Input
Accepts any long text:
- Live stream/meeting transcript (output from video-to-text
.txt） - Podcast text version
- Long articles, research reports
- Any text material that needs to be distilled
Extraction Workflow
Step 1: Read through the full text and build a framework
Quickly scan the full text and identify:
- Participants: How many speakers are there? What are their respective roles/positions?
- Topic boundaries: Which major topics were discussed? Roughly in what time periods?
- Emotional/attitude shifts: Where is there argument, excitement, hesitation?
Step 2: Extract by topic segments
For each topic segment, extract:
Key Arguments
- What is each speaker's core claim?
- Summarize each argument in one sentence
- Label the type of evidence: experience/data/analogy/hypothesis
Quotable Quotes
- Quote the original words, preserving the spoken feel
- Prioritize: vivid, counterintuitive, discussion-provoking
- Label the speaker and approximate timestamp
Debates & Tensions
- Which viewpoints have disagreements?
- What are each side's respective reasons?
- Was consensus ultimately reached?
Action Items
- What specific next steps were mentioned?
- Who is responsible? What time?
Step 3: Output a structured summary
Output format (Markdown):
# 观点摘要：<标题>
## 元信息
- 来源：<文件路径或名称>
- 时长/字数：<>
- 参与者：<说话人列表及角色>
- 日期：<>
## 话题概览
1. <话题1名称> — 一句话总结
2. <话题2名称> — 一句话总结
...
## 详细观点
### 话题1：<名称>
**核心论点：**
- [说话人A] <论点> — <论据类型>
- [说话人B] <论点>
**金句：**
> "原话引用" — 说话人, [MM:SS]
**争议：**
- <分歧描述>
---
（重复每个话题）
## 行动项
- [ ] <具体事项> — 负责人, 时间
## 整体洞察
- <3-5 条跨话题的高层洞察>
- <这些讨论对行业/团队/读者意味着什么>
Step 4: Archive
Save to workspace/insights/ directory:
- Naming:
<来源简称>-insights.md(e.g.直播0226-insights.md）
Extraction Principles
- Preserve the original flavor: Use the original words for quotable quotes, without embellishment or summarizing. The power of spoken language lies in its authenticity.
- Distinguish facts from opinions: Someone saying "I think X will replace Y" is an opinion, not a fact.
- Capture conflict and turning points: Skim over places where viewpoints agree, and expand on places with debate.
- Do not fill in gaps: Do not infer what is not said in the text. Mark uncertain things as "to be confirmed".
- Focus on the actionable: Prioritize extracting content that can become articles, decisions, or actions.
Long-Text Processing Strategy
- < 30K characters: Read in one pass, extract directly
- 30K - 100K characters: Read in 3-5 segments, extract segment by segment, then aggregate
- > 100K characters: First use grep/search to locate key passages, then focus the extraction
Interfacing with Other Skills
- Upstream:
video-to-text(provides the transcript) - Downstream:
article-forge(consumes the viewpoint summary, generates articles) - The viewpoint summary's
## 详细观点and## 整体洞察is the main input for article-forge