---

name: transcription-analyzer description: > Analyzes conversation transcripts using Supervisor Agent architecture. First classifies session type (MockInterview, CoachingSession, GenericMeeting), then routes to specialized analysis workflows. Features anti-hallucination protocol with confidence scoring and evidence citation for every claim. Use when reviewing mock interviews, coaching sessions, meetings, or saying "analyze my transcript". license: MIT metadata: author: vishnu-jayavel version: "2.0" categories: interview-prep, analysis, supervisor-agent

Transcription Analyzer v2.0

Analyze conversation transcripts with intelligent session type detection and specialized analysis workflows.

Triggers

• "analyze my transcript"
• "transcription-analyzer"
• "mock review"
• "review my transcript"
• "analyze this meeting"
• "coaching session review"
• "analyze topic flow"
• "topic flow analysis"
• "how did we deviate?"
• "show me our tangents"

---

Session Types Supported

| Type | Description | Analysis Focus | |------|-------------|----------------| | MockInterview.SystemDesign | System design practice | Technical depth, architecture, trade-offs | | MockInterview.Coding | Coding interview practice | Algorithm, complexity, edge cases | | MockInterview.Behavioral | Behavioral interview practice | STAR format, leadership, communication | | CoachingSession | Mentoring/advice session | Key tips, action items, scripts/patterns | | GenericMeeting | Any other conversation | Summary, decisions, action items | | TopicFlowAnalysis | Long multi-person discussions | Topic hierarchy, deviations, filler words, visualizations |

---

Anti-Hallucination Protocol (MANDATORY)

Every metric and insight MUST include confidence scoring and evidence citation.

Confidence Levels

| Level | Score | Criteria | |-------|-------|----------| | HIGH | 90%+ | Direct quote from transcript, explicit statement | | MEDIUM | 60-89% | Inferred from context, multiple supporting signals | | LOW | 30-59% | Single weak signal, ambiguous evidence | | NOT_FOUND | 0% | No evidence in transcript - explicitly state this |

Rules (Non-Negotiable)

1. Never fabricate - If not in transcript, output "Not found in transcript"
2. Cite evidence - Every claim needs line number or direct quote
3. Distinguish inference from fact - Mark clearly: [INFERRED] vs [EXPLICIT]
4. Aggregate confidence - Overall score = weighted average of components

See prompts/confidence_scorer.md for detailed methodology.

---

Phase 0: Transcript Metadata Extraction (NEW)

CRITICAL: Run this phase BEFORE any content analysis.

This phase prevents speaker confusion by:

1. Identifying all speakers and their aliases
2. Detecting role swaps (e.g., peer mock interviews)
3. Mapping who gives feedback to whom
4. Creating feedback_direction_rules to filter what's included in the report

See prompts/phase0_metadata_extraction.md for the full extraction prompt. See prompts/transcript_metadata_schema.json for the output schema.

Phase 0 Quick Reference

When to run: Always, before Step 3 (Session Type Classification)

Key output: transcript_metadata JSON with:

• participants[] - All speakers with is_primary_subject flag
• segments[] - Time/line-based segments with role assignments
• analysis_context.feedback_direction_rules[] - Which feedback to include

Role swap detection triggers:

• "can you start by introducing yourself" (after questions were answered)
• "now let me interview you"
• "your turn to ask me questions"
• "let's swap roles"

Example feedback_direction_rules:

{
  "feedback_direction_rules": [
    { "segment_id": 2, "feedback_from": "vishnu", "feedback_to": "anish", "include_in_report": false },
    { "segment_id": 5, "feedback_from": "anish", "feedback_to": "vishnu", "include_in_report": true }
  ]
}

Downstream Agent Rules (MANDATORY)

All analysis phases MUST:

1. Check segment role_assignments before attributing ANY quote
2. Only include feedback where include_in_report == true
3. Never attribute feedback FROM primary subject as feedback TO them

---

Step 1: Input Handling

If ARGUMENTS provided:

Use the provided file path directly.

If NO ARGUMENTS:

Use AskUserQuestion to request the file path:

What transcript file would you like me to analyze?

Please provide the full file path (e.g., /path/to/transcript.md)

---

Step 2: File Validation

1. Use the Read tool to load the transcript file
2. Validate the file exists and contains content
3. Count total lines for delegation decision (>500 lines = use subagent)

Error Handling:

If file not found:

Could not find transcript at: [attempted_path]
Please check the file path is correct.

If file is empty:

The transcript file appears to be empty.
Please provide a transcript with content to analyze.

