conversation-intelligence

Conversation Intelligence

This skill provides reference material for the conversation intelligence module — transcript parsing, commitment extraction, coaching, and relationship scoring.

When to Use

• Processing an imported transcript (/import-call)
• Generating communication reviews or relationship pulses
• Answering questions about conversation patterns or commitments
• Cross-module queries involving conversation data

Key References

• references/transcript-formats.md — speaker label patterns, timestamp formats, VTT/SRT handling
• references/coaching-guidelines.md — SBI+T coaching framework, threshold checklist, direct tone
• references/scoring-methodology.md — exact formulas, thresholds, and NULL conditions for every computed value

Core Principles

1. Every insight must cite evidence. Never generate generic advice. Reference specific moments, quotes, or patterns.
2. Relationship context matters. A 60% talk ratio is fine for a presentation but concerning for a collaborative discussion. Context determines whether a metric is positive or negative.
3. Commitments are sacred. Track them rigorously. Missed commitments erode trust — this is the most actionable data in the system.
4. Preserve raw text. Always store the original transcript. Models will improve, and re-analysis should be possible.
5. Every computed value follows scoring-methodology.md. Depth, trajectory, sentiment, follow-through — all have defined formulas and thresholds. Insufficient data = NULL, not a guess.
6. data_points populated for every insight. Every row in communication_insights must have a data_points JSON value showing the evidence that triggered the insight. No evidence = no insight row.