-
-
Sculpt
-
Clear Conversation View
-
Custom Coaches
-
Chatting
-
Takeaways
-
Memory
-
Profile
Inspiration
I've been working with an AI companion for over a year that completely changed how I approach my goals. Not through advice or therapy - through building actual systems. Invisible tracking of commitments, pattern recognition that calls out my excuses. When I saw Simon's brief about mobile AI coaching, I realized: this accountability architecture could work for anyone trying to level up. But most AI coaching tools felt like desktop productivity apps - complex, clunky, not something you'd actually use throughout your day. Simon's brief hit different: a beautiful mobile app where users could create and share their own coaches. That's when it clicked - what if the accountability systems I experienced could work for ANY coach someone wanted to create? Fitness, finance, relationships, whatever. And what if anyone could build one?
What it does
Sculpt is a mobile AI coaching platform that builds systems with you, not just gives advice. Each coach tracks your commitments through invisible behavioral tags, creates calendar checkpoints for follow-through, and generates session insights so you can see your progress over time.
How I built it
Built entirely in Swift/SwiftUI with CloudKit for the backend. The core innovation is the RAG memory system - coaches use invisible XML tags to track user behavior and commitments without cluttering the chat interface. All 10 coaches share a unified user profile that updates in real-time, so context persists seamlessly across conversations. Calendar integration uses native iOS APIs to create actual accountability checkpoints, not just reminders. The marketplace runs on CloudKit with upvote systems and personalized recommendations based on onboarding. Monetization via RevenueCat with 7-day trial and credit-based pricing. 80+ commits and 200+ hours compressed into 30 days of focused building.
Challenges we ran into
Memory gossip - Coaches were pulling old context and leading with it instead of responding to current messages. Had to rebuild the prompt architecture with stronger 'don't gossip' rules and cleaner separation between base system (format/tools) and coach personality (expertise/tone).
Domain sovereignty bugs - Fitness coach was refusing to give workout routines because I accidentally wrote liability guardrails that told it NOT to prescribe exercises (literally its core job). Had to rewrite prompts to balance safety with actual coaching utility.
Notifications completely broken - Didn't notice for 4-5 days that push notifications weren't working, which killed the accountability positioning since calendar integration is a key differentiator. Had to debug CloudKit + iOS notification permissions right before deadline.
Onboarding complexity - Initially tried building therapy-style deep intake flows, but realized Simon's professional audience needs fast/clean onboarding that gets them to value quickly, not 20 questions about their childhood.
Accomplishments that we're proud of
The invisible tracking system - Coaches build concrete accountability without users seeing any XML tags or technical implementation. It just feels like natural conversation that remembers and follows up.
Cross-coach memory that actually works - Your fitness coach knows about the work deadline your productivity coach is tracking. Context flows seamlessly across the entire platform.
Shipped a full marketplace in one sprint - Built the entire creator economy layer (coach creation, upvoting, personalized recommendations) in ~2 hours when I realized it was critical for Simon's vision.
Mobile-first UX that doesn't feel like a chatbot - WhatsApp-style conversation lists, clean coach profiles with Greek statue icons, haptic feedback, session insights. It feels like a real product, not a hackathon demo.
10 production-ready coaches - Each with distinct personality, domain expertise, and coaching approach. Not generic variations of the same prompt.
What we learned
Prompt architecture matters more than perfect prompts - Separating base system (tools/format/memory) from coach personality (expertise/tone) made everything cleaner. Principle-based prompts > scripted behavioral rules.
Mobile changes everything - Desktop AI coaching feels like work. Mobile coaching fits into your actual day - at the gym, commuting, making decisions in real-time. The UX completely changes the use case.
Community/marketplace isn't optional - Initially focused just on the 10 premium coaches, but realized the creator economy layer is what makes the platform scalable and engaging long-term.
Accountability requires real integration - Calendar checkpoints and session insights aren't just features, they're what separate actual coaching from chat. Without them, it's just another AI conversation.
Ship fast, validate with real users - Had a friend test comprehensively throughout the build. Caught bugs and UX issues I would've missed building in isolation.
What's next for Sculpt: AI Coaching
Creator monetization - Let coach creators earn revenue share when people use their coaches. Build a real marketplace economy, not just a sharing platform.
Advanced behavioral tracking - Deeper pattern recognition across sessions. Notice when users are stuck in loops, regressing on commitments, or avoiding specific topics.
Team/group coaching - Shared accountability with friends, couples coaching, manager/report dynamics. The memory system could track multiple people and their relationships.
Integration ecosystem - Connect with fitness trackers, calendar apps, habit tracking tools. Pull real data instead of relying on self-reporting.
Voice-first experience - Most coaching happens in-the-moment. Voice input/output makes it feel less like texting and more like having a real coach in your pocket.
Prompt refinement with Simon - As he mentioned in the brief, 'we'll fix prompts together if you win.' The architecture is solid, now it's about tuning each coach's personality and approach based on real user feedback.
Log in or sign up for Devpost to join the conversation.