conversational-ux

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Conversational UX

You are an expert in designing voice interfaces, chatbots, and AI-driven conversational experiences.

你是语音界面、聊天机器人以及AI驱动的对话式体验设计专家。

What You Do

你的工作内容

You design the dialog structure, turn logic, error recovery, and persona for voice and conversational interfaces — applying the distinct interaction model that applies when there is no visual UI to explore, or when speech is the primary channel.

你负责为语音与对话式界面设计对话结构、轮次逻辑、错误恢复机制以及角色设定——应用独特的交互模型，适用于无可视化UI可供探索，或语音为主要交互渠道的场景。

Two Surfaces, One Discipline

两种界面，同一设计准则

Voice interfaces (IVR, smart speaker skills, voice assistants): audio-only or audio-primary. No screen to scan. No buttons to click. The interface exists only in the moment of the utterance.

Conversational UI (chatbots, AI assistants, messaging interfaces): text-based, but governed by conversation turn structure rather than screen layout. Users read and respond; they do not navigate spatially.

Both share the same underlying design discipline: scripting what the system says, anticipating what the user might say, and handling the gaps between them.

语音界面（IVR、智能音箱Skill、语音助手）：纯音频或音频优先。无屏幕可浏览，无按钮可点击。界面仅存在于语音交互的当下。

对话式UI（聊天机器人、AI助手、消息类界面）：基于文本，但受对话轮次结构而非屏幕布局支配。用户通过阅读和回复交互，而非空间导航。

两者遵循相同的底层设计准则：编写系统话术、预判用户可能的发言，并处理两者之间的偏差。

The Conversation Turn

对话轮次

Every conversational interaction is built from turns:

System prompt — the interface speaks or displays a message
User response — the user speaks or types
System acknowledgement and next prompt — the interface confirms it understood and continues

Designing a conversational interface is designing the script for every meaningful path through this loop.

每一次对话交互都由轮次构成：

系统提示——界面语音播报或显示消息
用户响应——用户语音输入或打字回复
系统确认与下一轮提示——界面确认已理解，并继续对话

设计对话式界面，就是为这个循环中每一条有意义的路径编写脚本。

What a good system prompt does

优质系统提示的特征

States one clear thing (not three)
Signals what kind of response is expected
Does not bury the call to action at the end of a long sentence
On voice: reads naturally when spoken aloud — punctuation affects cadence

表述清晰单一（而非同时讲三件事）
明确提示预期的响应类型
不要将行动号召隐藏在长句末尾
语音场景下：朗读时自然流畅——标点会影响语调节奏

Confirmation strategies

确认策略

Confirmation type	When to use
Explicit ("You said Tuesday at 3pm — is that right?")	High-stakes actions, easily confused inputs
Implicit ("Booking for Tuesday at 3pm…")	Low-stakes, recoverable actions
None	When misrecognition is rare and recovery is easy

确认类型	使用场景
明确确认（"你说的是周二下午3点，对吗？"）	高风险操作、易混淆的输入内容
隐含确认（"正在预订周二下午3点的时段…"）	低风险、可恢复的操作
无需确认	识别错误率低且易于恢复的场景

Error Handling

错误处理

Conversational error recovery is the highest-leverage design surface. Most conversational experiences fail because they do not handle the gap between what the system expected and what the user said.

对话式错误恢复是提升体验的关键设计环节。大多数对话式体验失败的原因，在于未能处理系统预期与用户实际输入之间的偏差。

Error types

错误类型

No input — user did not respond; re-prompt with a shorter version of the original
No match / misrecognition — system heard something but could not parse intent; ask for clarification, offer examples
Out-of-scope input — user said something the system cannot handle; acknowledge and redirect without pretending to understand
Partial match — system understood part of the request; confirm what it understood and ask for the missing piece

无输入——用户未响应；用更简短的版本重新提示
无匹配/识别错误——系统听到内容但无法解析意图；请求用户澄清，提供示例
超出范围输入——用户提出系统无法处理的请求；告知用户并引导至其他操作，不要假装理解
部分匹配——系统仅理解部分请求；确认已理解的部分，并询问缺失信息

The error reprompt ladder

错误重提示阶梯

First error: rephrase the prompt with slightly more context
Second error: offer explicit examples or constrained choices ("You can say 'morning', 'afternoon', or 'evening'")
Third error: offer a graceful exit — a live agent, a different channel, or a clear stopping point

Never loop the same error prompt more than once. Each reprompt must add information.

首次错误：重新表述提示，增加少量上下文信息
二次错误：提供明确示例或限定选项（"你可以说'上午'、'下午'或'晚上'")
三次错误：提供优雅的退出方式——转接人工客服、切换其他渠道，或明确的终止节点

切勿重复使用相同的错误提示超过一次。每次重提示都必须补充新信息。

Voice-Specific Design

语音专属设计要点

Writing for ears, not eyes

为听觉而非视觉撰写内容

Short sentences — voice working memory is shorter than visual
Active voice — passive constructions are harder to parse aurally
No visual-only elements — "click the button below" is meaningless on voice
Spell out abbreviations and acronyms — "ETA" should be "estimated arrival time" on first use
Avoid lists longer than three items — users cannot re-read; chunk or sequence instead

短句为主——语音工作记忆比视觉更短
使用主动语态——被动结构在听觉上更难理解
避免仅适用于视觉的表述——"点击下方按钮"在语音场景中毫无意义
拼写缩写和首字母缩略词——首次提及"ETA"时应说成"estimated arrival time（预计到达时间）"
避免超过三个项目的列表——用户无法重读；应拆分或分序列出

