# Inspiration: Macro tracking is tedious, and most apps treat AI as an afterthought. We wanted to build something where Claude is present at every step — not bolted on.
How we built it:
Cloach uses Claude in four distinct ways: tool use with a strict JSON schema to generate personalized macro targets, Vision API to turn a food photo into per-item macro estimates, context injection to give the companion chat full awareness of the user's food log and weight history, and agentic tool calls (add_food, update_entry, delete_entry) so users can manage their log with plain language.
# Challenges: Keeping the injected context lean without losing depth, calibrating Vision confidence for ambiguous portions, and tuning tool schemas so Claude calls delete_entry instead of zeroing out an entry.
# What we learned: Claude's tool schema is a contract, not a suggestion — that reliability is what made natural-language CRUD safe. Rebuilding the system prompt as a live state snapshot on every message gives Claude session-level coherence without any memory infrastructure.
Log in or sign up for Devpost to join the conversation.