LLM: How to Avoid Sycophancy when Using AI

Being a responsible UX researcher means being a responsible AI user — namely, acknowledging and working within the constraints of current AI products.

As evidenced by our findings on prompt structure, chatbot limitations, and user behavior with AI products, AI responses to prompts constantly shift and change, leading to an unreliable and inconsistent user experience.

In particular, the user’s prompts influence responses throughout a chat session with a language model.

UX researchers should avoid sycophancy when using AI by:

1. Resetting conversations and sessions often. Beginning a new conversation or session with an AI language model will reduce how much the user’s opinions or inputs adversely influence the model’s responses.
2. Not expressing strong opinions or concrete positions during conversations with language models to avoid biasing the model’s responses. AI can supercharge confirmation bias, which is always dangerous for UX researchers.
3. Not relying exclusively on language models for fact-finding missions, especially in areas of knowledge that are unfamiliar to you. Catching false information is harder if you don’t already have some expertise in the topic. Treat AI bots as good starting points and keyword providers, but don’t trust them to do all your information seeking. Double-check anything you aren’t sure about.

From: Sycophancy in Generative-AI Chatbots by Caleb Sponheim

LLM Prompt Engineering: A Short Comprehensive Guide

In the evolving landscape of NLP, mastering prompt engineering is crucial for harnessing the full potential of LLMs. So, whether you’re aiming to perform sentiment analysis, named entity recognition, translation, text summarization, question answering, or complex reasoning, a well-crafted prompt is your key to unlocking the magic of Large Language Models.

Prompt Engineering Basics

LLM Types: Most modern LLMs are decoder-only transformers, like Falcon, GPT2, and others, but encoder-decoder models like Flan-T5 and BART are used for specific tasks like translation and summarization.

Starting with Prompting

Zero-shot vs Few-shot: Begin with a simple prompt and iterate. Zero-shot prompts work well with LLMs fine-tuned on instruction datasets. For more complex tasks, use few-shot prompts, which provide examples in the prompt to condition the model.

Best Practices

1️⃣ Choose the Latest Models: The most recent models tend to perform better.
Position Instructions: Place instructions at the beginning or end of prompts to ensure better model understanding.

2️⃣ Clear Separation: Clearly distinguish instructions from the text to which they apply.

3️⃣ Specific and Descriptive: Be precise about the task, specifying the format, length, style, language, etc.

4️⃣ Avoid Ambiguity: Craft unambiguous instructions and descriptions.

5️⃣ “What to Do” vs “What Not to Do”: Construct instructions that tell the model what to do rather than what to avoid.

6️⃣ Lead the Output: Guide the output by providing the first word or sentence to nudge the model.

7️⃣ Robustness Testing: Test prompts across different models to assess their robustness.

8️⃣ Version and Track: Keep track of prompt performance and versioning.

🚀 Advanced Prompting Techniques

➡ Few-shot Prompting: Provide examples in the prompt to condition the model for better results. Useful for more complex tasks.

➡ Chain-of-Thought (CoT): Encourage models to produce intermediate reasoning steps, enhancing results on complex reasoning tasks. This can be achieved by providing step-by-step instructions.

🔍 Prompting vs Fine-tuning

Fine-tuning Scenarios: Consider fine-tuning when your domain significantly differs from pre-trained data, working with low-resource languages, handling sensitive data, or constrained by budget, privacy, or infrastructure.

Dr Carmen Martinez on LinkedIn

Why not just use ChatGPT?

“Why don’t we just use ChatGPT?” 🤔

This is a question that is most likely being posed right now somewhere…

But here are five reasons you should not do this within a business context:

1. Absence of Direct Influence: Using ChatGPT means giving up control over your AI operations. It’s like a black box where you can’t provide input or influence its learning and results.

2. Lack of Specialised Learning: ChatGPT is a general AI assistant, so it will not understand specific terms and needs related to your business.

3. Personalisation Deficits: Customizing ChatGPT to match your business needs is difficult, unlike proprietary models made specifically for your requirements.

4. Legal and Security Risks: Data you put into ChatGPT might be used for its learning, which can raise concerns about confidentiality.

5. Limited Support: Facing complicated concepts or technical problems might make you struggle without enough help.

Entrusting business operations entirely to pre-trained models like ChatGPT is a risky venture.

A more balanced approach, wherein control and security are maintained while still harnessing the power of AI, may prove to be the optimal path forward.

James Crisp on LinkedIn

Prompt Instruction Cheat Sheet

Prompt instruction cheat sheet: Bot persona, user persona, verb, output format, length, tonality, edge case, conversation topics.

