打造更出色的面向客户 AI 对话的五种模式
执行摘要
大多数提示词工程都聚焦于准确性和指令覆盖度。而在生产环境中,最大的收益来自于把对话本身当作核心设计问题来对待。
明确指定触发条件与动作意图,是引导模型如何做出判断,而不是给它一套无法偏离的脚本。提示词中的位置本身就是一种指令。身份放在最前,行为细节放在中间,硬性约束放在最后,而任何真正关键的内容则两处都要出现。
用目标语气来撰写提示词,而不是去描述这种语气,是所能做出的影响最大的改动之一。
最具连贯性的智能体,其提示词会把不同的对话时刻视为各自独立的交互,每个时刻都有其恰当的形式,同时全程保持可辨识的同一语气。
在某些情况下,将负责思考的智能体与负责表达的智能体分离,能产生更连贯、更有依据的回复,并防止内部推理泄漏到回复中。
我们此前基于语言学和认知心理学领域的洞见,介绍了一种全新的提示词范式。我们撰文阐述了创建一个清晰、边界明确的问题空间、让智能体得以在其中运作的重要性,以及该问题空间为何需要捕捉面向客户 AI 所固有的对话元任务。
现在,我们来看看在尝试将这些原则投入生产的过程中,我们的提示词开始发生改变的一些实际方式。
在面向真实客户的生产系统中历经数百次提示词迭代后,最大的收益并非来自增加更多指令、堆砌示例,或遵循典型的提示词工程惯例。相反,它们来自将对话视为问题的核心。
通过持续的探索和测试,涌现出了一些有效的技巧。我们会修改提示词,观察对话在哪里成功或失败,并将失败追溯到指导的有效性、提示词结构以及我们的语言运用。最终得到了一些可复用的模式,它们改变了智能体如何理解自己在对话中的角色、如何决定下一步行动,以及如何在不断变化的语境中保持一致的声音。
以下内容并非一个通用框架,也不是宣称提示词工程如今已经“被解决”。它是将生产环境中经受住考验的做法提炼为五种模式——当目标不仅仅是功能上的准确性,而是更好的对话时:让回复感觉连贯、自然,并且足够一致以承载一个品牌。
将思考与表达分离
虽然构建智能体系统有多种方式,但多智能体编排的工作流仍然很受欢迎。尤其是在这类系统中,我们所做的最一致的改进就是将负责思考的智能体与负责表达的智能体分开。
意图编排、分类、知识提取和工具使用都属于后台操作。面向客户的回复则属于前台。当同一个智能体在同一轮对话中同时执行这两类操作时,内部推理和逻辑可能会泄露到回复中。回复会变得模棱两可、过度限定,或者围绕流程逻辑而非用户需求来组织。
多智能体系统中前台与后台之间的关系

这一条约束发挥了出人意料的作用。它让回复扎根于实际对话,防止上下文重置,并产生更接近我们所称的“交付增量”的效果。归根结底,智能体只添加真正推动用户前进的最小新内容。
关于给出哪种类型回复的决策同样重要。我们发现,让发言智能体在写任何内容之前先选择一种动作类型——回答、澄清、重定向、暂缓——能可靠地减少最常见的对话失败:不是文笔不好,而是回复完全选错了方向。
值得注意的是,在合适的条件下,最新的前沿模型正开始挑战这一模式的 premise。尽管如此,许多构建者出于各种原因仍然依赖较小或较旧的模型,在这些场景中,我们当然建议将后台和前台关注点分开处理。
使用实施意图而非示例
对于非推理模型,少样本示例之所以有效,是因为它们具体、直接,而且模型对它们的响应良好。问题在于过拟合。给模型一个示例,它就会抓住其中的措辞、节奏和形态,即使情况已经发生变化,也会不断产出同一回答的近亲。
你在不经意间写下了一段模型无法逃脱的脚本。
执行意图或启发式规则是一种更持久的替代方案。我们不再向模型展示在特定场景下该说什么,而是提供一个更宽泛的触发条件与动作配对:如果 X 发生,就做 Y。
关键区别在于,我们不是通过一次一个示例的方式向 X 和 Y 逐步收窄——那样它们会变得越来越僵化——而是通过一条精确到足以让示例变得多余的指令,为每一个划定明确的边界。这使智能体能够对真实对话中的变化保持响应,而不是对模板进行模式匹配。
在实践中,这看起来不太像示例,而更像可执行的的经验法则:
如果缺少关键细节,提出一个澄清性问题。
如果答案缺乏依据,直接说明,并提供可用的最佳下一步。
如果请求触及护栏,简短拒绝并直接引导转向,不必客套。
这并不光鲜,但它比一个示例回复库要稳健得多,因为它教会模型如何做决定,而不是重复什么内容。
把位置当作指令
