Four months of building side projects with AI: what works (vanonselenp.github.io)

Over four months the author transformed from an AI hobbyist into an active AI collaborator, using ChatGPT, Claude Pro and Copilot to build multiple side projects — notably a Magic: The Gathering “Jumpstart cube” solution, a portfolio site and a sprawling board/video game (about 24,000 lines of Python code). Paid subscriptions (Claude Pro, ChatGPT Pro) were central to the workflow: Claude for implementation and review, ChatGPT for exploration and brainstorming, Copilot for ticket-style handoffs. The experiment proved catalytic: what started as curiosity became a rapid-feedback loop of idea → AI-assisted generation → manual refinement that produced playable prototypes but also massive, hard-to-understand artifacts (a 10,000‑line IPython notebook, many files >2k lines). Key practical takeaways and technical implications: treat AI like a “junior engineer” — keep persistent specs as shared memory, break work into granular todos to hold the model accountable, commit very frequently (“git save‑scumming”), and enforce red‑green‑refactor cycles and tests to avoid runaway technical debt (the “refactor paradox”). Be wary of “tangent amplification” (AI makes scope creep feel achievable) and AI spatial blindness (orientation/rotation errors in game geometry), and use role-based selection (ChatGPT for ideation, Claude for coding/reviews, Copilot for handoffs). For the AI/ML community this means practical tooling patterns and guardrails (spec-driven workflows, test discipline) are as important as model capability when scaling productivity without sacrificing maintainability.