name: transcription-analyzer
description: >
Analyzes mock interview transcripts using multi-agent architecture with 4 parallel
analyst agents (Strengths, Mistakes, Behavioral, Factual) to produce confidence-scored
insights across 10 categories. Features anti-hallucination protocol requiring evidence
citation for every claim. Use when reviewing mock interviews, wanting interview feedback
analysis, or saying "analyze my transcript" or "mock review".
license: MIT
metadata:
author: vishnu-jayavel
version: "1.0"
categories: interview-prep, analysis, multi-agent

Transcription Analyzer

Analyze mock interview transcripts with comprehensive, confidence-scored analytics across 10 categories using multi-agent architecture.

Triggers

• "analyze my transcript"
• "transcription-analyzer"
• "mock review"
• "review my transcript"

Anti-Hallucination Protocol (MANDATORY)

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

Confidence Levels

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 references/confidence-scoring.md for detailed methodology.

Step 1: Input Handling

If ARGUMENTS provided:

Use the provided file path directly.

If NO ARGUMENTS:

Ask user for the transcript file path.

Step 2: File Validation

1. Load the transcript file
2. Validate the file exists and contains content
3. Count total lines for delegation decision

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 interview content.

Step 3: Interview Start Detection

Scan the transcript for trigger phrases that indicate when the actual interview begins (skip small talk):

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

Step 4: Interview Type Detection

Classify the interview type based on content signals:

System Design Signals

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

Coding Signals

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

Behavioral Signals

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

Output:
- Interview type with confidence level and evidence
- If unclear: "Unknown" with NOT_FOUND confidence

Step 5: Optional Diagram Analysis (System Design Only)

IF interview type is "System Design":

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:
Note that no diagram was provided and recommend saving diagrams for future review.

Step 6: Multi-Perspective Agent Analysis

IMPORTANT: Use parallel agents for comprehensive, bias-reduced analysis.

Launch 4 parallel agents to analyze from different perspectives. This prevents single-viewpoint blind spots.

Agent 1: Strengths Analyst

Find everything positive:
- Explicit praise from interviewer ("good", "nice", "I like that")
- Demonstrated competencies
- Strong moments and recoveries
- Communication wins

Output: {"positives": [{"title": "", "evidence": "", "confidence": "", "category": ""}]}

Agent 2: Mistakes Analyst

Find errors and problems:
- Technical errors corrected by interviewer
- Conceptual misunderstandings
- Communication issues (filler words, long pauses)
- Missed opportunities

Severity levels: CRITICAL (interview-ending), HIGH, MEDIUM, LOW

Output: {"mistakes": [{"title": "", "severity": "", "evidence": "", "confidence": "", "category": ""}]}

Agent 3: Behavioral Analyst

Assess Staff+ signals:
- Leadership presence (drove vs followed conversation)
- Trade-off articulation (made decisions, defended them)
- Depth of technical discussion
- Response to pushback/challenges
- Communication maturity

Output: {"behavioral": {"leadership": {...}, "tradeoffs": {...}, "depth_areas": [], "pushback_handling": {...}}}

Agent 4: Factual Verifier

Check technical accuracy:
- CORRECT: Technically accurate
- WRONG: Incorrect (cite the correction from transcript)
- NEEDS_VERIFICATION: Cannot determine from transcript alone

Only mark WRONG if interviewer explicitly corrected it.

Output: {"claims": [{"claim": "", "classification": "", "correction": "", "confidence": ""}]}

Synthesis

After all 4 agents return, cross-validate:
- If Strengths Agent found a positive but Mistakes Agent found related error → note the recovery
- If Behavioral Agent found leadership but Factual Agent found errors → assess net impact
- Resolve conflicts by citing evidence from both perspectives

Step 7: 10-Category Analysis

For each category, extract insights with confidence scoring and evidence citation.

Category 1: Scorecard

• Overall performance (1-10 scale) - Look for explicit feedback
• Level assessment (Junior/Mid/Senior/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: 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 5: Things That Went Well

• Explicit praise
• Demonstrated strengths
• Approaches that worked

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

Category 10: Interviewer Quality

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

Step 8: Output Formatting

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

Structure the report as:
1. Metadata (file, type, confidence)
2. Scorecard
3. Time Breakdown
4. Communication Signals
5. Things That Went Well (before mistakes!)
6. Mistakes Identified
7. Knowledge Gaps
8. Behavioral Assessment
9. Factual Accuracy Check
10. Action Items
11. Interviewer Quality
12. Confidence Summary

Include tables with evidence citations and confidence levels for each item.

Step 9: JSON Summary

After the markdown report, output a structured JSON summary with all categories for programmatic consumption.

Assets

• assets/sample_transcript.md - Example transcript for testing
• assets/sample_output.md - Example analysis output

References

• references/analyzer-prompt.md - Portable prompt for any LLM
• references/confidence-scoring.md - Confidence methodology