Clara Moves

Testing the brief

Arthritis NSW knew that moving more would greatly improve their members' quality of life. The brief was a progressive programme, with activities that step up as she improves. I built a version to test it before anyone committed. She chose every step, and I believed it could be made safe.

Staying outside medical-device rules was a requirement from the start. Once there was a first plan, the client sought formal regulatory guidance to be sure the programme would be safe and effective. I checked the design against it. Even with her making every choice, progression could look like prescribing movement. More safeguards wouldn't change that, so I looked for another way to reach their goal.

The new approach

The client's exercise physiologists have a deep understanding of their members, which showed us that confidence and the beliefs around it were the main things holding them back.

We then focused on the barriers they describe: fear of making their condition worse, pain, fatigue, or well-intentioned family telling them to rest.

The programme runs on what she tells Clara. At onboarding, she's asked "How is movement feeling today?" and what is stopping her. After that, it checks in regularly to ask whether moving is getting easier, harder or staying the same.

Her answers decide what Clara does next: celebrate progress, where trying counts, or help her break a barrier down or see it differently.

I researched behaviour-change techniques and mapped them to each barrier. If she fears making her arthritis worse, Clara offers a trustworthy article that corrects the myth. Then it suggests a small, safe movement and asks if she'd like to try. Every movement has a smaller version and a point to stop. She decides each time, and Clara never advises on symptoms or judges her effort.

This became the preferred approach.

Measuring progress

The client needed to know the programme was working for their members. They wanted an inactivity label at the start and end, and a scored questionnaire every fortnight. They also asked me to drop the barriers question, concerned it might steer her towards activities involving her sore areas.

I made the case for a different approach. Precise scoring could weaken the regulatory position. It would also turn a supportive experience into a judging one, and a reluctant woman might stop. And Clara needed the barriers answer to know how to respond.

I kept it. In place of the label, the feeling-today question lets her answer against her own normal. "Nothing feels realistic today" is one answer.

The cost was the clean baseline, the clearest evidence of whether it worked. Relative change is harder to report. In return, she can answer honestly on a bad day.

Rules for an AI coach

Clara's words are generated, so nobody can script them. I specified it as rules: what it may know about her, what it may offer, and what it must never say or suggest. It never advises on symptoms or judges her effort, and points her to the client's material or her GP when a question needs clinical judgement.

The rules lived in one set of master files. I used an AI-assisted setup to build a version for each audience. The client's clinical advisers reviewed the reasoning throughout.

One shared document would have been simpler. Separate versions meant more upkeep, and they held when the programme changed direction.

I worked with the engineer to strip the design and clinical rationale down to rules the AI follows and specs he could test.