From Maximilian Vogel

Do we still need Conversation Designers?

Automation can replace individual processes or steps. However, as in many other areas, it is only one component in complex systems and processes.

As a rule, artificial and human intelligence complement each other fantastically here, as they are fundamentally different in the way they work.

As in other areas, generative AI will add many capabilities to conversational AI, but will not be able to replace humans in conversational AI design.

1.    𝐂𝐨𝐦𝐩𝐥𝐞𝐱 𝐒𝐜𝐞𝐧𝐚𝐫𝐢𝐨𝐬 𝐚𝐧𝐝 𝐏𝐫𝐨𝐛𝐥𝐞𝐦-𝐒𝐨𝐥𝐯𝐢𝐧𝐠:
Conversations aren’t always straightforward.
They can involve intricate problem-solving and multi-step interactions.
Conversation designers excel at designing these complex dialog flows, ensuring that AI-guided conversations remain logical, coherent, and goal-oriented even when navigating intricate scenarios.

2.    𝐄𝐭𝐡𝐢𝐜𝐚𝐥 𝐚𝐧𝐝 𝐑𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐥𝐞 𝐀𝐈:
Conversation designers can navigate potentially sensitive topics, address bias, and avoid harmful content that AI might inadvertently generate.
Their expertise helps maintain user trust and prevents AI systems from promoting misinformation or harmful ideologies.

3.    𝐈𝐭𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐈𝐦𝐩𝐫𝐨𝐯𝐞𝐦𝐞𝐧𝐭 𝐚𝐧𝐝 𝐅𝐞𝐞𝐝𝐛𝐚𝐜𝐤 𝐋𝐨𝐨𝐩:
Conversation designers can analyze user interactions, identify pain points, and iteratively refine the conversation models.
This human oversight ensures AI systems continuously evolve, becoming smarter and more user-friendly over time.

4.    𝐇𝐮𝐦𝐚𝐧-𝐂𝐞𝐧𝐭𝐫𝐢𝐜 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠:
Conversation designers bring a deep understanding of human emotions, cultural nuances, and contextual subtleties that AI might miss.
They can craft dialogues that resonate emotionally, ensuring interactions are compassionate, empathetic, and culturally sensitive, ultimately leading to more meaningful interactions between users and AI systems.

5.    𝐏𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐀𝐝𝐚𝐩𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲:
While generative AI can produce responses, conversation designers excel in crafting personalized interactions.
This personal touch creates a more engaging and tailored user experience that fosters long-term engagement.

Lisa Precht on LinkedIn

Why chatbots?

Top 9 Benefits for agencies using chatbot:

1. Customer Support
Chatbots can provide 24/7 customer support, answering FAQs instantly without human intervention. No more waiting for responses.

2. Lead Generation
With predefined scripts and qualifying questions, chatbots can effectively generate leads, capturing information seamlessly.

3. Booking Appointments
Schedule appointments with ease. Chatbots can access calendar data, provide available slots, and book appointments, all in real-time. Ideal for service-based businesses.

4. Surveys and Feedback
Gather customer feedback conversationally and engagingly, thereby improving response rates.

5. Multilingual
Serve global customers in their native language. Chatbots can be programmed in multiple languages, providing a personalized experience for diverse audiences.

6. Cost Savings
Chatbots save businesses money. They reduce the workload of customer support teams, handle multiple queries simultaneously, and require no rest.

7. User Engagement
Increase user engagement with interactive chatbots. From quizzes and games to interesting facts, chatbots can entertain as well as educate your customers.

8. Data Collection
Chatbot tracks user interactions, providing valuable insights into customer behavior and preferences.

9. Training and Onboarding
Chatbots can simplify the onboarding and training process for new employees. With access to databases and resources, they can provide instant answers to queries, help with system navigation, and ensure smooth assimilation into the company.

From Harsh Makadia’s LinkedIn post

‘Politeness’ in bot interaction

Politeness in its core essence is about three things: Respecting the other, treating the other as a unique person (being smart and using information about the other and about the context of the conversation to help the other), and being consistent — that is, engaging with a degree of integrity — with the other.

RESPECT

A voicebot that treats users with respect is more likely to win the cooperation of its users than one that does not. Here are some instances of how the voicebot should behave respectfully towards the user:

Respect the user’s time: for instance, avoid having the user suffer through long prompts; proactively tell the user how long they need to wait for an agent; offer to the user the option to be called back; let the user interrupt.

