01The problem
An AI that always agrees with you is the one that abandons you. That's the trap GodGPT had to avoid: it's an emotional-support LLM, and mainstream assistants are trained to please, so they validate and never resolve. It needed a voice that could hold a hard conversation and move someone forward, and earn enough trust for them to open up.
02The thinking
I designed against the easy behavior. The product is built around the conversations that matter, grief, doubt, spiralling, not the happy path. The interface does as much work as the model: paced slow, given space, nothing like a productivity tool, so it reads as a place to reflect. Trust here is earned on the hardest moments, so I designed those first.
03Key decisions & tradeoffs
Resonance over agreement
Responses that move you forwardnotAffirmations that keep you comfortable
ImpactAvg. session length: ~3× a generic assistant
A reflective surface
Calm, space-led chatnotA standard assistant layout
Design the hardest moments
Built for grief and doubtnotThe easy, happy path
04Driving the call
Designing an AI to not simply agree runs against every engagement instinct, and that was the argument I had to win. The easy path boosts the numbers that look good in a demo, so I pushed the team to treat resonance, not agreement, as the actual product, even where it meant the AI says the harder thing. I set the voice and the interaction principles the whole experience was built on.
05Outcomes
- A distinct voice: resonance over agreement
- A surface calmer than a standard assistant
- Patterns built for the hardest moments
06Reflection
A product that gives emotional guidance carries real responsibility, so the thing I'd invest in earlier is the guardrails: clear boundaries for when the AI should step back and point someone to a human. The voice was right, but safety has to be designed with the same care as the tone, not bolted on after.
