Paceback

I started Paceback after I developed runner’s knee. My physiotherapist explained exercises during our appointments, but I struggled to remember the plan and follow it each day. At the next visit, I could not easily recall what I had done. I needed a record I could use between appointments and bring back to the physiotherapist.

I designed and built Paceback, first as a place to plan and log my exercises. My work covered research, product direction, design, and development of the web and iOS app. The focus grew to the relationship between runners and professionals. That raised a design question: when should AI explain, ask for direction, or prepare a change, and how should that conversation work beside the plan?

I kept the AI control in the top-left corner throughout the app, so runners always know where to find it. On desktop, it opens into a sidebar beside the current screen. They can ask for help with the plan or record they are viewing without leaving it.

The first prototype was a rehabilitation log in December 2024. I later designed and built weekly AI suggestions. The flow used activity logs, progress, journal notes, and plan adherence to suggest increasing, reducing, or maintaining the plan. In my own use, it worked reasonably well for a long time.

Conversations with physiotherapists made me reconsider recommendations based on a runner’s health. I wanted this product to support their decisions. I also wanted to avoid taking it into medical-device scope. I reworked the app so physiotherapists and coaches can manage client plans and progress, while runners can import a plan, follow it, and ask questions about their records.

For injured runners, AI provides information and insights for the runner and professional to use. Runners who confirm they are healthy have a separate path with training plans and recommendations. The earlier automatic weekly flow is no longer running. The later design uses conversation to explain records and prepare allowed actions for review.

I defined three response patterns for the core chat. The response depends on the request and the plan’s rules. Conversation sits beside the schedule, so a request can lead to a visible change in the product.

• Answer: explain the plan or recorded activity.
• Ask: seek human direction when the request involves a health-related decision.
• Draft: show an allowed plan change for review before it affects the schedule.

I put the proposed change in the schedule so the runner can see its effect. Apply changes the active plan; Dismiss keeps it as it is. In this example, a request moves the long run to Sunday while keeping the weekly distance at 18 km.

I treated an answer, a draft, and an applied change as separate states. The agent can prepare an allowed change, but the runner decides whether to apply it. This schedule example does not show AI making a recommendation in response to pain or injury.

A record of completed activity gives the runner and professional something concrete to discuss at the next appointment. I used text to explain activity and charts to compare periods. Chart values come from recorded data; AI writes the explanation. Missing information stays unknown. A source sheet names the records and dates behind the report, which can be saved outside the chat.

A runner may already have a plan from a physiotherapist or coach. I used AI to help read a PDF, image, or text file so the runner does not have to enter every session manually. The imported plan can then be checked against the original. The example shows a returning user reviewing one plan at a time, with Apply and Dismiss beside the schedule.

I designed separate flows for planned sessions and recorded activity. Strava, Apple Health, and Withings bring activity and health data into Paceback. Planned sessions go out to calendar apps through a subscription link. A separate action sends each supported workout to Apple Watch. The aim was to keep the plan and its records connected to the tools runners already use.

Calendar subscriptions and Apple Watch export were implemented as preview features. Apple Health and watch export require the iOS app. Watch export is a per-session action, so sending a workout remains an explicit choice.

Goals and preferences can help the app respond, but a passing comment should not become a lasting assumption. I added Remember and Not now to memory suggestions. The runner can inspect, edit, or remove saved goals, preferences, and limitations in settings.

I took Paceback from a personal exercise log through design, development, and an iOS App Store release. The iOS app is now unpublished from the store, and development is paused. The web demo remains available. The main change was in its purpose: helping runners and professionals work from the same plans and records.

My own use and feedback from runners and physiotherapists shaped the product. I would next test whether people can explain a proposed change, apply or dismiss it, identify a report’s source, and use the record in a conversation with a professional.

Listening to physiotherapists changed the role I gave AI. Building the app let me carry that decision into interaction patterns: answer a question, ask for direction, or prepare a draft. Plan review, report sources, memory controls, and import review all make the agent’s work visible and open to correction.