---

Step 3: Session Type Classification (SUPERVISOR AGENT)

CRITICAL: Classify session type BEFORE any detailed analysis.

Follow the classification algorithm in prompts/supervisor_classifier.md:

3.1 Scan for Signal Patterns

Scan the transcript for these signals (case-insensitive):

MockInterview.SystemDesign signals:

• "design a system", "scalability", "database", "load balancer", "API design"
• "high availability", "microservices", "CAP theorem", "partitioning"

MockInterview.Coding signals:

• "write a function", "time complexity", "algorithm", "test cases", "edge cases"
• "optimal solution", "brute force", "data structure", "Big O"

MockInterview.Behavioral signals:

• "tell me about a time", "STAR", "leadership", "conflict", "situation"
• "difficult situation", "disagree with", "mentor"

CoachingSession signals:

• "advice", "tips", "here's what I'd recommend", "you should try"
• "feedback on your", "let me coach you", "I suggest", "my recommendation"

3.2 Calculate Confidence Score

1. Count signal matches per session type
2. Normalize to percentage (0-100)
3. Determine confidence level:

   • = 70%: HIGH confidence → Route directly

   • 50-69%: MEDIUM confidence → Suggest confirmation
   • < 50%: LOW confidence → Require confirmation

3.3 User Confirmation (if needed)

If confidence is LOW or multiple types have similar scores, use AskUserQuestion:

**Session Type Classification**

I detected signals for multiple session types. Please confirm which best describes this transcript:

Options:
- Mock Interview - System Design
- Mock Interview - Coding
- Mock Interview - Behavioral
- Coaching/Mentoring Session
- General Meeting/Conversation

Detected signals:
[List top signals found with line numbers]

3.4 Output Session Type Badge

Display at the top of every report:

**Session Type:** MockInterview.SystemDesign [Confidence: HIGH 92%]

or

**Session Type:** CoachingSession [Confidence: MEDIUM 68%]

or

**Session Type:** GenericMeeting [Confidence: LOW - Default]

---

Step 4: Route to Specialized Analyzer

Based on session type classification, route to the appropriate workflow:

If MockInterview.*:

→ Continue to Step 5: Mock Interview Analysis → Use prompts/mock_interview_analyzer.md

If CoachingSession:

→ Jump to Step 10: Coaching Session Analysis → Use prompts/coaching_analyzer.md

If GenericMeeting:

→ Jump to Step 11: Generic Meeting Analysis → Use prompts/meeting_analyzer.md

If TopicFlowAnalysis:

→ Jump to Step 12: Topic Flow Analysis → Use prompts/topic_flow_orchestrator.md

---

Step 5: Interview Start Detection (MockInterview only)

Scan the transcript for trigger phrases that indicate when the actual interview begins:

| Trigger Phrase | Context | |----------------|---------| | "go design" | System design prompt | | "let's get started" | Formal interview start | | "the problem is" | Coding problem introduction | | "design a system" | System design prompt | | "let's dive into" | Technical start | | "first question" | Interview structure cue | | "walk me through" | Technical prompt |

Record:

• Line number where interview starts
• If no trigger found: analyze from beginning, flag LOW confidence on timing

---

Step 6: Mock Interview Subtype Context

Adjust analysis emphasis based on detected subtype:

SystemDesign:

• Emphasize architecture decisions, scalability trade-offs, component design
• Focus on design patterns and system thinking

Coding:

• Emphasize algorithm choice, complexity analysis, edge case handling
• Focus on code quality signals and optimization discussion

Behavioral:

• Emphasize STAR format usage, leadership signals, communication quality
• Focus on storytelling and impact articulation

---

Step 7: Optional Diagram Analysis (System Design Only)

IF interview type is "MockInterview.SystemDesign":

Ask user if they have an architecture diagram to analyze alongside the transcript.

IF diagram provided: Analyze:

• Components identified (services, databases, caches, queues)
• Data flow clarity (request paths, async flows)
• Missing components vs. verbal description
• Naming quality
• Diagram Quality Score (1-10)

IF no diagram:

[Confidence: NOT_FOUND] No diagram provided for analysis.
Tip: Save diagrams from future interviews for more comprehensive review.

---

Step 8: 10-Category Mock Interview Analysis

For transcripts over 500 lines, delegate to subagent using Task tool with subagent_type: "Explore".

