CalorieFlow started as a personal problem: every calorie tracker I tried made logging feel like data entry. So I designed and built my own — an AI-first tracker where you describe your meal in plain language and the assistant estimates calories and macros. I designed the product end to end and built it by directing an AI coding agent through structured, constraint-driven prompts.

With the MVP live, I audited it like a client's product. The critical findings weren't visual — they were failures of data integrity. For a product whose entire value is numbers, that's where trust lives or dies.

Activity counted twice. Maintenance calories used a multiplier that already included exercise, yet logged workouts also added calories back. I redesigned the model: the multiplier now covers lifestyle movement only; workouts are logged separately on top. The fix was mostly microcopy — the option labels could no longer mention exercise at all.

The projection lied politely. Forecasts assumed constant metabolism, but BMR drops as weight drops — so linear projections are systematically optimistic. I specified an iterative model that recomputes maintenance from projected weight at every step, verified with a 120→70 kg edge case where loss must visibly slow near the goal.

No guardrails. The app calculated a healthy BMI range, then silently accepted goals below it — even labeling the plan "RECOMMENDED." I added inline warnings for out-of-range goals, a 1,200 kcal/day floor, and a mirrored surplus cap for weight-gain mode. Warnings inform, never block. For a health product this is ethics, not polish.

The AI chat is the core input — and the easiest place to lose trust. Three rules shaped it:

Show assumptions. "Muffin — 350 kcal" is a guess in a lab coat (a real muffin is 150–600). The estimate card states its portion assumption ("1 medium, ~110 g").

Make outputs editable. Calories and macros are tap-to-edit before logging — not just the meal category, the one field that didn't matter.

Make actions reversible. Logging shows the exact effect ("Added to Lunch · +350 kcal") with a 5-second Undo.

My real deliverables were prompts: scoped specs with explicit constraints and verification criteria. Method that emerged:

Constraints first — every prompt opens with what the agent must not do, or it helpfully redesigns everything it touches. Audit before fix — "report findings" and "fix only what the audit shows" are separate steps. Verify the render, not the report — one bug survived two claimed "fixes" until I required printed evidence at each step of the data path: database → API → chart props. Kill duplicated logic — the plan calculator diverged from the recommendation three times, always because two copies of the same formula drifted; the durable fix was one shared projection function.

The audit surfaced 20+ issues; four broke the core model's math — all fixed and live. The app now handles loss, gain, and maintenance with mode-aware math, copy, and guardrails. What I'd do differently: define the calculation model before building UI around it — every expensive bug traced back to math being an afterthought. Not yet validated: real users. Next step is a usability round on the estimate flow — do people actually use editability and undo, or trust the AI by default?