The first thing we decided about Coach Niramai was that she should never sound like a textbook. That ruled out every off-the-shelf persona template we looked at. The second thing: she should never make a student feel stupid. That ruled out blunt correction models.
The persona brief
We gave Niramai three constraints: warm but not sycophantic, direct but not blunt, patient but not passive. Those constraints sound obvious until you try to operationalise them in a system prompt and test them against 500 edge-case inputs.
Her name came from a Thai word root meaning "good guidance." The choice was deliberate — names in Thai signal respect relationships, and we wanted students to feel they were working with someone who understood their world.
The feedback architecture
Niramai's feedback has three layers:
- Prosodic awareness — rhythm and stress pattern, flagged first. Students hear the pattern, not just the rule.
- Lexical choice — suggested alternatives offered as options, not corrections. "You could also say…" not "That's wrong."
- Grammar note — only surfaced after the student has had a chance to self-correct. Minimises shame avoidance.
How the 40 mechanics connect
Each of our 40 retention mechanics is tagged to a tonal tier — emotional register the student should be in for that mechanic to land. Niramai's role is to read the tier and modulate her warmth accordingly. A student deep in a streak gets a different energy than one recovering from a missed session. The model knows which context it's in via the session payload.
The result: students in the Q2 cohort described Niramai as "a friend who happens to be a teacher" — which was exactly the emotional target.