Respect the user’s freedom: let the user opt-out if they don’t want to interact with the voicebot; let them get back to the voicebot while waiting.

Don’t lie to the user: for example, don’t tell them that you are going to route them to a human and then have them interact with the voicebot.

Don’t blame the user: in cases where an error occurs, the voicebot should always take the blame.

Never terminate an interaction unilaterally: the act of ending a conversation unilaterally is the ultimate act of disrespect in the context of a conversation. Always make sure that the decision to end the dialog is consensual.

Tell the user what you are going to do. For instance, the voicebot should always tell the user that it needs to pause the dialog interaction for a few seconds to execute a back end action (e.g., retrieve something from the database); the voicebot should always tell the user that it is transferring the user to an agent.

INTELLIGENCE

On the flip side, the best way to win a user’s respect and, therefore, their cooperation is by acting intelligently. Here are some examples:

Know the user’s preferences: if the customer has selected English in previous interactions, don’t keep asking them what language they wish to use every time they engage. Note the language preference, remember it, and default to it.

Know the user’s level of expertise: treat frequent users who know the voicebot differently from first-time or infrequent users.

Anticipate the user’s requests: if a user has recently placed an order or submitted a ticket, the chances are that they are calling to inquire about that order or that ticket. Offer the user the status of that order the next time they call.

Detect and act on request spikes: if the voicebot is experiencing a sudden spike in interactions with humans, have the voicebot adapt its behavior in light of that spike: for instance, if the first three weekdays of every new month experience a spike in users engaging with the voicebot to inquire about their checking balance, then during the first three days of the month, have the voicebot volunteer to offer the user’s balance before lapsing to the main menu.

CONSISTENCY

Nothing unnerves a user more than an irrational machine. Every instance of inconsistency by the voicebot will occasion the user to ask, “Why is it behaving like this? Did I miss something, or is this thing just badly designed?” Obviously, such questioning can only hurt the user’s confidence in the voicebot’s ability to help them solve their problem.

In Language: be consistent in how you refer to objects, properties and actions across prompts and menus. Don’t use “ticket” in one prompt and “case” in another; “incorrect” in one and “invalid” in another; “log in” in one and “sign on” in another.

In modality: if the user can speak their answer in part of their engagement with the voicebot, don’t take that ability away from them in some other part unless you explain to them why you are taking it away from them.

Across contexts: if users are responding to an infomercial and the infomercial tells the viewers that by calling the line, they will get to a sales agent, then make sure that the voicebot does not offer options that have nothing to do with sales: e.g., offering them to be connected to the help desk or to billing.

Ahmed Bouzid

Voice Principles

This collection of design principles and rules was wholesale copied from Voiceprinciples (compiled by Ben Sauer).

Voice Platforms

Amazon Alexa

Design checklist

• Make it clear how customers can benefit from your skill
• Make sure customers can find your skill
• Design for natural language conversation
• Use good interaction design practices
• Handle unexpected user utterances gracefully
• Watch customers try to use your skill

How Alexa responds

• Be brief
• Speak and write naturally
• Prompt with guidance for the user
• Use conversation markers
• Add variety
• Use parallel language
• Remember what was said
• Provide definitive choices
• Use brevity, arrangement, and pacing when listing options
• Handle problems
• Provide contextual help

Legacy guidelines

Google’s Conversation Design Guidance

Principles

• Keep it short. Respect users’ time. Get to the point and get out of the way. 
• Give users credit. People know how to talk. Don’t put words in their mouth. 
• Be relevant and sensitive to context. 
• Delight the ear without distracting the mind. When adding personality, be sure it’s not over the top.
• Engage beginners and attract experts. Designing for many people doesn’t mean designing for the lowest common denominator.
• Take turns. Just asked a question? Stop talking. 
• Don’t read minds. Give them the facts and let them decide.

