- Company
- Foreverlearning
- Timeline
- Jan 2026 β Present
- Role
- Founding Product Designer
- Team
- Founder, AI director, engineers, marketing, and me
- Contribution
- Research, product strategy, rapid prototyping in Figma and Claude Code, design systems, go-to-market screens
Results
- Real coaching sessions we studied before writing a single prompt
- 22Real coaching sessions we studied before writing a single prompt
- From a trainer's public profile to a working twin
- 48 hrsFrom a trainer's public profile to a working twin
- Trainers signed up across 5 countries, and counting
- 50+Trainers signed up across 5 countries, and counting
Overview
TrainerTwin builds a working copy of a trainer: one that probes, coaches and debriefs a learner the way they actually would, so their time can reach more people than their calendar ever could.
I'm the founding product designer here, in the room since the first rough prototype. TrainerTwin grew out of what we learned building Intervoo, our first product, which has its own case study.
Why TrainerTwin
Every coach is a Drona
The gap we found while building Intervoo, good coaching that can't scale, is the whole reason TrainerTwin exists. There's an old story that explains it better than any pitch deck: Eklavya couldn't get into Dronacharya's gurukul, so he built a clay model of his guru and practiced in front of it until he became a great archer anyway.
We build the talking clay models. Every great coach is a Drona, reachable by only a handful of Arjunas who can afford the seat or the fee. We want more Eklavyas to reach them, and the Dronas to be better off for sharing what they know, not worse.
A twin isn't bookish knowledge, frozen at the moment we built it. People keep learning, so their twin has to as well.
We never claimed the twin would be right from day one. Trust gets built the same way it does with a person: a little, over time, through feedback.
Studying the coach, not just the content
Before we wrote a single prompt, we needed to understand how a trainer actually decides what to say next, not just their catchphrases. So we sat through 22 real coaching sessions, more than ten hours of tape, and broke every exchange into three parts.
Those sessions explained the trainer, but only for the learners who happened to be in the room. So we built a matrix of learner types instead: confident but shallow, sharp but inarticulate, evasive, honest about what they didn't know. We tested the twin against all of them before it ever talked to someone real.
Observation
What did the learner just say, and what did it reveal, skip past, or quietly get wrong?
Decision
Does the trainer push for evidence, drop a hint, teach outright, or let it go for now?
Behaviour
How does that decision actually land: the tone, the phrasing, the thing they say right before the hard question?
Before the screens
The high-level journey
Before jumping into screens, we mapped the whole journey at a high level: what our team builds as the twin for a coach, and what the coach gets out of it. It starts with their account and tools, the knowledge and context the twin needs, and the trainer setup itself: coaching style, persona and the learners they work with.
On the left, we do that setup for the coach. On the right it's DIY: the coach sets up the roleplay and the agent, with the scenario, the twin's instructions and the goals to evaluate learners against, then saves it to a library, tests it and assigns it to a learner. Agreeing on this before any solution screens meant we all knew what a coach puts in and what outcomes come out.
Figma for structure, Claude Code for speed
Most of my week still goes into research and sitting in on sessions, not Figma. Once I know what I'm building, I block out the structure there, then move into Claude Code, wired up with Codex and Paper over MCP, to get something real in front of the team the same day.
Most mornings start with a new prototype. By afternoon it's been through engineering, our AI director, the founder and marketing, and it's usually been rebuilt at least once before the week is out.
Two ways to prototype the same screen
Here's one flow, built two ways: one shaped first in Figma, one built straight in code. More of our early exploration happens in the second one lately. It's faster to get wrong fast, which is really what a prototype is for.
Putting it together
The complete flow, end to end
Once we'd arrived at the complete flow, I took it into Claude Code and generated the whole thing as one end-to-end prototype of the twin and roleplay product: the builder, calibration, the learner experience, assigning scenarios and analytics. Everything in one place.
The journey runs from the trainer setting up and improving their twin, to sharing it, to learners using it, to the trainer reviewing what happened.
What trainers actually asked for
We ran discovery calls with over 50 trainers across 5 countries, and the pattern was loud: nobody wanted an AI twin for its own sake. They wanted to cut their camera time, offer more practice between live sessions, and reach learners they couldn't get to in person.
A couple of things we got wrong, and fixed: our sign-up page read like a manifesto: too much AI jargon, not enough "what does this actually do for me." And the original form only asked two questions, so job-seekers and future-trainer hopefuls were applying right alongside real coaches. Both are fixed now.
Told us how important Socratic-method adherence is in the twin we build: not just answers, better questions.
Pushed us to make the twin usable for sales roleplay that can be measured against a real framework, not just "felt" better.
Built in 48 hours, checked by hand
We built a pipeline that turns a trainer's public profile and training material into a working twin in about two days. Most of that time is us studying their style closely enough to verify it actually sounds like them before a learner ever talks to it.
Three trainers are live with their twins right now. More than ten are in the queue, waiting on us.
What we're actually measuring
A fluent conversation isn't the same as a good one. We learned that the hard way when early tests sounded right but made the wrong call. So we track four things separately.
Underneath all of it sits one quieter metric: how long a learner waits before the twin responds. In a voice conversation, a second of silence feels like ten, so latency isn't an infrastructure footnote here. It's part of the experience we're designing.
- Learner fidelity: did the simulated learner stay in character?
- Decision fidelity: did the twin make a trainer-like call?
- Behaviour fidelity: did it sound like the trainer when it did?
- Conversation quality: was the session actually useful?
Prototype
To experience the prototype, go here.
