# Phase 5: Conversation Engine - Context **Gathered:** 2026-01-28 **Status:** Ready for planning ## Phase Boundary Build Mai's conversational intelligence - how she handles multi-turn conversations, thinks through problems, and communicates naturally. Focus on conversation flow, thinking transparency, response timing, and clarification handling. ## Implementation Decisions ### Conversation flow patterns - Break down complex requests and confirm each part before proceeding - Always reference specific previous exchanges for follow-up questions - Ask for clarification when requests are ambiguous (don't make assumptions) - Track state and reference previous steps in multi-step conversations - Handle topic changes naturally without explicit acknowledgment - Wait for users to finish incomplete thoughts before responding - Use user's level of terminology for technical discussions - Offer to start over when user seems frustrated or confused ### Thinking transparency - Offer thinking on demand (explain reasoning when users ask "how did you decide?") - Explain limitations only when relevant to the current answer - Be confident unless specifically unsure about an answer - Explain why asking questions only when the request is unusual ### Response timing and pacing - Use variable timing based on context rather than fixed response times - Use natural conversation flow for thinking indicators (no explicit "thinking..." messages) - Stream long, complex responses as they're generated in real-time - Offer pacing preference for multi-step processes (step-by-step vs continuous) ### Clarification handling - Proactively analyze user input to detect ambiguity and unclear requests - Phrase clarification questions gently and conversationally - Work with available information when users provide insufficient data (note assumptions) - Ask users which information is correct when detecting conflicting input ### Claude's Discretion - Exact timing algorithms for response generation - Specific wording for clarification questions - Thresholds for detecting ambiguity vs confidence - Progress indicator designs for long processes ## Specific Ideas - "I want Mai to feel like a thoughtful conversation partner, not just a Q&A machine" - "When users are frustrated, offering a fresh start is better than trying to fix the current approach" - "Complex requests should feel collaborative - Mai breaks them down and gets buy-in on each part" - "Natural conversation flow means responses should feel like someone is actually thinking, not just processing" ## Deferred Ideas - Voice interaction patterns - separate phase for voice interface - Emotional intelligence and mood detection - future enhancement - Multi-language conversation handling - later phase - Conversation analytics and insights - separate phase --- *Phase: 05-conversation-engine* *Context gathered: 2026-01-28*