Greetings and goodbyes

• Tell users who you are
• Give the right amount of information
• End conversations appropriately

Conversational dialogs

• Sound natural
• Be cooperative
• Take turns

Conversation repair

• Prevent errors by expecting variations
• Provide helpful reprompts or pivot to another question
• Be prepared to help at any time
• Let users replay information
• Fail gracefully

Persona

• Reflect your unique brand and identity
• Keep users coming back
• Stay consistent

Google’s Conversation Design Best Practises

Be Co-operative, like your users

• Understand recognition grammars and repair prompting
• Accommodate diverse user speaking styles
• Let people know what they can say, intuitively

Unlocking the power of spoken language

• Communicate what the system understood
• Offer meaningful examples when letting people know what they can say
• Avoid stating the obvious
• Give instructions only if needed

Instilling user confidence through confirmations and acknowledgements

• Use explicit confirmation for clarity around high-risk requests
• Use Implicit confirmation for speed around simple requests
• Avoid “Go back” instructions
• Leverage acknowledgers to help reassure people they’ve been heard
• Randomize acknowledgers to avoid monotony and gimmickry

In conversation, there are no errors

• Don’t treat technical error “events” as users misbehaving
• Handle different types of error events with the appropriate strategy
• Prevent errors by providing help in the moment
• Know when to give up
• Make the success path more robust to “disguise” errors

Google’s principles of human conversation

• Give your VUI a personality. All voices project a persona whether you plan for one or not.
• Move the conversation forward. Try to anticipate moments when your VUI can keep the conversation going by offering more information and recognizing informative answers from users.
• Be brief, be relevant. Keep messages short and relevant. Let users take their turn. Don’t go into heavy-handed details until or unless the user will clearly benefit.
• Leverage context. Keeping track of the conversation and remaining “aware” of the user’s context will effectively advance the perception of human intelligence.
• Direct the user’s focus through word order and stress. To focus the user’s attention on what’s important, leverage their expectations of word order and stress placement.
• Don’t teach “commands”. Speaking is intuitive. If you have to explain a command, something’s wrong.

Apple’s Human Interface Guidelines: Siri

• Strive for a voice-driven experience that doesn’t require touching or looking at the screen. 
• Respond quickly and minimize interaction. 
• Take people directly to content. 
• Be relevant and accurate. 
• Be appropriate. 
• When a request has a financial impact, default to the safest and least expensive option. 
• Increase accuracy with custom vocabulary. 
• Provide example requests. 
• Don’t advertise. 

Microsoft’s Cortana: Voice Design Best Practises

• Design for the most common scenario. 
• Tasks should feel quick and easy. 
• Keep tasks glanceable.
• Help guide the user’s attention to the current focus of the conversation. 
• Clearly communicate forward progress during task completion. 

IBM: Conversation design guidelines

Achieve mutual understanding

• Recipient design. Adapt the dialogue as we do in everyday conversation: different topics or levels of detail depending on the person.  
• Minimization. Strive to minimize details for the user without sacrificing understandability.
• Repair. Build robust repair mechanisms, so that your conversational agent does not need to always get it right on the first try.

Practises

• Onboard users. Conversational agents should always be able to talk about what they can do or what they know.
• Progressive disclosure. Provide next steps sequentially and break down a process into bite-sized chunks. 
• History. Relay the current state of the conversation to the user. In voice-based interactions, use repetition tactfully to not only provide feedback but also to mark location.
• Artifacts. Whenever possible leverage the medium to facilitate the conversation. For example: a map with an X can better relay complicated instructions than words.
• Multimodal feedback. Give the user feedback through the conversation and the visual user interface (if present) to illustrate whether or not a request was heard or an action took place.
• Fail gracefully. Don’t be afraid to let the conversational agent admit a lack of understanding. Sometimes humans don’t understand each other either.
• Personality. Construct the persona of your agent somewhat like you would for your user. How serious or professional do you want your agent to be?

Voysis

• Have a compelling reason to use voice. Voice is a relatively new way to interface with apps and devices and users have habits that are developed already with those form factors. The solution should present a compelling reason to use voice over those existing habits (i.e. touch, click, etc.).
• Set user expectations and build trust. The user should understand what is and isn’t possible with the system and if the system doesn’t understand or can’t respond it should handle those situations in an empathetic, honest and helpful way.
• Make it naturally discoverable. Traditional GUIs have labels and constraints that help the user to understand where to go and what the system can do. VUIs need to be more flexible and at the same time allow users to understand what is possible when they are trying to complete a task through natural discovery.
• Mimic how people naturally speak. A good way to help users complete their task is to ensure, as completely as possible, the system can understand and respond using natural language. Most systems are fairly constrained in what they can process and how they can respond.
• Create context through modality and state. If a GUI is available it is essential to take advantage of it to help create a good voice experience. Modality allows the user to go back and forth between using voice and traditional interaction methods and gives the user visual feedback. It is also important that the user understands what state they are in. Is the system listening? processing? responding? Audibly or visually it’s important to give the user an indication.

