Interview Analysis Skill

Analyze customer interview transcripts in parallel using the Teresa Torres Continuous Discovery framework.

Input: $ARGUMENTS

Parse arguments:

• folder-path: Directory containing interview files (required)
• --output: Path to findings markdown file (default: <folder>/Interview-Analysis-Findings.md)
• --batch-size: Number of interviews per parallel agent (default: 5)

Data Sensitivity

Interview transcripts contain customer PII — names, roles, companies, and potentially contact details or business-sensitive information.

• Anonymise by default. In the output findings file, replace customer names with identifiers (e.g., "Company A", "Participant 1") unless the user explicitly requests real names.
• Flag PII in quotes. When extracting direct quotes, strip or redact names, email addresses, and phone numbers that appear in the transcript.
• Output file location matters. The findings file persists on disk. Avoid writing to shared or synced directories (e.g., Dropbox, Google Drive) unless intended.
• Temporary agent files. Parallel agents write intermediate results to /tmp/. These aren't automatically cleaned up — note this to the user after analysis completes.

Execution Strategy

1. Discovery Phase

First, enumerate all interview files:

find <folder-path> -name "*.rtf" -o -name "*.txt" -o -name "*.md" | sort

2. Parallel Dispatch

CRITICAL: Use the Task tool to dispatch multiple agents IN A SINGLE MESSAGE.

For each batch of files, create a Task with:

• subagent_type: "general-purpose"
• prompt: Include the specific files AND the analysis template below
• run_in_background: true (for large batches)

Example dispatch pattern:

// Dispatch 3 agents analyzing 5 files each - ALL IN ONE MESSAGE
Task("Analyze interviews 1-5", files=[...], template)
Task("Analyze interviews 6-10", files=[...], template)
Task("Analyze interviews 11-15", files=[...], template)

3. Analysis Template (for each agent)

Each parallel agent receives this template:

For each interview file, extract:

1. Metadata
   • Company/Participant/Role
   • Interview type (Discovery/Concept Validation/Usability)
2. Pain Points (Unprompted)
   • Direct quotes with context
   • Classify: operational, integration, data, training, compliance
3. Workflow Use Cases
   • Trigger → Condition → Action patterns mentioned
   • Quote the exact customer language
4. Integration Systems Mentioned
   • System name, type, priority signal
5. Terminology Reactions
   • "Workflow" interpretation (SOP vs automation)
   • "Agent" interpretation
   • Alternatives suggested
6. Unprompted Questions (Strongest Signal)
   • Questions customer asked without prompting
   • These are highest-confidence insights
7. Leading Question Flags
   • Questions that may have biased responses
   • "Does this resonate?" patterns
8. Key Takeaway (1-2 sentences)

Output Format: Structured markdown with direct quotes.

4. Aggregation Phase

After all agents return:

1. Collect all outputs using TaskOutput tool
2. Merge into findings document:
   • Add each interview to the log table
   • Append detailed analysis sections
3. Update Emerging Patterns:
   • Tally pain points by frequency
   • Aggregate workflow use cases
   • Compile integration systems
   • Summarize terminology reactions

5. Pattern Synthesis

After individual analyses, update the Emerging Patterns section:

### Pain Points (by frequency)
| Pain Point | Count | Companies |
|------------|-------|-----------|

Count occurrences across all interviews for:

• Pain points
• Workflow use cases
• Integration systems
• Terminology confusion patterns
• Leading question patterns

Parallel Dispatch Rules

1. Independence: Each agent works on different files - no shared state
2. Batch sizing: 5-7 files per agent is optimal
3. Single message: Dispatch ALL agents in one tool call block
4. Background mode: Use run_in_background: true for >3 files/agent
5. Output files: Check with Read tool when agents complete

Example Invocation

/interview-analysis ./interviews/customers --output ./findings.md --batch-size 5

Error Handling

• If a file can't be read, log it and continue
• If an agent fails, note the batch and retry individually
• Always verify file count matches interview count at the end

Output

The skill produces:

1. Interview Log Table: All interviews with metadata
2. Detailed Analysis: Per-interview findings with quotes
3. Emerging Patterns: Aggregated cross-interview insights
4. Recommendations: Actionable takeaways from patterns

Verification

After completion:

1. Count files in folder
2. Count entries in log table
3. Ensure all accounted for (note multi-part interviews)