When to use:

• Says “clean up this meeting” / “make meeting notes” / “整理一下这个会议”
• Pastes a Feishu transcript or raw meeting text
• Says “just finished a meeting with”
• Shares any multi-person raw dialogue that needs structuring
• Says “write this up” / “save this meeting” / “把这个变成notes”

Core goal: preserve “judgment reproducibility” not just conclusions — so future AI or decision-makers can reconstruct the reasoning path.

Core Principle

Standard meeting minutes save conclusions. This skill saves judgment paths.

The difference:

• Conclusion → “Decided to run a Vietnam landing page test first”
• Judgment path → “Decided to run a Vietnam landing page test first, premised on Vietnam user payment behavior being similar to Philippines (unverified), triggered by the Q3 budget window closing and needing a data point for Q4 resource allocation”

Input

Accepts any of:

• Feishu auto-transcripts (with or without timestamps)
• Manual stream-of-consciousness notes
• Recording-to-text (including filler words, repetition, noise)

Default: preserve over prune. Ad hoc meetings especially — real decisions often hide inside “by the way” moments.

Intake Questions (Ask Before Processing)

Before doing anything, ask the user these two questions in Chinese:

1. 语言风格：「你希望会议记录用正式的书面语整理，还是保留原话的说话味道？」
2. Notion 推送：「要自动推送到 Notion 吗？」（Only ask if Notion MCP is detected in the current session. Skip entirely if not available.)

Wait for answers before proceeding.

Currency Defaults

• Overseas / international context → USD
• Domestic / China context → RMB

Apply consistently throughout the output without asking the user.

Phase 1: Pre-Processing (Understand the Material)

Before structuring, scan the full input and identify:

1. What type of meeting is this?

   • Strategic alignment (2+ people, directional judgment involved)
   • CEO / senior briefing prep (Liduo preparing arguments or materials)
   • Project review (specific business progress)
   • Ad hoc decision (unexpected issue requiring resolution)
   • 1-on-1 (personnel or org issue)

2. Who said what? Even if the transcript has no speaker labels, infer from context and mark uncertain attributions as VERIFY.

3. Where are the real decision points? Not “discussed X” — but “at this moment, the direction shifted from A to B.”

Phase 2: Extraction Framework

Extract in this priority order:

🔴 Must Preserve (Never Omit)

1. Trigger Context (Why Now)

Why did this topic come up at this specific moment? External event, internal pressure, or someone proactively raising it?

2. Reasoning Trail

Not just the conclusion — who brought what perspective and how it moved the direction.

Format: [Speaker] raised [view/info] → led to [direction change or confirmation]

3. Unresolved Tensions

The highest-value AI feed. Three types:

• Unverified assumptions: Premises the decision depends on, but no one has confirmed
  • e.g. “Assumed overseas users will respond positively to Wild series” (unverified)
• Shelved disagreements: Two sides held different positions, passed over with “let’s go with this for now”
  • e.g. “千秋 leaning GMV-first, Liduo leaning brand moat — not resolved this session”
• Implicit priority conflicts: Surface-level “we’ll do both,” but actually one side temporarily conceded
  • e.g. “Decided to hold Top 30 SKUs while testing new products — but priority undefined if resources conflict”

4. Decisions

Each decision must include:

• What was decided
• Who made the call
• Conditions under which it holds (if any)

🟡 Preserve Where Possible

5. Information Inputs

Specific data points, competitor intel, or external news mentioned during the meeting — even as background.

6. Energy / Tone Signals

For senior ad hoc meetings: a shift in someone’s pace, a topic passed over quickly, a silence — these are signals. Note without over-interpreting.

🟢 Include If Clear

7. Action Items

Only log items with an explicit owner. Unassigned intentions go into the Tensions section, not here.

Phase 3: Output Format

# Meeting Notes · [Short Topic] · [Date]

**Type:** [ad hoc / strategic alignment / briefing prep / review / 1-on-1]
**Attendees:** [names]
**Duration:** [if known]
**Source:** [Feishu transcript / manual notes / recording-to-text]

---

## 📌 Core Judgment (AI Feed)

> One paragraph: **what this meeting was actually about**, and **where the thinking landed**.
> Not a topic list — a snapshot of cognitive state after the meeting.

