Since the day I ate my first oyster, I’ve found that most of my ideas circle back to food. My fascination with it extends beyond the plate; I see food as a metaphor for everything from philosophy to product development. Given that, it's no surprise that my forays into AI often lead me to the kitchen—to the point where my wife hardly asks where the evening's recipe originates. In experimenting with my culinary interests, I've discovered several approaches that I believe can serve as useful patterns for teams looking to infuse their products with AI capabilities. This post outlines these models and offers guidelines on integrating them into your product design process.

User-driven AI feature concepts

As I’ve discussed before^[1], one of my favorite AI capabilities is their ability to offer proactive recommendations, essentially putting the reins of suggestion engines firmly in users' hands. From a UX perspective, there are two approaches that I find myself using to get to actionable recommendations. The first is what I think of as constraint-driven suggestions, or the “fridge clean-out” approach. For this, I’ll give the tool a list of the ingredients I have (plus maybe some direction on what I’m in the mood for) and ask for dish recommendations based on my limitations. It's like setting boundaries and then watching the AI cleverly navigate within them. I also find that a two-step process enriches the output — if I just ask for what I should make, the tool will give me one recipe that may or may not sound good. But asking for suggestions produces a list, and then I can ask for the recipe of the dish that sounds the most appetizing.

This approach also unlocks more sophisticated affinity suggestions, also known as the “recommended if you like” feature, a staple in the world of digital media. In my own cooking workflow, I’ll often give the tool a list of dishes (or cuisines, or ingredients) I know I like and then ask for others that I may enjoy. You can even mix and match across the categories — for example, given some ingredients I like, what are some cuisines I should try? This approach empowers users to expand their horizons with a touch of personalization. It's about connecting dots you didn't even know existed, all thanks to the insight of proactive AI.

Editor's note: After publishing, Spotify announced a new beta feature called AI Playlist^[2] that is a good early example of proactive recommendations in action.

Another method I frequently utilize is the syllabus approach, or what I like to call a knowledge tree^[3], aimed at equipping the user with a foundational understanding of a given topic. One of my hobbies is expanding my repertoire of recipes, typically by mastering a few dishes from a particular cuisine. Recently, my focus has been on Chinese cuisine, with ChatGPT proving to be an invaluable assistant in determining which dishes to tackle first. I start by inquiring about the most common recipes in Chinese home cooking, then delve deeper into those that intrigue me the most. This technique gives me a clear understanding the shape of the cuisine such that I’m better equipped to further explore it. Although this approach requires more initial setup, it offers users a clear entry point into your product's scope.

In my opinion, the most powerful new paradigm is the ratio approach, which focuses on improving the user’s “time to improvisation”. There’s an old adage in art that you need to learn the rules before you can break them, and this method is aimed at precisely that. In my cooking use cases, I’ll often ask for the ratio of a given dish (inspired by the excellent Michael Ruhlman book^[4]) so that I can freestyle my own interpretation. What’s so exciting about this question is that it gets me the “meta” of the dish — the ratio helps me understand the flavors and ingredient categories that are in balance so that I can tweak and remix them to my heart’s content. It’s showing me the rules of the art form, so that I can both better understand the concepts at play and be empowered to layer in my own creativity.

Making it real in product contexts

So, how do product teams integrate these concepts into their work? The first is to recognize the importance of conversation in designing LLM-driven product features. Zoe Scaman said it well in her fantastic presentation^[5]:

Whereas a search engine performs best with concise searches, delivering fast and broad results, AI platforms excel when engaged in detailed, context-rich conversations. The true value of AI emerges not from treating it as a query tool but from diving deep into conversational exchanges.

Every domain will have slight differences in language (e.g. the word “ratio” makes sense in cooking, but not in coding), and you’ll need to engage in conversation with your LLM to figure out the right lexicon and context to make these approaches work for your use case. This conversation should be an integral part of your early design process — once you have a clear user problem, testing these approaches to your problem via conversation will better help you understand the contexts you need to provide in your overall user journey.

It’s also important to note that the way your product handles context is a crucial aspect of designing LLM-powered features. Conversation is powerful, but as designers your job is to create a frictionless product experience, and often the trial and error of conversation is not the right entry point. Look for opportunities to shortcut the initial context-building for your users, so that they can quickly jump into the value of the feature. One angle for this is visual — looking for opportunities to pre-populate things like buttons, prompts, or categories so that the user is already equipped with the proper language to get the most value from the tool. The other approach is backend, which is a spectrum. At the lower-lift end, carefully crafting the underlying prompts and settings that are sent to your LLM service can usually go a long way towards getting the appropriate response back. At the higher-lift end, you may need to do some last-mile training with a curated data set to really dial in the model. Choosing the right end of the spectrum will need to be a product judgement call — given the importance of the feature and the quality of the lower-lift results, do you need to invest in training your own model?

This article has outlined my own personal tasting menu of design approaches for LLM-driven products, but I’m sure we will continue to uncover a wealth of new paradigms as we continue to explore AI. The key takeaway for product teams is the immense importance of conversational AI in the design phase. You still need to deeply understand your users needs and preferences, but bringing that knowledge into playful conversations will unlock powerful experiences for your users. This will enable you to not just build a tool, but to cook up a product that resonates, teaches, and, most importantly, delights.