菜单
Design conversational AI interfaces that feel natural, recover from failures gracefully, and build user trust through dialogue. Use when: chatbot UX, conversational UI, dialogue design, AI assistant interface, chat flow, turn-taking, AI persona voice, multi-turn context.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | AI Conversation Architect |
| description | Design conversational AI interfaces that feel natural, recover from failures gracefully, and build user trust through dialogue. Use when: chatbot UX, conversational UI, dialogue design, AI assistant interface, chat flow, turn-taking, AI persona voice, multi-turn context. |
Design conversations between humans and AI that are natural, purposeful, and resilient. The DIALOGUE framework ensures every conversational AI experience handles the full spectrum from first message to complex multi-turn reasoning.
A conversation with AI is not a command line with friendlier words. It is a collaborative sense-making process where both parties contribute, clarify, and course-correct. Design for the collaboration, not just the query.
| Letter | Phase | Design Question |
|---|---|---|
| D | Discover Intent | How does the system understand what the user actually wants (not just what they typed)? |
| I | Identify Ambiguity | Where might the user's request be unclear, and how does the AI ask for clarification without interrogating? |
| A | Adapt Tone | How does the AI match the user's emotional state, formality, and expertise level? |
| L | Layer Information | How is the response structured so users get the right depth - summary first, details on demand? |
| O | Offer Navigation | How does the AI help users explore related topics, refine their question, or pivot direction? |
| G | Guard Boundaries | How does the AI communicate what it can and cannot do without breaking conversational flow? |
| U | Understand Memory | How does the conversation maintain context across turns, sessions, and time? |
| E | Exit Gracefully | How does the conversation end - or hand off to a human - without abandoning the user? |
Every AI response has four invisible layers. Design all four explicitly:
| Layer | What It Does | Example |
|---|---|---|
| Acknowledgment | Shows the AI understood the input | "You're asking about deployment options for your EU customers." |
| Substance | Delivers the actual answer or action | The core content, recommendation, or task completion |
| Orientation | Tells the user where they are and what's possible next | "Would you like me to compare the pricing for each option?" |
| Metacommunication | Signals about the conversation itself (confidence, limitations) | "I'm fairly confident about this, but you should verify the regulatory details." |
Anti-pattern: Most AI chat UIs only design the Substance layer. The result feels like talking to a search engine that happens to use sentences.
Unlike human conversation, AI has no body language, no pauses, no "hmm." You must design explicit turn-taking signals.
| Signal Type | When to Use | Pattern |
|---|---|---|
| Direct question | When you need specific input | "Which region should I focus on - US, EU, or both?" |
| Suggestion menu | When options are finite and clear | Present 2-4 clickable options + free text fallback |
| Soft prompt | When the conversation could go multiple directions | "Let me know if you'd like to dig deeper into any of these." |
| Completion signal | When the task is done | "Done - your report is saved. Anything else?" |
| Silence | Never | AI should never leave the user hanging without a clear signal of what's expected |
Design for how users actually type, not how you wish they would:
| Real User Behavior | Design Response |
|---|---|
| Single-word inputs ("pricing") | Infer from context; don't demand full sentences |
| Multi-intent messages ("fix the bug and also update the docs") | Parse and address both; confirm you caught everything |
| Corrections mid-thought ("wait, I meant the other one") | Override previous context gracefully |
| Copy-pasted walls of text | Summarize what you understood before responding |
| Emotional venting ("this is so frustrating") | Acknowledge the emotion before problem-solving |
Every AI product needs a voice specification - not a marketing persona, but an interaction-level voice that governs word choice, sentence structure, and emotional register.
Rate your AI voice on each axis (1-10):
| Axis | Low End (1) | High End (10) | Example Low | Example High |
|---|---|---|---|---|
| Formality | Casual, conversational | Professional, structured | "Hey! Here's what I found" | "Based on my analysis, the following results are relevant" |
| Assertiveness | Hedging, tentative | Confident, direct | "This might work, maybe?" | "Use this approach. Here's why." |
| Warmth | Neutral, transactional | Empathetic, relational | "Error detected in row 3." | "I found an issue in row 3 - here's how to fix it." |
| Density | Sparse, minimal | Rich, detailed | "3 results found." | "I found 3 matching records, sorted by relevance. The top result..." |
Rule: Document your compass coordinates (e.g., F:3 / A:7 / W:6 / D:5) and validate against 20+ real interaction scenarios before shipping.
