StudyMate

Zevi Arnawitz was traveling in Japan with his wife when Claude 3.5 Sonnet was released. Watching YouTube videos of developers building apps with Bolt and Lovable felt like a lightbulb moment: "it basically felt like someone came up to me and said, hey, Zevi, there's this cool new technology you should check out. You should really give it a try. Oh, and by the way, you have superpowers now." He had zero technical background—music in high school, not a tech unit in the Israeli army—but the promise of AI-powered development was irresistible. The moment he got home, he opened Cursor and began building.

Zevi started with Bolt and Lovable but quickly realized these tools, while excellent for beginners, were too opinionated for more complex features. When he tried to integrate payments, he "kind of started losing it." He graduated to Cursor with Claude Code, gaining full control over architecture and decision-making. He developed a sophisticated workflow that mimics working with a CTO: first creating a ChatGPT project with a custom system prompt that positioned an AI as a technical co-founder, then evolving that into a series of slash commands within Cursor—reusable prompts for each phase of development.

The workflow includes: `/create_issue` (quickly capture bugs or ideas into Linear while mid-development), `/exploration_phase` (deep technical understanding and clarifying questions), `/create_plan` (structured markdown plan with status tracking), `/execute_plan` (build the feature), `/review` (have Claude review its own code), and `/peer_review` (have multiple AI models review each other's work). Each model has distinct strengths: Claude as the communicative, opinionated CTO; GPT as the brilliant but silent coding expert; Gemini as the artsy, terrifying-to-watch but design-gifted innovator.

The peer review system proved game-changing. Instead of struggling to review code himself ("code is terrifying"), Zevi has multiple AI models critique each other's work, then runs `/peer_review` to have the dev lead model respond to criticism—defending decisions or fixing issues. This turned his biggest weakness (not being technical) into a strength: different models catch different bugs, creating a more robust review than any single reviewer would provide.

Zevi also discovered that iteration on prompts matters enormously. When Claude makes a mistake, he asks it to reflect on what in the system prompt caused it, then updates the tooling to prevent recurrence. This creates a virtuous cycle where the AI gets smarter with each feature. Documentation updates are critical—he's learned that updating slash commands and system prompts after every build prevents the same mistakes from happening again.

StudyMate launched with multiple-choice quizzes where students upload PDFs, select pages, set difficulty levels, and receive AI-generated quizzes with deep explanations. During this conversation, Zevi live-built a new feature: fill-in-the-blank questions with drag-and-drop interfaces. What would take a dev team weeks took him minutes with Composer. He calls these "time machine moments"—simultaneously localizing the app from Hebrew to English (two days, normally weeks), building a personal site (90 minutes from zero to live domain), and adding new features, all running in parallel while AI agents work. The feature went from Linear issue to working code in under an hour. His goal now is to inspire others: "If people walk away thinking how amazing you are, you failed. If people walk away and open their computer and start building, you've succeeded."