Designing for Conversational AI: Prompt Flows

First of all, I want to let you know that I'm back to writing and sharing content, so all your feedback is very welcome — without it, I don't even know if it's worth continuing!

Second, I'd like to thank everyone who has subscribed to this newsletter and let you know that the current address is uxsnack.com (uxsnack.design will no longer be linked to this publication).

Now then, let's go!

Designing for conversational AI opens up new challenges and opportunities to experiment with innovative methods. It's worth creating a separate post on how to start designing for conversational, but today we'll focus only on prompt flows.

We designers are familiar with user flows, right? Prompt flows are similar, but focused on dialogue examples.

Some reasons why you might want to use this method:

• No need to draw screens and mockups

• You can explore dialogues much more easily and quickly

• Macro view of the conversation, but also allows you to go into detail and see the branches

• Facilitates alignment between design, development, and product teams

But conversational is not deterministic

Even when we think about designing just the flow, we can fall into the trap of thinking we're designing, as before, for deterministic experiences — but in reality we can't think that way. Experiences with generative AI, such as conversational, are probabilistic. This means we don't know what the exact outcome of the interaction will be, as it's impossible to calculate all possible scenarios.

Still, this shouldn't stop you from exploring possible scenarios and example dialogues that help you not only prepare better for prompt engineering, but also create control guidelines (guardrails).

How to get started

• Start by identifying the main use case(s)

• Define the "baseline" of the first dialogue

• Explore possible branches and error scenarios

• Think about rich content, not just text (we'll talk about this in another post)

• Take the opportunity to think about what your AI Agent will need to do, or what data it will need to retrieve

• Consider accessibility: how will the agent respond to users with different needs?

• Test with diverse users to identify language patterns and cultural preferences

Prompt richness and intent

Capturing the user's intent is not always easy and may lead to "disambiguation". You can also try to capture strategies in prompt flows to handle these situations.

Simply put, "prompt richness" refers to the quality and clarity of the user's prompt — whether it can be well understood by the conversational agent, whether disambiguation is possible, or whether it's truly impossible to understand.

Examples of different levels of prompt richness:

• High: "I need a table for 4 people tomorrow at 8pm, we have a child who is allergic to gluten"

• Medium: "Table for tomorrow night"

• Low: "Dinner"

I'll talk about strategies and frameworks for getting better results with conversational AI agents in a separate post!

A communication and technical alignment tool

We used to use mockups and screens to discuss with technical teams, in addition to that being our main handoff. When designing for conversational, you need to go beyond that — prompt flows are an excellent communication tool for technical teams and can even serve as a test.

Tips for technical alignment:

• Document the types of data needed for each response

• Specify required APIs or integrations

• Define success metrics for each interaction

• Establish timeouts and fallbacks for error scenarios

• Consider rate limiting limitations and API costs

Once again, conversational is probabilistic — we don't know what will happen, but we should still anticipate some scenarios in order to instruct the LLMs and conversational agents according to what we consider to be the best experience. Prompt flows are allies in that mission.

How far to guide the conversational agent

We can get into an endless discussion about how far to "guide" conversational agents, since good human-AI conversations in terms of conversation also want to feel natural and intelligent. For that, we need to keep in mind that we must leave some freedom for the agent to respond in the best way — not just rigidly following our rules.

Considerations on diversity and inclusion:

• The agent should recognize and respect different forms of communication

• Avoid assuming genders, family relationships, or cultural contexts

• Provide alternatives when communication is unclear

• Take into account different levels of digital literacy

Testing and iteration

Recommendations for testing your prompt flows:

• Test with real users from different backgrounds

• Simulate stress scenarios (multiple requests, contradictory information)

• Validate with technical stakeholders the feasibility of the interactions

• Document failure patterns to improve guardrails

• Measure metrics such as resolution rate, satisfaction, and interaction time

As mentioned earlier, I'll have a dedicated post on this topic where we'll discuss some strategies such as the fallback ladder and other advanced techniques.