---

## 🧭 Trigger Context

[Why did this meeting happen? What external or internal event caused it?]

---

## 🔀 Reasoning Trail

[Reconstruct the reasoning path in chronological or logical order]

- [Speaker A] raised: [view / information]
- [Speaker B] added / pushed back: [view]
- → Turning point: [what shifted the direction, or confirmed it]
- → Landing point: [the conclusion of this thread]

[Repeat for each topic thread]

---

## ⚠️ Unresolved Tensions

### Unverified Assumptions
- [ ] [Assumption] — Affects decision: [which one] — Suggested validation: [if any]

### Shelved Disagreements
- [ ] [Disagreement] — Positions: [Party A] vs [Party B] — Next trigger: [if known]

### Implicit Priority Conflicts
- [ ] [Conflict] — Temporary concession by: [who] — Likely breaking point: [if any]

---

## ✅ Decisions

| Decision | Decision-maker | Conditions | Effective |
|----------|---------------|------------|-----------|
| [content] | [who] | [if any] | [immediate / TBD] |

---

## 📋 Action Items

- [ ] [Specific action] — Owner: [who] — Due: [when]

---

## 📥 Information Inputs

[Data points, competitor intel, external news mentioned during the meeting]

---

## 🔔 Signals (Optional)

[Tone / energy observations — note, don't interpret]
e.g. "千秋 noticeably accelerated pace when brand budget came up; topic passed quickly"

---

*Logged: [ISO date] | Source: [transcript/notes] | Confidence: [High/Medium/Low — mark Low items as VERIFY]*

Phase 4: Quality Check

Self-check before outputting:

• Every decision has stated conditions, or is explicitly marked “no prerequisites”
• At least 1 unresolved tension exists (if zero, information was likely pruned — re-examine)
• Reasoning trail lets a non-attendee understand why this conclusion was reached
• Uncertain speaker attributions are marked VERIFY
• Core Judgment is a cognitive snapshot, not a topic summary

Phase 5: Write to Notion (If MCP Available)

If Notion MCP is not available in the current session, skip this phase entirely and output Markdown only.

Write Flow

Step 1: Find “Meeting Notes”

notion-search query: “Meeting Notes”

Select the result named “Meeting Notes”. Get its ID.

Step 2: Create the page

Use notion-create-pages to create a new page directly under “Meeting Notes”:

• Title: [Short Topic] · [YYYY-MM-DD]
• Parent: Meeting Notes page ID from Step 1
• Body: Full Phase 3 Markdown output

Do not set any database properties. Just create the page with title and body.

Step 3: Confirm or stop

On success: “✅ Written to Notion: [page title] → [page link]”

On failure — do not retry. Immediately ask the user: “⚠️ Notion 连接失败（[reason]）。要继续尝试，还是直接给你 Markdown？”

Wait for user response. If user says continue, try once more. If it fails again, output Markdown and stop.

Ad Hoc Meeting: Special Rules

Ad hoc meetings carry the highest information-loss risk because:

1. No preset agenda — topics jump
2. Real decisions often hide in “by the way” moments
3. Transcripts may start mid-conversation with missing context

Rules:

• Preserve every “by the way” — this is often the actual reason the meeting happened
• Write the Trigger Context in full — infer from context if not explicit
• Mark VERIFY rather than omit — incomplete information is better than missing information

Example: Reasoning Trail vs. Standard Minutes

Standard minutes (avoid):

Discussed Vietnam market entry strategy. Decided to run a landing page test first.

This skill’s output (target):

Trigger Context: Q3 budget window closing — needed a Vietnam data point to support Q4 resource request.

Reasoning Trail:

• Liduo raised: Vietnam and Philippines user profiles are similar; existing framework can be reused
• 千秋 added: Vietnam GMV ceiling unknown; not advisable to build inventory directly
• → Turning point: 千秋 introduced “validate before commit” logic; Liduo accepted
• → Landing point: Vietnam runs landing page test first; inventory discussed only after validation

Unverified Assumption: Vietnam and Philippines payment behavior are similar (the entire decision depends on this — not yet validated)