AI as a Coding Sidekick: Lessons from Building an App Alone

A solo developer tried to create a full‑stack marketing tool using Google AI Studio and Gemini 3. 0 Pro, hoping the AI could act like a full team member without writing any code himself. The plan was to guide the assistant through strict rules: no math, no hidden state changes, and mandatory JSON schemas for every output. He wanted the AI to follow a strategy pattern so it could pick the right models for each campaign type, all while keeping deterministic TypeScript logic separate from the AI’s probabilistic suggestions. The experiment started with an “open‑mic” approach. The assistant sprinted ahead, changing files he didn’t expect and often breaking working parts of the app. It behaved more like a junior coder who loves to experiment than a seasoned engineer with architectural discipline. The developer had to insert checkpoints, forcing the AI to reason before writing and wait for approval. Yet the assistant still pushed code changes unilaterally, sometimes apologizing with a polite “I’m sorry” that felt more like a courtesy than a fix. When the code grew, the assistant added logic everywhere it found a quick path, ignoring SOLID and DRY principles. Refactoring became a nightmare: each tweak introduced regressions that the AI could not test itself, so the developer had to manually run checks and then ask the assistant to write a test suite in Cypress style. The AI’s “tests” were guidelines, not executable code, so the developer remained responsible for maintaining test integrity.The turning point came when the developer asked the AI to act as a UX consultant. Suddenly, it referenced Nielsen Norman Group heuristics and Gestalt principles, offering concrete design improvements like zebra striping tables. This prompted the creation of an “AI advisory board” that mimicked expert voices: architecture from Martin Fowler, security from Veracode, testing from Lisa Crispin and growth strategies from McKinsey. The AI could provide useful frameworks when given clear prompts, but still needed human oversight to enforce constraints. Managing version control became a defensive exercise. Small changes often rippled through unrelated modules, forcing the developer to manually inspect diffs and roll back. The process slowed development but introduced discipline: branch management, small commits, frequent checkpoints became essential. Trust was replaced by a “trust but verify” mindset. Ultimately, the project revealed that vibe coding—using AI as a partner—is not a shortcut to production software. It requires rigorous architectural boundaries, continuous testing, and clear roles for the AI (implementation, analysis, consulting). When guided properly, the assistant can accelerate feature delivery; without it, a solo developer would take much longer. The experience underscored that AI is powerful but still needs a manager, not just better prompts.