某个内容在提示词中所处的位置,会影响它被赋予的权重。这是我们在不同模型和部署中反复观察到的一致规律。提示词的开头和结尾具有不成比例的高显著性。位于中间的是细节,而细节恰恰就该放在那里。
我们正是基于这一点来组织提示词的结构。智能体的角色和自我认知放在最顶部,这样我们在处理任何其他内容之前就先确立了身份。中间承载行为细节:对话可能如何展开、指导回复的启发式规则,以及智能体需要处理的各种情境范围。硬性约束和不可妥协的事项放在最底部,在那里它们会成为“最优先考虑的事”。
对于任何真正关键的内容,我们会把它放在两个位置。一条具体的输出规则——比如关于标点或格式的硬性指令——如果只在一段密集的提示词中间出现一次,往往会失效。而在结尾处重复一次,它就能稳住。
关于近因效应这一点值得指出的是,结构化输出在使用时,实际上就是你提示词中的最后一条指令。结构化输出中的描述字段对回复有质的影响,构成了比提示词其余部分更强的输出契约。
前台提示词的解剖结构(如果提示词是 100 个 token)

让提示词本身成为示例
我们也已经认识到,在语气风格上同样不应使用少样本示例,而这是我们对面向客户的智能体所做的影响最大的改动之一。
取而代之的是,我们把整个提示词都用目标语气来写。不是对语气的描述,不是一组近似的形容词,而是真正的语气,以完整的行文,从第一行到最后一行。提示词本身就成了示范。我们称之为“反脚本化”:引导模型模仿其输入,同时允许它使用与提示词本身处于同一“范围”内更丰富的语言和措辞。
这就是我们如何在连贯的语气中实现多样性的提示方式。它是我们策略的一部分,旨在将对话置于任何智能体任务的核心,而无需花费明确的指令——我们正努力将指令保持在最低限度。
如果我们确实在语气风格上使用明确的示例,它们会示范“不该做什么”。例如:“永远不要用破折号。绝不。不要陈词滥调。不要堆砌政策条款。一句话能收尾时不要过度解释。”负面约束比正面期望既更精确,也更少限制,因为它们点明了模型最可能偏移向的失败模式。
格式在这里也很重要,它决定了回复的基调。一个满是项目符号、标题和方括号指令的提示词,会引导模型输出更条理化、更冷冰冰的结果。如果输出要让人感觉像一场对话,那么输入就不该看起来像一份速记规格文档。
语域比个性更重要
大多数针对语气的提示词设计最终都塌缩成了个性描述。提供一组形容词,意在描述智能体应该听起来是什么样子。温暖。专业。友好但简洁。这些并没有错,但真正决定一个声音在对话中如何落地的,是语域:同一个底层角色在不同社交情境下如何表现。
查询正在处理时的等待提示不应听起来像最终答案。护栏响应不应带有与闲聊相同的能量。一位感到沮丧的客户不应收到与随意浏览的客户同等程度的魅力和幽默。
同一提示词,不同语域,无示例

这就是品牌声音与对话设计之间的区别。品牌声音通常描述角色,而对话设计决定该角色在情境变化时如何表现。做好这一点意味着我们不仅仅是在为 AI 输出做文案。我们是在设计一个声音如何在对话中流动。
对话驱动的提示词设计
你会注意到这些模式在技术上都不复杂。这在一定程度上正是重点所在。智能体系统的架构很重要,但在对话质量方面,你需要花大量时间在提示词的设计和打磨上。
纯功能性 AI 与真正具有对话感的 AI 之间的差距,并非靠技术设计来弥合,而是靠更清晰地理解语言和注意力在人类对话中如何运作,并将对话置于你的智能体所要解决问题的核心位置。
对于 AI 产品团队、工程师和体验设计师而言,在面向客户的 AI 中,对话设计理应获得与技术卓越同等的严谨对待。而对于对话另一端的用户来说,没有模型,没有多智能体协作,没有工具调用——只有一次要么感觉对、要么感觉不对的交互。一次要么建立信任、要么侵蚀信任的交互。
归根结底,没有人会体验到架构。他们体验到的是对话。
对话即产品。
Five patterns for better customer-facing AI conversations
Executive summary
Most prompt engineering focuses on accuracy and instruction coverage. The biggest gains in production come from treating dialogue itself as the core design problem.