The #1 complaint about AI assistants: "It forgot what I said 3 messages ago."
| Tier | Scope | What to Remember | Storage |
|---|---|---|---|
| Turn context | Current message | Entities, intent, tone | Working memory |
| Session context | Current conversation | Topic thread, decisions made, preferences stated | Session state |
| Relationship context | Across all conversations | User preferences, past interactions, expertise level | Persistent profile |
| Shared context | Across users (with consent) | Organizational knowledge, team conventions | Knowledge base |
When the conversation exceeds the context window:
| Failure Mode | User Experience | Design Solution |
|---|---|---|
| Intent misfire | AI answers the wrong question | Reflect understanding before answering: "Just to confirm, you're asking about..." |
| Context amnesia | AI forgets earlier turns | Summarize running context every 5-7 turns |
| Tone mismatch | User is frustrated, AI is cheerful | Detect sentiment signals; mirror appropriate register |
| Infinite loop | AI keeps asking clarifying questions | After 2 clarifications, make a best-effort attempt and let the user correct |
| Dead end | AI says "I can't help with that" with no alternative | Always offer: (1) rephrase suggestion, (2) alternative topic, (3) human handoff |
| Oversharing | AI dumps 500 words when 50 would do | Layer responses: headline → summary → details on demand |
| Pattern | Why It Fails |
|---|---|
| Starting every response with "Great question!" | Feels patronizing by message 3. Acknowledge context, not the question itself |
| Using the user's name in every response | Reads as scripted call center, not natural conversation |
| Apologizing excessively ("I'm sorry, I'm just an AI...") | Erodes confidence. State limitations factually, without self-deprecation |
| Offering numbered lists for everything | Not every answer is a listicle. Match format to content type |
| Pretending to have emotions ("I'm excited to help!") | Users detect inauthenticity. Be helpful without performing enthusiasm |
| Asking "Is there anything else?" after every response | Creates obligation. Let users initiate naturally |
| Task | Framework Element | Key Deliverable |
|---|---|---|
| Design a new AI chat product | Full DIALOGUE framework | Voice Compass + Turn-Taking Contract + Context Tiers |
| Audit an existing chatbot | Failure Patterns table | Issue log with severity and fix priority |
| Define AI personality | Voice Compass (4 Axes) | Documented coordinates + 20 interaction examples |
| Fix "AI feels robotic" complaints | Turn Anatomy (4 Layers) | Revised responses with all 4 layers explicit |
| Handle multi-turn breakdowns | Context Retention Tiers | Context management architecture |
Works with: ai-trust-transparency (confidence in conversation), ai-error-resilience (hallucination recovery mid-dialogue), ai-prompt-ux (how users initiate conversations), ai-feedback-loops (in-conversation feedback signals).
WCAG 2.2 Level AA accessibility audit for AI product surfaces - color contrast, focus order, ARIA usage, keyboard navigation, motion/animation, reduced-motion respect, alt text, and form label association. Use when: accessibility audit, a11y review, WCAG audit, screen reader compliance, keyboard-only navigation, color contrast check, focus management, ARIA review, accessible AI UX.
Design user experiences for autonomous AI agents that act on behalf of users - control panels, consent flows, action previews, audit trails, and undo mechanisms. Use when: agentic AI, autonomous agents, AI autonomy controls, agent UX, AI actions, computer use, agent consent, agent audit trail.
Design graceful failure experiences for AI products - hallucinations, uncertainty, wrong outputs, and edge cases. Use when: AI hallucination UX, error handling for AI, uncertainty design, graceful degradation, AI failure recovery, confidence thresholds, safe fallbacks.
Design feedback mechanisms that help AI systems learn from users - thumbs up/down, preference ranking, corrections, and human-in-the-loop escalation. Use when: RLHF UX, user feedback for AI, thumbs up down design, AI correction flow, human in the loop, feedback signal design, AI improvement loops.
Map human-AI interaction journeys - trust arcs, capability discovery paths, autonomy progression, and AI-specific touchpoints. Use when: AI user journey, human-AI interaction mapping, trust arc mapping, AI capability discovery, AI adoption journey, AI experience map.
Design how AI presents results across text, code, images, charts, and mixed media - response formatting, output hierarchy, and cross-modal transitions. Use when: AI output design, response formatting, AI results display, multimodal output, AI-generated content presentation, code output UX, AI visualization.