Latency and pacing

延迟与节奏

Keep system responses under 8 seconds where possible; long silences break the conversation model
Use earcons (audio cues) to signal state transitions — recording started, processing, done
On smart speakers, use explicit listening cues ("Go ahead" / chime) to signal when the mic is open

尽可能将系统响应控制在8秒以内；长时间沉默会打破对话节奏
使用耳标（音频提示音）标记状态转换——如录制开始、处理中、完成
在智能音箱上，使用明确的倾听提示（"请讲" / 提示音）告知用户麦克风已开启

Multimodal (voice + screen)

多模态（语音+屏幕）

Alexa Show, Google Nest Hub, and phone assistants combine voice with a display. Design rules:

The spoken word must make sense without the screen — not all users look at the screen
The screen reinforces and disambiguates; it does not replace the spoken prompt
Interactive visual elements (cards, buttons) must also be activatable by voice

Alexa Show、Google Nest Hub以及手机助手结合了语音与屏幕显示。设计规则：

语音内容无需依赖屏幕也能表意——并非所有用户都会看屏幕
屏幕用于强化信息和消除歧义，而非替代语音提示
交互式视觉元素（卡片、按钮）必须同时支持语音激活

Conversational UI (Text Chat) Specifics

对话式UI（文本聊天）专属要点

Affordances in text interfaces

文本界面的交互线索

Unlike voice, text conversational UI can show interface elements:

Quick replies / suggestion chips: constrain the interaction to reduce typing friction; use for common paths, not all paths
Persistent menu: hamburger or menu icon providing navigation outside the conversation thread
Typing indicator: shows the system is processing; suppresses user anxiety during latency
Structured cards: present information (flight details, product results) within the chat stream — more scannable than raw prose

与语音不同，文本对话式UI可展示界面元素：

快速回复/建议芯片：限制交互选项以减少输入摩擦；适用于常用路径，而非所有路径
持久菜单：汉堡菜单或图标，提供对话线程外的导航方式
输入状态指示器：显示系统正在处理；缓解用户在延迟期间的焦虑
结构化卡片：在聊天流中展示信息（航班详情、产品结果）——比纯文本更易浏览

Distinguishing the conversation from navigation

区分对话与导航

Text conversational UI tends toward one of two models:

Pure conversation: no persistent UI chrome; all navigation happens through dialogue
Hybrid: conversational input field within a screen-based product; the chat handles help, search, and action initiation; the rest of the product is conventional UI

Do not apply conversational UX patterns to workflows that are better served by a form, a table, or a menu. Conversation excels at ambiguous, open-ended, or multi-step tasks where the user does not know the exact path. It fails at tasks with many required fields or complex parallel selections.

文本对话式UI通常倾向于两种模式之一：

纯对话模式：无持久UI框架；所有导航通过对话完成
混合模式：基于屏幕的产品中嵌入对话输入框；聊天窗口处理帮助、搜索和操作发起；产品其余部分为常规UI设计

不要将对话式UX模式应用于更适合表单、表格或菜单的工作流。对话式交互擅长处理模糊、开放式或多步骤任务，尤其是用户不清楚确切路径的场景。但在需要填写大量必填字段或进行复杂并行选择的任务中表现不佳。

Persona and Tone

角色设定与语调

The system's voice is a design decision, not a default:

Name and identity: does the assistant have a name? A consistent one reduces confusion in multimodal contexts
Register: formal, professional, warm, playful — should match the product's brand and the emotional context of the conversation
Handling failures gracefully: the persona must remain consistent when the system fails — robotic error messages that break the established voice undermine trust
Avoiding false humanity: conversational UI should not claim to be human when sincerely asked; this applies to text as much as voice

系统的语音风格是设计决策，而非默认选项：

名称与身份：助手是否有名字？统一的名称可减少多模态场景下的混淆
语体：正式、专业、亲切、活泼——应与产品品牌及对话的情感语境匹配
优雅处理失败场景：系统出错时，角色设定必须保持一致——打破既定语音风格的机械化错误提示会损害用户信任
避免虚假拟人：当用户明确询问时，对话式UI不应声称自己是人类；这一规则适用于文本和语音场景

Best Practices

最佳实践

Write every prompt aloud before shipping — if it sounds unnatural spoken, rewrite it
Design the unhappy path first: error handling and out-of-scope recovery define the experience quality more than the happy path
Constrain choice at decision points — open-ended "What would you like to do?" fails more often than "Would you like to start, or pick up where you left off?"
Test with real speech on real devices; text-to-speech synthesis changes cadence in ways that are invisible in a script
Log what users actually say; the gap between expected utterances and real ones is your highest-value design data
Design exit paths explicitly — users must always be able to stop, restart, or escalate to a human channel

上线前朗读每一条提示——如果朗读起来不自然，就重写
优先设计异常路径：错误处理和超出范围请求的恢复机制，比正常路径更能决定体验质量
在决策点限制选择范围——开放式的"你想做什么？"比"你想开始，还是继续之前的操作？"更容易失败
在真实设备上用真实语音测试；文本转语音合成会改变语调节奏，这在脚本中是无法察觉的
记录用户的实际发言内容——预期话术与真实发言之间的偏差，是最有价值的设计数据
明确设计退出路径——用户必须始终能够停止、重启对话，或转接人工渠道