Adaptive AI trainer for a coaching practice
A coach-facing workflow for drafting and reviewing weekly plans from structured activity records, with explicit escalation boundaries.
What changed. Weekly-plan drafts and activity records are collected for coach review; roster capacity, time saving and health outcomes are not established here.
Stack. structured client state · weekly plan generation · load adaptation · coach dashboard
Context
A coaching workflow needs to reconcile the planned week with the activities a participant actually completed. Reusing a static calendar can leave missed sessions or a changed condition out of the next draft. The coach remains responsible for deciding whether a proposed plan is appropriate.
Task
Organise the activity record, draft the next week and present exceptions to the coach. The system should make the relevant state reviewable without implying that a generated plan is medical clearance or a substitute for professional assessment.
What we built
The workflow separates structured state, a weekly-plan draft, activity logging and a coach-facing summary. The next draft uses the recorded state instead of assuming every planned session happened. The coach can review exceptions alongside the proposed changes.
Pain or a reported injury is an escalation condition to validate in the workflow, not an invitation for the model to invent an alternative exercise or treatment. The intended stop boundary needs explicit acceptance checks and a responsible human decision. A prompt instruction alone does not prove that the boundary is enforced.
What changed
The design brings plan drafting, recorded activity and exceptions into one review path. This case does not establish a supported client-roster size, a measured reduction in drafting time or a health outcome. Those conclusions would require dated observations, a sample and an appropriate assessment method.
What we would do differently
Define the stop-and-review behaviour before generating more detailed plans. Test missing state, conflicting entries and an escalation fixture with synthetic data. Verify that the final interface actually exposes the exception to the coach rather than merely storing a flag.
Related service: AI consultant and chatbots