Alonso RosadoAI shipped product · Mobile
BJJ training apps are either bloated with competition stats or too generic to reflect how people actually train. I built a lightweight alternative around a quick logging flow and an AI coach that remembers previous weeks, then tested it with a small group of practitioners to see what actually held up.
Tatami
Live app · Log form in guest mode
Constraint
Hobbyists have the most retention potential and the least tolerance for friction. The coaching layer also needs rich data to work.
Competitors may be more willing to pay, but hobbyists significantly outnumber them and may be the stickier audience. They are also the least tolerant of a logging flow that feels like work.
That created a real tension: keep the log fast enough to use after training while still gathering enough context for an AI coach to say something useful.
Decisions
Each decision traded a richer default for the harder choice that kept the log fast enough to actually get used.
I cut everything non-essential and collapsed the date field because most logs happen the same day. An AI coach is only as good as its data, and a form people skip produces worse data than a form they complete. Completion rate beats field richness.
Guest-mode walkthrough · Example session
Choose 60 minutes and Closed Guard, then Save. Skip the optional review and open Progress to see the session recorded. Reflections stay empty.
Instead of regenerating a stateless summary, I stored each week’s coaching recommendation and injected the last three as context. For hobbyists without a coach present, continuity is the trust mechanism. Without memory, it feels like a novelty. With it, it starts to feel like a relationship.
Every comparable app includes it. I kept it out of the training-room experience because turning practice into a scoreboard changes the social dynamic. The product’s job is to support the training room, not surveil it.
AI insight
The first coaching prompt sounded punitive. ‘Direct and honest’ reliably collapsed into authoritative and judgmental whenever the model had negative data to work with.
The fix was not softening the copy. I reframed the persona from a coach reviewing performance to an experienced training partner sharing an observation.
“LLM tone is not a dial between formal and casual. It is a function of the role the model believes it is playing.”
Final product
Choose a duration and techniques, then save. Add reflections when you have more to say.
See training consistency without turning the gym into a scoreboard.
Use saved recommendations from previous weeks as context for the next coaching response.
Beta signal
Initial beta feedback described core logging as taking under fifteen seconds. The coaching layer can refer to saved recommendations from previous weeks, while optional reflections add context about the session. The design tradeoff is keeping the log quick without losing the detail that helps make coaching useful.
Next time
I’d test a separate post-session reflection prompt, such as a notification two hours later, so adding context doesn’t compete with the quick logging flow.