Understand the training request and its context.
FORM / INDEPENDENT AI PROJECT
An AI training agent built around human control.
A working agent that adapts strength-training plans to real life, while keeping the person using it in charge of every change.
THE PRODUCT
Train with a plan that can change.
When schedules shift or recovery changes, Form helps rethink a session without starting a program from scratch.
- My role
- Independent designer and builder
- Scope
- Concept, UX, interface design, implementation, iteration and testing
- Tools
- Claude Code, AI agent tools and code-based prototyping
- Status
- Working prototype, not a public production launch
THE PROBLEM
Training plans are static. Real life isn't.
Most strength programs assume each planned session will happen as written. In reality, people miss workouts, need more recovery, face changing schedules and work around personal constraints.
Editing a plan manually means balancing exercises, session length and the rest of the week. I wanted to explore whether an agent could help without making decisions on the user's behalf.
How might an AI agent turn an ambiguous request into a useful, reviewable training change that respects the rest of the plan?
THE WORKING EXPERIENCE
From request to actionable change.
The agent interprets a request, retrieves the current plan and preferences, proposes a targeted adjustment, validates it against explicit rules and asks for approval.
THE AGENT WORKFLOW
Tool use, not just conversation.
Read the plan, profile, preferences and notes.
Suggest a focused change with reasoning.
Check explicit constraints and plan structure.
Let the user review, accept or dismiss.
DESIGN EVOLUTION
The first AI-generated interface wasn't the finished product.
I used Claude Code to build Form, then worked through the experience as a designer. The first version was functional, but its information hierarchy, comparisons and interaction feedback needed deliberate refinement.
WHAT I REFINED
Craft is in the decisions.
Hierarchy and consistency. Made the plan easier to scan and prioritized the current session over interface noise.
Transparent changes. Made proposed adjustments easier to compare with the existing plan before approval.
Accessibility and behavior. Improved interaction states, feedback and layout instead of accepting the initial generated output.
DESIGNING THE BOUNDARIES
Not every decision belongs to AI.
The model handles ambiguity. Code-enforced rules protect the structure of the plan. The user makes the final decision.
AI judgment
Interprets feedback, chooses relevant tools and explains proposed changes.
System rules
Checks exercise exclusions, session limits and scheduling conflicts before changes can be saved.
Human control
Reviews the before-and-after state, approves or dismisses, and can undo applied changes.
These checks protect the plan's structure. They are not a medical safety guarantee.
WHAT I LEARNED
Building faster doesn't remove design responsibility.
Claude Code helped me build a working agent, but its behavior required separate decisions about feedback, boundaries, persistence and reversibility. I had to evaluate what the interface communicated, not just whether the underlying code could produce an answer.
Form reinforced a principle I bring to AI product work: useful autonomy depends on clear responsibilities. A system can investigate and recommend, but people need to understand what changed and retain control over the result.
Form is an independent working prototype. This case study does not claim validated long-term fitness outcomes.
Explore more of my product work
Next project: Billing Assistant