81 unique personas.
Three role engines and four decision layers, for a platform that needed many distinct AI companions that each stay themselves.
- The problem
- Design a platform that generates many distinct companions, from coaches to friends to teachers, without them blurring together or drifting over long conversations.
- Client · Role
- Client: EmotionMachine.ai · My role: Persona architect, consulting
- Led
- Persona architecture · three role engines · drift prevention · prompt compression · scored evaluation checklists
- Result
- 81 stable persona variants from one architecture, each engine with its own validation checklist.
A system for creating a cast.
EmotionMachine needed a way to make diverse characters that felt distinct. I designed the architecture that generates them, and the tests that prove each one holds.
Long persona descriptions don't survive contact.
Write a persona as a paragraph and it fades. The model compresses it, repeats itself and drifts toward a generic assistant. Multiply that across dozens of personas and they all start to sound the same.
Don't describe a personality. Give it rules it can follow.
Each decision multiplies the character.
Role, then voice, then interaction style, then mode emphasis: 3 → 9 → 27 → 81.
Role type: Coach, Friend, Teacher · Voice cluster: e.g. Strategic, Reflective, Performance coach · Interaction style: Collaborative, Directive, Co-creative · Mode emphasis: Insight, Exploration, Clarification
Three engines, one rulebook.
Coaching engine
a warm, non-directive partner that leads with reflection and stays steady when cut down.
Friend engine
a voice-led identity with yes-and co-creation; its personality survives compression.
Teacher engine
level checks, structured explanation and recall; clarity comes before flair.
The variants differ in real behaviour: authority vs partnership, expressiveness vs precision, how many questions, emotional energy, mode frequency and turn pacing.
Compression as a stress test.
I cut every persona down to the fewest words it needed. If it still sounded like itself, the architecture was sound. Each engine came with a scored evaluation checklist, so persona QA became a test rather than a matter of taste.
- priority order
- mode boundaries
- turn structure limits
- repetition rules
- consolidated safety blocks
"…modular prompts and evaluation scaffolding meaningfully informed our platform."