💡 Conversational Skill Discovery: Avatar-Fronted Capability Extension via Claude Skills API

[github-actions[bot]]

Summary

Leverage the newly GA Claude Skills API to let TalkTerm's avatar discover, explain, and activate new capabilities through natural conversation. Non-technical users never see a "marketplace" or "plugin store" — instead, the avatar proactively suggests relevant skills during workflows ("I notice you work with spreadsheets often. Would you like me to learn some new tricks for data analysis?") and handles installation transparently. The user experiences capability growth as the avatar "learning new things," not as software configuration.

Market Signal

Claude Skills API reached General Availability on August 20, 2026, alongside Computer Use, Browser Use, and the Files API. Claude Code already has a growing community skill ecosystem with skills for API development, code review, security analysis, and more. The Skills API enables agents to load versioned procedures and domain-specific capabilities at runtime. No consumer-facing product currently surfaces skill discovery through a conversational avatar interface — every existing skill/plugin marketplace (ChatGPT, VS Code, Slack) uses a traditional browse-and-install UI pattern. This is a blue-ocean UX opportunity.

User Signal

No existing TalkTerm Discussion covers Skills API integration. Meanwhile, TalkTerm's onboarding flow is actively being built (Issues #333-#338 — Guided Interactive Onboarding phases 1-4). Skill discovery naturally extends onboarding: after the initial tutorial, the avatar can progressively introduce new capabilities as the user's needs expand. This aligns with the emerging agentic UX pattern of "Goal-First Onboarding" (Gartner, Eleken) where users see outcomes, not tutorials. The avatar's progressive capability expansion mirrors the PRD's preference learning system (FR51) — the system adapts to the user, not the other way around.

Technical Opportunity

The Claude Agent SDK supports settingSources: ['project'] which loads skills from the working directory. TalkTerm's workspace selection (FR52-53) already establishes a working directory per session. Skills can be installed to the workspace's .claude/skills/ directory. The AgentBackend abstraction layer can expose skill metadata via a new SkillRegistry port interface. IPC channels (following the agent:* namespace convention) can add skill:discover, skill:install, skill:activate events. The IPC message bridge (architecture) already handles streaming agent state to the renderer — skill availability is additive metadata on this channel.

Assessment

Dimension	Score	Rationale
Feasibility	high	Skills API is GA. Claude Agent SDK already supports skill loading. Workspace directory structure is planned. Implementation extends existing architecture patterns without new dependencies.
Impact	med	Extends TalkTerm's capability surface without engineering new features. Each community skill multiplies what the avatar can do. However, value depends on the community skill ecosystem's growth and quality.
Urgency	med	Skills API just went GA. Early integration positions TalkTerm as the first avatar-fronted skill consumer. But the ecosystem is still nascent — value compounds as more skills are published.

Adversarial Review

Strongest objection: Non-technical users don't need or understand "skills" — they just want to get work done. Adding a capability layer creates complexity without clear user value.

Rebuttal: That's exactly why the avatar abstracts skills completely. The user never sees "skills" — they see the avatar offering to become more capable. The avatar says "I can now help you with budget forecasting" not "I installed the finance-analyst skill." This mirrors how humans learn new capabilities: progressively, contextually, and without metacognitive overhead. The technical complexity is hidden behind TalkTerm's core UX contract: the avatar handles everything. The user experiences a smarter assistant, not a software configuration step.

Suggested Next Step

Define a SkillRegistry port interface in src/shared/types/ports/. Test loading a community skill into a Claude Agent SDK session via settingSources. Build a prototype conversation flow where the avatar detects a user pattern (e.g., repeated spreadsheet questions), suggests a relevant skill, and activates it — measuring whether the user perceives capability growth or configuration complexity.