight 1>: Concise explanation.
* <Insight 2>: Concise explanation.

## 2. Detailed Structure / Key Discussion

### <Section A>

* <Key point + source reference>
* <Mermaid diagram if applicable>

### <Section B>

* <Key point + source reference>

## 3. Socratic Questions

* <Challenging question 1>
* <Challenging question 2>

## 4. Related Concepts

* [[Concept 1]]: Why it relates to this content.
* [[Concept 2]]: Connection explained.

## 5. Critical Analysis (Deep mode only)

### Strengths
* ...
### Weaknesses / Limitations
* ...
### Potential Improvements
* ...
### Future Directions
* ...

Rules

• Read deeporbit.json from the workspace root to determine the interaction language. Use this language for all your responses and generated note contents (e.g. zh-CN). The Obsidian folder paths themselves will ALWAYS remain in English.
• Use the run_command tool to execute obsidian open path="<absolute_path>" for every Markdown file you create or modify. See the do.obsidian-open skill for details.
• Use do.mermaid skill for all diagram syntax. Do not inline Mermaid rules.