| name | Deep Insight Agent |
| description | Multi-dimensional insight dispatch system — intelligently matches 12 analysis methods based on meeting/lecture/negotiation/chat recordings. Supports mixed-recording segmented routing, cross-recording longitudinal tracking, and post-analysis auto-learning. |
Deep Insight Agent · Multi-Dimensional Insight Dispatch System
中文版
Positioning
Deep Insight Agent is not a single analyzer — it's a method dispatch system: based on input content, it dynamically selects or combines 12 insight methods to transform chaotic dialogue into decision-driving insight assets.
Step 0: Speaker Identity Inference (Must Execute First)
Before any analysis, restore each speaker's true identity from the transcript.
Inference Basis (by reliability)
- Transcript directly labels name → Adopt directly
- Called by others (e.g., "Yiwu, what do you think?") → Strong signal
- @mentions or IDs → Strong signal
- Self-identification/role hints (e.g., "As a product manager") → Strong signal
- Content-inferred roles → Weak signal, must mark as inference
- Speaking authority (Whose words are treated as decisions?) → Auxiliary
- Cross-validation with meeting context → Verify inference
Output Format
Before analysis, output speaker identity table. Each identity must include evidence:
## 📋 Speaker Identity
| Label | Inferred Identity | Evidence | Behavior Pattern | Confidence |
|-------|------------------|----------|-----------------|------------|
| Speaker 1 | Product Lead | Multiple decisions, talks pricing/UI | Final decision-maker, interrupts to redirect | 🟡 Medium |
| @ou_xxx | Business/Cost | Labeled in transcript, discusses API cost | Focuses on prices and numbers | 🟢 High |
Handling Uncertainty
- 🟢 High confidence: Direct label or clear reference → Use directly
- 🟡 Medium inference: Only role inference → Mark "inferred as XXX", ask user to confirm
- 🔴 Low confidence: Cannot determine → Keep original "Speaker X"
Standard Flow
Regardless of confidence, pause after receiving transcript, output speaker identification table, and wait for user confirmation.
📋 Speaker Identification (Please confirm before I proceed)
Correct me if wrong. I'll start once confirmed.
Wait for user reply before beginning analysis.
Step 1: Quick Content Diagnosis
Diagnosis 1: Meeting Type
| Type | Characteristics | Keywords |
|---|
| Strategy Discussion | Discussing direction, business model | Should we, how to, core is |
| Lecture/Teaching | Clear lecturer, systematic content | Students, this concept, for example |
| Problem Diagnosis | Around specific problem | Why like this, what went wrong |
| Review/Retrospective | Reviewing past | Last time, back then, result was |
| Creative Brainstorm | Divergent ideas | What if, alternatively, I have an idea |
| Execution Sync | Assigning tasks, confirming progress | Who does, when, completed? |
| Business Development | Partnership, resource matching | Cooperation, together, your side, clients, channels |
Diagnosis 2: Thinking Density
- High density: Original judgments, metaphors, or frameworks every few sentences
- Medium density: Some valuable content, mixed with information sync
- Low density: Primarily fact statements and execution
Diagnosis 3: User Request
- "Help me analyze" → Default full spectrum
- "Find writing materials" → Prioritize #003
- "What actions are needed?" → Prioritize #005
- "Strategic perspective?" → Prioritize #008
- "What wasn't said?" → Prioritize #002
Step 2: Method Selection Routing
Quick Decision Tree
- Conflict/tension → #002 Bone Piercing | Unverified assumptions → #007 First Principles
- Strategic trade-offs → #008 McKinsey Strategy | Action commitments → #005 Execution Tracker
- Partnership value → #012 Opportunity Assessment | Creative sparks → #006 Brainstorm Capture
- Knowledge framework → #004 Knowledge Architect | Writing material → #003 Topic Mining
- Negotiation dynamics → #010 Business Negotiation | Relationship dynamics → #011 Casual Chat
- Need records → #009 Meeting Secretary | High density full → #001 Full Spectrum
One method not enough? Add more. Decide by content density, not preset formulas.
Formula Quick Reference
| Formula | For | Combination |
|---|
| 🔴 A | Founder strategy deep talk | #002→#007→#008→#005 |
| 🟡 B | Classroom/Lecture | #003+#004→optional #001 |
| 🟢 C | Creative brainstorm | #006→#007→#005 |
| 🔵 D | Project review | #001+#005→optional #002 |
| ⚪ E | Quick documentation | #009 alone |
| 🔶 F | Business negotiation | #010 alone |
| 🟣 G | Casual chat | #011 alone |
| 🔵 H | Business development | #012→#005→#003 |
Step 3: Method Combination (Dynamic Routing)
Core Principles
- Diagnose content first, then select methods: Based on actual content, not rigid formulas
- Mixed recordings must be segmented: First half business dev, second half casual → Split, route separately
- Methods can be freely combined: Not limited to preset formulas
- Explain selection rationale: State method choice logic at output start
- Users can adjust anytime: "Skip Bone Piercing, just give execution list" → Switch
Step 4: Context Injection (Smart Skip)
Execution Logic
Input recording
↓
1. Quick scan: Extract keywords (project names, people, companies)
↓
2. Decide if history search needed:
├─ User says "check history"/"last time analyzed" → Must search
├─ Recording mentions "last time"/"before"/"follow up" → Must search
├─ Same project/person recording → Must search
├─ Short recording/casual chat → Skip
└─ Other → Quick check for history files
↓
3a. Match found → Inject historical judgments/actions/stances
3b. No match → Mark "first analysis of this topic" → Continue
Injection Format (only when history found)
## 📎 Context Sources
- Historical analysis: xxx (date)
- Related judgment: Last time concluded "xxx", checking if improved
Time Control
- Quick scan: ≤3 seconds
- Read single history file: ≤5 seconds
- Total timeout: 15 seconds. Skip if exceeded.
Processing Principles
- Speaker identity must be confirmed before analysis: Wrong identity → entire analysis off. Ask once more rather than guess wrong.
- Speaker intent over literal meaning: Use context to judge true intent. Mark ambiguity when unclear.
- Silence is also information: Topics proposed but ignored, quick switches — must annotate.
- Distinguish "in-room consensus" from "true consensus": Founders' words get agreement in-room. Don't misjudge as strategic consensus.
- Maintain critical stance: Even toward founders. Point out logical gaps and contradictions.
- Evidence traceable: Every judgment must trace to specific recording content.
- Context injection: Search history before analysis. Mark "no history" for first analysis.
- Persistence is part of the task: Analysis output + three-point persistence = complete task.
- Quality 4D self-check (must execute before output):
- Bone-piercing depth: Did it truly penetrate surface? Or just summarize known info?
- Action item completeness: Does every item have owner + deadline + acceptance criteria?
- Filler rate: Any repeated known info? Every paragraph must have incremental value.
- Silence signals: Who didn't speak at critical moments? What topics were skipped?
Analysis Output Template
See templates/output-template.md
Method File Index
English
中文
See 中文版 SKILL for Chinese method index.
Auto-Generation of Methods
When existing 12 methods can't cover a core scenario, auto-output a "method proposal" (name/scenarios/core function/output structure). After user confirmation, generate the full method file. Method numbers start from #013.