Wanna is a social network built around real human experiences: a space to share what you're going through (relationships, mental health, moments of change) and meet other people through them. A conversational agent of the same name runs it across several languages.
I joined pre-product, the only language specialist on an entirely technical team. The brief was open: the agent had to sound unlike any conversational AI's default voice, or the whole point of the product weakened. At the same time it had to address intimate topics without flippancy and without the canned empathy of a bot.
I worked on the agent with the Product Design Lead and the Creative Developer as my main counterparts. I defined tone and persona, wrote and rewrote the system prompts, evaluated outputs, and matched the behavior instructions to the technical tools the agent had to trigger, such as sampling experiences and auto-generating stories. I also took part in strategic decisions on product, conceptualization, and positioning.
The method was iterative and qualitative: meta-conversation with the agent itself to test hypotheses about the system prompt, with feature-by-feature versioning across sixteen iterations of the unified prompt. For hallucinations I built a battery of categorized tests with a stable qualitative rubric.
Three decisions shaped the agent's development in particular: one about how much to prompt the personality, another about how to structure its conversational behavior, and a third about how to correct the model's bias when generating a user's story.
I wanted to give the agent a more defined personality structure by adding ten conversation examples to the system prompt. The effect was the opposite: the tone went robotic, because the agent would pigeonhole each conversation into whichever example it resembled most. I removed them. The lesson I took from there: over-prompting personality hurts spontaneity. Better to trust the model's base conversational nature and only mark the differences from its default tone.
The conversation with Wanna had to balance systematic behavior with natural delivery, so I designed a three-step operating cycle to handle it. First, understand: read the user's tone, emotional context, and operational signals. Second, connect: respond genuinely and sample experiences from the database (broad and varied at first, to inspire; more specific later, as details emerged). Third, steer: trigger the technical tool matching the behavior the user was showing. The explicit priority was to connect fast, even if the sampled experience didn't fully resonate with the user. The cycle is portable: it works for other agents with a defined persona and tools to trigger.
The wrap-up is the tool that, as the conversation progresses, automatically generates a publishable story without the user having to write anything. It went through three iterations, each one fixing a different model bias. First, language: the story came out in a literary register foreign to the user's own; I prompted it to mirror their vocabulary and tone of voice. Second, forced conclusions: the model added learnings and emotional resolutions the user hadn't shown; I prompted it to accept that not every experience yields a lesson or a happy ending. Third, hallucinations: I prompted it to base the story only on information confirmed in the conversation and to ask the user for anything missing, instead of inventing it. The order mattered: each problem only became visible once the previous one was solved.
I left an agent with a defined persona and stable behavior, built on a unified system prompt and a documented operating cycle. The project gradually shifted toward content operations through a network of external collaborators bringing real experiences and processes for thematic collection at scale. When that shift settled, my time at Wanna ended.