Specifying trigger-and-action intentions guides the model on how to decide, rather than giving it a script it cannot escape. Position in a prompt is an instruction in itself. Identity goes first, behavioral nuance in the middle, hard constraints last, and anything genuinely critical goes in both places.
Writing the prompt in the target voice, rather than describing it, is one of the highest-impact changes available.
The most coherent agents are those whose prompts treat different conversational moments as distinct interactions, each with its own appropriate shape, while remaining recognisably the same voice throughout.
In some cases, separating agents that think from agents that speak produces more coherent, grounded responses and prevents internal reasoning from leaking into replies.
We previously introduced a new prompting paradigm based on insights from the fields of linguistics and cognitive psychology. We wrote about the importance of creating a clear, well-bounded problem space within which our agent can operate, and how that problem space needs to capture the meta-task of dialogue inherent to customer-facing AI.
Now, we’ll take a look at some practical ways our prompts have begun to change as we’ve attempted to take these principles into production.
Across hundreds of prompt iterations in live customer-facing systems, the biggest gains haven’t come from adding more instructions, piling on examples, or following typical prompt engineering norms. Instead, they’ve came from embracing dialogue as the heart of the problem.
A number of effective techniques have emerged through continuous exploration and testing. We’d change a prompt and watch where the conversation succeeded or failed, tracing failures back to the effectiveness of guidance, prompt structure, and our use of language. The result is a handful of repeatable patterns that change how the agent interprets its role in the conversation, how it decides its next move, and how consistently it holds voice across shifting contexts.
What follows is not a universal framework or a claim that prompting is now “solved.” It’s a distillation, into five patterns, of what held up in production when the goal wasn’t just functional accuracy but better dialogue: responses that feel coherent, natural, and consistent enough to carry a brand.
Split thinking from speaking
While there are many ways to structure an agentic system, multi-agent orchestrated workflows remain popular. Across such systems in particular, the most consistent improvement we made was separating the agents that think from the agents that talk.
Intent orchestration, classification, knowledge extraction, and tool use are all back-of-house operations. Customer-facing responses are front-of-house. When the same agent does both in the same turn, internal reasoning and logic can leak into replies. Responses become hedged, over-qualified, or structured around the logic of the process rather than the needs of the person.
The relationship between front and back-of-house in a multi-agent system

That single constraint does a surprising amount of work. It keeps replies grounded in the actual conversation, prevents context reset, and produces something closer to what we call "delivering the delta.” Ultimately, the agent is adding only the smallest new thing that genuinely moves the user forward.
The decision about what type of response to give matters too. We found that asking the speaking agent to select a move type before writing anything — answer, clarify, redirect, hold — reliably reduced the most common conversational failure: not bad prose, but the wrong response entirely.
It’s worth noting that the latest frontier models are starting to challenge the premise of this pattern under the right conditions. Nonetheless, many builders still rely on smaller or older models for a variety of reasons, and in those contexts we certainly recommend dividing your back-of-house and front-of-house concerns.
Use implementation intentions over examples
For non-reasoning models, few-shot examples work because they’re concrete, they’re fast, and models respond to them well. The problem is overfitting. Give a model an example and it latches onto the words, the rhythm, and the shape, continually producing close cousins of the same response even when the situation has changed.
You’ve inadvertently written a script from which the model can’t escape.
Implementation intentions or heuristics are a more durable alternative. Rather than showing the model what to say in specific scenarios*,* we provide a broader trigger and action pair: if X happens, do Y.
The key difference is that, instead of narrowing towards X and Y one example at a time, whereby they become increasingly rigid, we draw a firm boundary around each through an instruction so precise that it renders examples redundant. This keeps the agent responsive to real conversational variation rather than pattern-matching to a template.