Extract insights with confidence scoring and evidence citation for each category:

Category 1: Scorecard

• Overall performance (1-10 scale) - Look for explicit feedback
• Level assessment (E5/E6/E7/Staff+) - Look for explicit statements or infer
• Dimensions: Communication, Technical Depth, Structure, Leadership
• Readiness % = 100 - (P0_gaps × 15) - (P1_gaps × 5) - (CRITICAL_mistakes × 20) - (HIGH_mistakes × 10) - (MEDIUM_mistakes × 3)

Category 2: Time Breakdown

• Total interview duration
• Phase timings: Requirements, High-Level Design, Deep Dives, Q&A
• Time-related feedback from interviewer

Category 3: Communication Signals

• Talk ratio (candidate vs interviewer)
• Long pauses, filler words (um, uh, like, you know, basically)
• Clarifying questions asked
• Course corrections after feedback

Category 4: Things That Went Well

IMPORTANT: Show positives BEFORE mistakes (ADHD-friendly ordering)

• Explicit praise from interviewer
• Demonstrated strengths
• Approaches that worked

Category 5: Mistakes Identified

For EACH mistake:

• Title and description
• Severity: 🔴 CRITICAL, 🟠 HIGH, 🟡 MEDIUM, ⚪ LOW
• Category: Fundamentals, API Design, Patterns, Domain Knowledge, Communication
• Direct evidence with line number

Category 6: Knowledge Gaps

For EACH gap:

• Area/topic
• Category: Fundamentals, API Design, Patterns, Domain
• Priority: P0 (must fix), P1 (important), P2 (nice to have)

Category 7: Behavioral Assessment (Staff+ Signals)

• Leadership presence
• Trade-off discussions
• Depth areas
• Handling pushback

Category 8: Factual Claims

For EACH technical claim:

• The claim
• Classification: ✅ Correct, ❌ Wrong, ❓ Needs Verification
• Correction if wrong

Category 9: Action Items

• Explicit recommendations from interviewer
• Resources recommended (books, sites, problems)

Category 10: Interviewer Quality

• Feedback actionability (1-5 scale)
• Specific examples given (count)
• Teaching moments

---

Step 9: Mock Interview Output Formatting

Structure the report as:

## Mock Interview Analysis

**File:** [filename]
**Session Type:** MockInterview.[subtype] [Confidence: X%]
**Date Analyzed:** [timestamp]

---

### 1. Scorecard
[Overall score, level assessment, dimensional scores, readiness %]

### 2. Time Breakdown
[Duration, phase timings]

### 3. Communication Signals
[Talk ratio, filler words, clarifying questions]

### 4. ⭐ Things That Went Well
[Positives with evidence - BEFORE mistakes]

### 5. Mistakes Identified
[Severity-coded mistakes with evidence]

### 6. Knowledge Gaps
[Priority-coded gaps]

### 7. Behavioral Assessment
[Staff+ signals]

### 8. Factual Claims
[Verification status]

### 9. Action Items
[Recommendations and resources]

### 10. Interviewer Quality
[Actionability assessment]

---

### Confidence Summary
[Overall confidence, by-category breakdown, data quality notes]

After the markdown report, output a JSON summary for programmatic consumption.

→ END of MockInterview workflow

---

Step 10: Coaching Session Analysis

ONLY execute this step if session type is CoachingSession

Use prompts/coaching_analyzer.md for detailed extraction.

6-Category Extraction:

Category 1: Session Context

• Identify coach/mentor and mentee names
• Extract session topic(s) discussed

Category 2: Key Advice/Tips

• List all explicit recommendations from coach
• Categorize by domain: communication, technical, career, behavioral, other
• Include direct quotes with line numbers

Category 3: Scripts/Patterns

• Extract reusable phrases or frameworks taught
• Format as quotable, ready-to-use text
• Note context for when to use each script

Category 4: Action Items

• Explicit tasks assigned by coach
• Implicit tasks (things to practice, review)
• Prioritize by urgency: 🔴 High, 🟡 Medium, ⚪ Low

Category 5: Questions Raised

• Topics needing further exploration
• Unanswered questions from session

Category 6: Session Quality

• Actionability score (1-5)
• Concrete examples count

Coaching Output Format:

## Coaching Session Analysis

**File:** [filename]
**Session Type:** CoachingSession [Confidence: X%]
**Coach:** [name] | **Mentee:** [name]
**Topics:** [topic1], [topic2]

---

### 1. Key Advice & Tips
[Categorized by domain with direct quotes]

### 2. Scripts & Patterns
[Quotable text with usage context]

### 3. Action Items
[Explicit and implicit tasks with urgency]

### 4. Questions Raised
[Topics needing exploration]

### 5. Session Quality
[Actionability score, examples count]

---

### Confidence Summary

→ END of CoachingSession workflow

---

Step 11: Generic Meeting Analysis

ONLY execute this step if session type is GenericMeeting

Use prompts/meeting_analyzer.md for detailed extraction.