Books

Designing Voice User Interfaces

by Cathy Pearl

• Conversation design. Humans rarely have conversations that only last one turn. Design beyond that one turn; imagine what users might want to do next.
• Set user expectations. Don’t ask a question if you won’t be able to understand the answer.
• Confirmations. Make sure that users feel understood, and let them know when they weren’t.
• Conversational markers. Let the user know where they’re at in the conversation.
• Error handling. Design for when things go wrong, because something will always go wrong.
• Don’t blame the user. 
• Novice and expert users. Adapt to the experience and expertise of the users. 
• Keep track of context. People don’t repeat terms in conversation, they use pronouns like ‘she’ after the subject has been established. Make sure your system understands the context of user input. 
• Help and other universals. Include a set of universals at every state: e.g.  repeat, main menu, help, operator, and goodbye.
• Latency. Use audible or visual cues to communicate unavoidable system delays to the user.
• Disambiguation. If a user gives ambiguous information, use contextual clues to make a smart guess or ask for clarity. 
• Accessibility. Design experiences for everyone, no matter their abilities. Make interactions: time efficient, provide context, and prioritise personalization over personality. 

Don’t Make Me Tap! Notes on Design 

by Bouzid / Ma

Respect

• Respect the user’s time
• Respect the user’s freedom
• Don’t lie to the user
• Don’t blame the user
• Never terminate an interaction unilaterally
• Tell the user what you are going to do
• Don’t switch modalities on the user without telling them

Intelligence

• Know the user’s preferences
• Know the user’s level of expertise
• Anticipate the user’s requests
• Detect and act on request spikes

Consistency…

• …in language
• …in voice
• …in modality
• …across exchanges
• … across contexts

Articles and Links

Grice’s maxims

“These maxims may also be understood as describing the assumptions listeners normally make about the way speakers will talk, rather than prescriptions for how one ought to talk.”

• Maxim of quality: Try to make your contribution one that is true.
• Maxim of quantity: Make your contribution as informative as is required (and no more).
• Maxim of relation: be relevant.
• Maxims of manner: Avoid obscurity of expression; be brief and orderly. 

BBC: Principles for designing a voice experience for children

Asking questions

• If you’re asking a question, end the sentence with it and listen for an answer immediately.
• Don’t tell children to ‘say this’ or ‘say that’, simply ask the question.
• Don’t ask rhetorical questions; children will answer them.
• Ask questions that have distinctive, easy-to-say answers.

Listening for answers

• When offering a choice, provide no more than three options.
• Strive to present options that are balanced in their appeal to children.

Handling errors

• Don’t keep children stuck in error loops. Turn a bad situation good by progressing them even when they are misunderstood.
• Don’t use language or tone to make the child feel as though they are to blame.

Writing content

• Use a real voice to speak to children in a warm, friendly tone – avoid cold, monotone, synthesised voices.

Fjord: Six principles for designing for voice UI

• Conversation as user interface. Conversations require cooperation between participants and mutual acceptance, understanding and respect. Same with a voice interface.
• The interface of least resistance. A user will typically look for the easiest way to complete a task, but their definition of “easy” can vary wildly depending on the context and situation.
• Everything happens in sequence. Effective conversational design follows familiar sequences and implements conversational structure and familiar queues to increase empathy, maintain a relationship with the user and motivate for further engagement.
• Context is key. Voice interfaces are most useful in more private, controlled environments, such as the car or home.
• What’s your name again? Using emotive expressions to address the user and conversational context to recognize them helps build trust.
• Mind your manners. The importance of helpful words, and all that implies, applies to voice UI, too. 

Back to top ↑

Fjord: six principles for designing engaging voice interfaces

by Michael Levy

• Create a conversation bubble. Build a small pool of shared memory between the user and the system.
• Keep it simple. Users look for the path of least resistance.
• Guide users from A to B. Use repeatable patterns of behaviour to build familiarity into interactions. 
• Know when to be seen, not heard. Understanding context is key to creating an engaging voice UI.
• Build empathy through personality. An increased level of empathy is beneficial to both parties.
• Open access to all. A truly engaging voice interface is one that doesn’t feel like it’s for some people and not others.

Jellyvision’s Jack Principles

Maintaining Pacing

1. Give the user only one task to accomplish at a time 
2. Limit the number of choices the user has at any one time 
3. Give the user only meaningful choices 
4. Make sure the user knows what to do at every moment 
5. Focus the user’s attention on the task at hand 
6. Use the most efficient manner of user input 
7. Make the user aware that the program is waiting 
8. Pause, quit or move on without the user’s response if it doesn’t come soon enough.