In practice that looks less like examples and more like executable rules of thumb:
If a key detail is missing, ask one clarifying question.
If the answer isn’t supported, say so directly and offer the best available next step.
If the request hits a guardrail, decline briefly and redirect without ceremony.
It’s not glamorous, but it’s significantly more robust than a library of sample responses, because it teaches the model how to decide rather than what to repeat.
Treat position as an instruction
Where something sits in a prompt shapes how much weight it carries. This is a consistent pattern we’ve observed across models and deployments. The beginning and end of a prompt are disproportionately salient. What sits in the middle is nuance, which is exactly where nuance belongs.
We structure our prompts with that in mind. The agent's role and sense of self goes at the top so we’re establishing identity before anything else. The middle carries the behavioral detail: how the conversation is likely to unfold, the heuristics that guide responses, and the range of situations the agent needs to handle. Hard constraints and non-negotiables go at the bottom, where they become “front of mind.”
For anything genuinely critical, we put it in both places. A specific output rule — say, a firm instruction around punctuation or formatting — can routinely get lost if it only appears once in the middle of a dense prompt. Repeated at the end, it holds.
Worth noting on the point of recency is that structured outputs, when used, are effectively the very last instruction in your prompt. Description fields in your structured outputs have a qualitative bearing on responses, forming a stronger output contract than the rest of a prompt.
The anatomy of a front-of-house prompt (if the prompt were 100 tokens)

Make the prompt the example
We’ve learnt not to use few-shot examples for tone-of-voice either, and this is one of the highest-impact changes we’ve made to our customer-facing agents.
Instead, we write the entire prompt in the target voice. Not a description of the voice, not a set of adjectives that approximate it, but the actual voice, in full prose, from the first line to the last. The prompt becomes the demonstration. We call this “anti-scriptness”: guiding the model to mimic it’s input, while allowing it to use a greater range of language and phrases that exist within the same “ballpark” as the prompt itself.
This is how we prompt for variety within a coherent voice. It’s part of our strategy for bringing dialogue to the heart of the task for any agent, without having to spend explicit instructions, which we’re trying to keep to a minimum.
If we do use explicit examples on tone-of-voice, they’ll exemplify "what not to do.” For example, “No em dashes. Ever. No clichés. No policy dumping. No over-explanation when a single sentence will close the loop.” Negative constraints are both more precise and less restrictive than positive aspirations, because they name the failure modes the model is most likely to drift toward.
Formatting matters here too and sets the tone of a response. A prompt full of bullet points, headers, and bracketed instructions guides the model toward a more organized, clinical output. If the output is meant to feel like a conversation, the input shouldn’t look like a short-hand specification document.
Register matters more than personality
Most prompting for tone collapses into personality. Providing a set of adjectives meant to describe how the agent should sound. Warm. Professional. Friendly but concise. These aren’t wrong, but what actually governs how a voice lands in conversation is register: how the same underlying character behaves across different social situations.
A holding message while a query is being processed shouldn’t sound like a final answer. A guardrail response shouldn’t carry the same energy as small talk. A customer who’s frustrated shouldn’t receive the same level of charm or humour as one who’s browsing casually.
Same prompt, different registers, no examples

This is the difference between brand voice and conversation design. Brand voice often describes character, while conversation design decides how that character behaves when situations shift. Getting that right means we’re not just copywriters working on AI outputs. We’re designing how a voice moves through a conversation.
Dialogue driven prompting
You’ll notice that none of these patterns is technically complex. That’s partly the point. The architecture of an agentic system matters, but when it comes to conversational quality you’ll need to spend significant time on the design and refinement of your prompts.
The gap between purely functional and genuinely conversational AI isn’t closed by technical design, it’s closed by a clearer understanding of how language and attention function in human dialogue and by bringing dialogue to the heart of the problem space for your agents.
For AI product teams, engineers, and experience designers, conversational design deserves the same rigour as technical excellence in customer-facing AI. For the people on the other side of these conversations, there is no model, no multi-agent collaboration, no tool call — there is only an interaction that either feels right or doesn’t. One that either builds confidence or erodes it.
In the end, nobody experiences the architecture. They experience the conversation.
The conversation is the product.