2026 · Enterprise SaaS · 6 Weeks
Natural language + agentic AI unlocks key gym staff workflows
Helping martial arts instructors confidently track student progress, schedule belt promotions, and focus on coaching over software.
Claude Prototype
AI Workshop
Web & Mobile
DISCOVERY RESEARCH
Uncovering Top Friction
Looking for our next initiative, we started gathering data sets.
Claude NPS dashboard
I had Claude create a dashboard of our last 3 quarters of NPS feedback (a high-response survey where users often take the opportunity to note their frustrations with the software).
Belt ranking and testing emerged as a consistent pain point, though the details were vague.
Follow-up survey
A survey dove deeper into the belt and promotions area to narrow in on where the highest friction was coming from.
KEY INSIGHT
Student progress tracking, promotion eligibility, and test registration emerges as top friction areas.
UNDERSTANDING
Contextualizing the Friction
User interviews unlocked our understanding of the problems, personas impacted, and the current user journey.
User interviews
Key workflows were identified that our product wasn't solving for—and it needed to.
Primary impacted persona
My user personas showed that while owners and front desk are affected, these problems center staff instructors.
Current screens (workarounds)
Interviews uncovered the following workarounds to identify student belt ranks and register them for tests:
DEFINE
Determining the Project
Turning research insights into a clear problem, a focused goal, and the opportunities worth exploring.
It was clear from the research:
The problem
Instructors are responsible for deciding when students are ready for promotion, but our software makes them piece together the information they need across individual profiles, reports, and promotional tests.
They have no efficient way to evaluate student progress at a group level, identify who is ready, or act on that information—turning a routine part of their job into a time-consuming administrative process.
The goal
Help staff quickly understand student progression, identify who need supports, and register eligible students for tests so they can focus on coaching.
How might we…
Statements that helped me frame the direction of the work.
How might we…
Reduce manual steps?
Decrease the number of steps and screens instructors need to identify student belt rank and progress at scale.
How might we…
Provide scaled analysis?
Instructors can identify students who may need attention based on factors like time at rank, attendance, or skill progression.
How might we…
Increase registration efficiency?
Reduce the time required to go from “Who is eligible for a test?” to “These students are registered for a test.”
*And add feature parity to our Staff Mobile App!
IDEATION
Concepts & "UI or AI" workshop
Quick wireframes for feedback
Fast concepts built with MagicPatterns helped me convey basic ideas for a new Ranks area of the platform. This sparked creativity, quickly going from where should we start to what could be better?
Reviewing designs led to a key question:
UI or AI? workshop
When AI possibilities felt limitless, this workshop helped draw the line between AI, human, and UI. It aligned on design principles, then mapped what the UI should handle and where AI could solve friction the UI alone couldn’t.
Design
Web and Native App Solutions
A new entry point
Users now have a dedicated space for their organized programs and rank structures within each. No more digging through individual profiles or building complex reports!
Ranks and student progress
Instructors can now view students by their belt rank, with key stats on their progression that help them make coaching decisions. AI actions are strategically placed throughout the screen where questions and actions are frequent.
Assistant coach (AI chat)
This is where prior long flows turn into one simple query. The AI chat has access to data on student profiles, skills, attendance, classes, and more to give Instructors answers where they need it most. Suggested prompts help the user learn what the AI can do, and provide common interaction starting points.
AI Interaction
Claude Code Prototype
Prototyping with Claude quickly surfaced areas we hadn't considered, such as response framing, timing, negative tones, and other parameters we then defined further below. Getting feedback and refining the interaction details was more streamlined than ever before.
AI definition and guardrails
Voice & Tone
Defining how it will speak to users
Scope & Boundaries
What information it has access to, and how it says "no"
Response Schema
Prioritizing brevity and defining response UI display
Guardrail Tests
Extra precaution to ensure the AI doesn't go off track
Results & Beta Testing
Some improvements were immediate and measurable from day one. Others will be tested over time.
Immediate impact
Reduced workflow complexity
20+ to 3 clicks
To schedule a test and register eligible students. (Compared to the current fastest path starting at report creation.)
Key information at a glance
Minutes to Seconds
To view students by their belt rank at scale, and identify eligibility.
Expanded coverage
Web & Mobile App
Providing a core staff workflow in the native apps for instructors on-the-go.
What we'll measure in a closed beta
8 studios · 12 weeks (to ensure multiple promotion cycles)
Feature adoption
70%+ weekly adoption
Users visit the new ranks experience at least once weekly, focused on repeat visitation.
Tracked via Pendo unique user visits
AI engagement
3 submissions per week
Average chat queries submitted per user, per week (in-page actions or manual chats).
Tracked via Pendo event trigger
AI test workflow completion
75%+ tests scheduled
Tests drafted via AI chat that result in a successfully scheduled test.
Tracked via Pendo funnels and event triggers
Customer satisfaction
4.2/5 avg score
Post-beta instructor survey using a 5-point satisfaction scale.
Tracked via CSAT survey, in-app Pendo Guide
AI response quality
85%+ helpful responses
Users rate the AI chat interaction helpful in a yes/no survey post-interaction.
Tracked via Pendo survey, triggered no more than once weekly
Qualitative experience evaluation
Open feedback
Conduct user interviews to gather more in-depth experience feedback.
Calls scheduled via personalized email outreach
What I Learned
There was a lot to this project, and it moved incredibly fast! Here are my personal takeaways.
Lesson 1
AI prototyping really helped define detailed interactions
Building a live prototype with Claude turned Figma screens into something I could actually use, and that exposed decisions a static mock never would. For example, the AI chat answers came back so fast they felt unearned, so I added a “thinking” state and loading cards. New chats felt more natural when the inquiry slid to the top of the window. Long cards needed to stop scrolling before their top went out of view. Iterating with realistic data revealed more edge cases. Instead of feeling stressed about ensuring I catch every detail, I felt more empowered to shape the experience than ever as a designer.
Lesson 2
AI parameters are now a core part of UX work
Much of the experience was in how the AI behaves: its tone, the words it never uses about students, how it says no to billing questions, the fact that it drafts instead of acting, and the typed response cards it must return. Those choices decide whether coaches trust it as much as any other screen. Designing the system prompt, guardrails, and response schema alongside the UI is now part of the job, and leaving this to engineering alone means leaving a big piece of the user experience undesigned.
Lesson 3
Keeping our users in control is key for trust, and usage
Early versions let the AI schedule tests and register students as soon as it was asked. Coaches didn’t just distrust it, they found it a little insulting. Deciding who’s ready for their next belt is the heart of their expertise, and an AI acting on that judgment felt like it was overstepping. So the AI’s role changed. It now prepares the work: who’s eligible, a suggested date, a drafted test. The coach makes the call, confirming in one tap, editing the full event, or undoing afterward. Users were much more willing to use it when they felt it support their expertise, not replace it.