Creating the Illusion of Awareness

Specifically Respond with Human Intelligence and Emotion to: 

1. The user’s actions 
2. The user’s inactions 
3. The user’s past actions
4. A series of the user’s actions
5. The actual time and space that the user is in
6. The comparison of different users’ situations and actions

Maintaining the Illusion of Awareness

1. Use dialogue that conveys a sense of intimacy 
2. Make sure characters act appropriately while the user is interacting
3. Make sure dialogue never seems to repeat 
4. Be aware of the number of simultaneous users
5. Be aware of the gender of the users
6. Make sure the performance of dialogue is seamless
7. Avoid the presence of characters when user input cannot be evaluated

Designing Voice Experiences: Guiding Principles

by Lyndon Cerejo

• Onboard the user and help them get started
• Keep conversation exchanges brief to reduce cognitive load
• Examples work better than instructions
• Delight without interfering with the task
• Use explicit confirmations for important actions, and implicit for less risky
• Design for failure
• Respect the user’s privacy and security 

The Interaction Design Foundation’s Guidelines for Designing Voice User Interfaces

• Provide users with information about what they can do.
• Help users understand where they are in the system.
• Express intentions in examples.
• Limit the amount of information.
• Use visual feedback. 

NNG: Classic Usability Principles for Voice Interaction 

By Kathryn Whitenton

(The complete list of usability heuristics that these are taken from)

• Error prevention: systems should prevent errors from occurring in the first place.
• Flexibility and efficiency of use: allow for flexible use of the device (e.g. wherever they are);  make interactions efficient (e.g. avoid forcing the user to speak repetitive commands that they wouldn’t say in conversation).  

Voice User Interface Principles

by Stephen Gay 

1. Get to know each other: introduce yourself and establish communication preferences.
2. Be approachable: make it easy for users to start the conversation at any time.
3. Listen closely: show that you’re listening, through both words and actions.
4. Mind your manners: be responsive and socially sensitive in your interactions.
5. Talk like a native: convey meaning by what you say and how you say it.
6. Adapt your speaking style: anticipate what kind of conversation will suit the situation.

CX Partners: 7 reasons people aren’t using your voice UI

by Fabien Marry

1. Be clear about what it can do.
2. Don’t ignore privacy and social contexts.
3. Take care who is speaking.
4. Remember the context of conversation: what did the user just say?
5. Don’t rely on voice alone.
6. Consider the environmental audio (e.g. car, kitchen, music playing)

10 Best Practices when Designing for Voice

by Jess Williams

1. Manage users’ expectations.
2. There doesn’t need to be a hierarchy!
3. Consider the linguistics. A great voice app needs to cater for differences in the way people express themselves.
4. Keep Alexa’s responses short.
5. Don’t have too many steps in the conversation.
6. Try not to answer a question with a question.
7. Spend time on the edge cases and half happy paths.
8. Minimize choice.
9. Minimize pressure. Always give the user an option that buys them more deciding time.
10. AUDIO, AUDIO, AUDIO. The best skills on Alexa have audio content in them — a piece of audio that isn’t her voice.

Basic Guidelines for Successful Voice Design

by Cheryl Platz

1. Reincorporation is key. When the customer has given you data in their utterance, reincorporate it to confirm recognition.
2. GUI parity is not the goal. Speech interfaces are good at certain things: search, frequently repeated actions, sets of unique values. Speech interfaces are bad at screen-by-screen navigation and data-heavy interactions.
3. Brevity is the soul of voice UI. Every word of your response will increase the time your customer must spend listening. Be particularly strong-handed with edits on frequent responses.
4. Choose personality moments wisely. Only inset personality if you believe your customer has time to spare. Avoid in repetitive tasks. Best used in response to open-ended questions, i.e. “How are you?”
5. Use questions to guide multi-turn interactions. Don’t just open up the mic and hope for the best. If you don’t have enough information to act, give the customer a starting point in the form of a question to set them up for success.
6. Test drive your sample dialogs in audio form. Your TTS system may mispronounce common words or generate odd intonations, and your utterances may be awkward when spoken. If possible, generate audio comps with both sides of the conversation recorded.
7. Consider earcons, but use sparingly. Earcons (audio icons) can lead to more streamlined interactions, especially for repeated tasks. There are additional considerations like speaker quality, cohesion and acoustics that are best discussed with a sound designer.

Voice in technology

From An introduction to conversation design for UX designers | VUX World

Natural Conversation Framework: Pattern language

From Dr Bob Moore’s talk: Conversational UX