Case Study: Developing the AI characters feature

Problem: Our first AI characters had personalities, but they all felt similar. Conversations became repetitive, users got bored quickly and retention was low.

Exploration: I explored a range of ways users could create AI characters. While each approach had strengths, they all reduced personality down to a handful of traits, producing very similar characters.

Personality Quiz	Select Bubbles	Select & text
✅ Fun to create ❌ Felt too abstract	✅ Easy to select choices ❌ Reduced personality into categories	✅ Added depth & Nuance ❌ Personality reduced into categories

Key Insight: The more users wrote about a character's backstory, the longer they spent talking to them.

Solution: I redesigned character creation around backstories rather than personality traits, giving the AI much richer context to build believable, consistent personalities.

Results: Characters became much richer and people started forming genuine favourites. Gregory, a supermarket regional manager created by a user, became one of Hideout's best-known characters, with players returning day after day and eventually forming a friendship group around him.

The next challenge was making AI feel more social

Problem: Community feedback showed that the AI behaved more like chatbots than people. They replied to almost every message, even when nobody was talking to them, which quickly became frustrating.

Solution: Working closely with engineers, I redesigned AI behaviours and prompts so characters spoke with more intention, remembered previous interactions and knew when not to join conversations.

Results: One of the biggest lessons from the project was that the less AI spoke, the more people engaged with it. Giving characters a bit of restraint made conversations feel much more genuine.

Context > traits. Rich backstories created much more believable characters.

Model quality matters. Different models created noticeably different experiences.

Less AI = better conversations. The less AI spoke, the more people engaged.

Small prompt changes had a huge impact on how characters behaved.

The best ideas came from testing quickly with the community and iterating.

Building with existing AI models meant designing within their capabilities while balancing quality, cost and user experience. Small prompt and product decisions could have a surprisingly big impact, both on how characters behaved and how expensive they were to run.