💡 Adaptive Complexity Scaling: Expertise-Aware Avatar Communication That Grows With the User

[github-actions[bot]]

Summary

The avatar automatically adjusts the complexity of its explanations, the granularity of presented options, and the depth of technical detail based on the user's demonstrated expertise — per domain, not globally. New users see 2-3 simple choices with plain-language explanations. Experienced users see nuanced options with contextual details. The system adapts based on interaction signals (skip speed, modification frequency, vocabulary used) and resets per domain — Sarah can be expert-level in brainstorming but beginner-level in architecture reviews.

Market Signal

2026 UX industry trend: "The best products adjust their interface based on user behaviour, expertise level, and context in real time" (confirmed by Parallel HQ, UXPin, YUJ Designs). Progressive disclosure of AI reasoning and adaptive content density per user expertise are identified as top-3 design challenges for AI products in 2026. 75% of knowledge workers now use AI daily — spanning an unprecedented expertise range that one-size-fits-all interfaces can't serve. The AI avatar market is growing at 33.1% CAGR ($0.80B to $5.93B by 2032), and differentiation through adaptive intelligence is a key competitive lever.

User Signal

TalkTerm's PRD defines two personas: Sarah (PM, non-technical, wants simplicity) and Marcus (developer, wants depth). The current architecture treats all users identically. No existing discussion addresses per-domain communication complexity adaptation — #320 (Progressive Trust Escalation) adjusts agent AUTONOMY, not COMMUNICATION style. #452 (Adaptive Voice Expressivity) adjusts speech STYLE, not content COMPLEXITY. The distinction matters: autonomy is about what the agent does on its own; complexity scaling is about how the agent communicates what it's doing.

Technical Opportunity

The renderer/context/ state management layer (React Context + useReducer) can track per-domain expertise scores as part of user session state. The agent backend abstraction layer (AgentBackend interface in shared/types/ports/) can accept a complexity parameter that shapes response formatting via the system prompt. The overlay card system (UX-DR2 ActionCards) already supports variable detail levels — this adds automatic level selection. The preference memory system (FR51, context-scribe) provides the persistence mechanism for cross-session expertise tracking. Dreaming (when available via SDK) can refine expertise models across sessions.

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires expertise signal taxonomy, per-domain tracking state, and response templating layer; moderate architecture impact
Impact	high	Directly solves the dual-persona challenge (Sarah vs Marcus) that defines TalkTerm's market positioning
Urgency	med	Not blocked by regulatory deadlines, but earlier implementation influences architecture decisions for response formatting

Adversarial Review

Strongest objection: Getting expertise calibration wrong is worse than not trying — users feel patronized when underestimated or lost when overestimated. The calibration problem may be unsolvable.

Rebuttal: Start conservative (always default to simpler) and only escalate on unambiguous signals: user explicitly says "show me more detail", user skips intermediate explanations, user uses domain-specific vocabulary. The avatar relationship makes miscalibration recoverable — "Would you like me to explain this differently?" feels natural from a colleague, not from a menu system. Always allow manual override: "Give me the technical version." Per-domain scoping prevents cross-contamination (being expert in one area doesn't assume expertise in all areas).

Suggested Next Step

Define the expertise signal taxonomy (which user behaviors indicate higher/lower expertise per domain). Design the per-domain expertise tracker as a renderer-side state management concern. Create mockups showing the same decision point at 3 complexity levels (beginner, intermediate, expert).