6-Category Extraction:

Category 1: Meeting Context

• Identify participants (if detectable)
• Identify meeting purpose/topic

Category 2: Summary

• 3-5 sentence executive summary
• Most important points discussed

Category 3: Decisions Made

• All explicit decisions reached
• Owner for each decision (if mentioned)

Category 4: Action Items

• All tasks assigned
• Owner and deadline (if mentioned)

Category 5: Open Questions

• Unresolved topics
• Topics needing follow-up

Category 6: Key Quotes

• Memorable or important statements
• Speaker attribution and line number

Meeting Output Format:

## Meeting Analysis

**File:** [filename]
**Session Type:** GenericMeeting [Confidence: X%]
**Participants:** [list]
**Purpose:** [detected purpose]

---

### 1. Executive Summary
[3-5 sentence summary]

### 2. Key Decisions
[Decisions with owners]

### 3. Action Items
[Tasks with owners and deadlines]

### 4. Open Questions
[Unresolved topics needing follow-up]

### 5. Key Quotes
[Memorable statements with attribution]

---

### Confidence Summary

→ END of GenericMeeting workflow

---

Step 12: Topic Flow Analysis

ONLY execute this step if session type is TopicFlowAnalysis

Use prompts/topic_flow_orchestrator.md for the full workflow.

Three-Phase Map-Reduce Approach:

Phase 1: Map (Chunk Analysis)

• Split transcript into 300-line chunks with 50-line overlap
• For each chunk: extract topics, speakers, timestamps, deviations, filler words
• Output intermediate JSON per chunk

Phase 2: Reduce (Merge & Reconcile)

• Merge topic hierarchies across chunks
• Deduplicate and reconcile overlapping segments
• Build global topic tree with parent-child relationships

Phase 3: Synthesize (Visualizations & Insights)

• Generate Topic Hierarchy (tree structure)
• Generate Sankey diagram data (topic flow)
• Generate Timeline visualization data
• Identify deviation patterns and tangent analysis
• Aggregate filler word statistics by speaker
• Produce actionable insights

Topic Flow Output Format:

## Topic Flow Analysis

**File:** [filename]
**Session Type:** TopicFlowAnalysis [Confidence: X%]
**Duration:** [total time] | **Speakers:** [count]

---

### 1. Topic Hierarchy
[Tree structure of main topics → subtopics]

### 2. Flow Visualization (Sankey Data)
[JSON for Sankey diagram: topic transitions with weights]

### 3. Timeline
[Chronological topic progression with timestamps]

### 4. Deviations & Tangents
[Where conversation deviated from main topics, duration, return points]

### 5. Filler Word Analysis
[Per-speaker breakdown: um, uh, like, you know, basically, etc.]

### 6. Insights
[Patterns, recommendations, conversation quality metrics]

---

### Confidence Summary

→ END of TopicFlowAnalysis workflow

---

Subagent Delegation (Large Transcripts)

For transcripts exceeding 500 lines, use the Task tool to delegate analysis:

Task tool with subagent_type: "Explore"
Prompt: [Include appropriate analyzer prompt + transcript content]

Validate JSON response structure before displaying results.

If subagent fails:

1. Attempt direct analysis with truncation warning
2. Analyze available portion
3. Note incomplete analysis in output

---

Portability Constraints

This skill MUST remain portable and dependency-free:

PROHIBITED references:

• mcp__mem0__* - No memory service
• mcp__obsidian__* - No note-taking service
• /Users/vishnu/ - No personal paths
• study plan - No integration with other skills
• gotcha - No gotcha tracking

ALLOWED tools only:

• Read - File reading
• AskUserQuestion - User interaction
• Task - Subagent delegation (with subagent_type: "Explore")

---

Examples

• examples/sample_transcript.md - Example mock interview transcript
• examples/sample_output.md - Example analysis output

Prompts

• prompts/phase0_metadata_extraction.md - Phase 0: Speaker/role extraction (run first!)
• prompts/transcript_metadata_schema.json - JSON schema for metadata output
• prompts/supervisor_classifier.md - Session type classification
• prompts/mock_interview_analyzer.md - Mock interview analysis
• prompts/coaching_analyzer.md - Coaching session analysis
• prompts/meeting_analyzer.md - Generic meeting analysis
• prompts/topic_flow_orchestrator.md - Topic flow analysis (Map-Reduce)
• prompts/confidence_scorer.md - Confidence methodology