The Beautiful Boring: Refactoring with AI (www.petervanonselen.com)

This piece lays out a practical, disciplined workflow for “vibe coding” with AI—treating models as partners but engineering around their predictable weaknesses. The author prescribes several repeatable patterns: a Planning Folder Pattern of numbered spec files that act as persistent cross-session context; a Todo Accountability System that converts specs into granular checkboxes to hold the model to concrete acceptance criteria; and a Git “save-scumming” habit of frequent commits so you can roll back when an AI overwrites a working solution. Combined, these techniques preserve institutional memory, make progress auditable, and mitigate hallucination, context loss, and momentum-driven tangents. It also recommends role-based model selection (ChatGPT for brainstorming, Claude for implementation and code review, Copilot/Codex for ticket-style handoffs) and a “Discipline Override” mindset: enforce refactor cycles, write tests even when the AI seems to have solved the problem, and constantly ask “Is this the MVP?” Finally, a minimal but powerful prompt rule: always end prompts with “Ask me questions.” Forcing the AI to query you turns one-shot outputs into iterative dialogue, surfacing missing assumptions and producing much more reliable, refactorable code. The net effect: boring engineering practices supercharged by AI, reducing risk while improving collaboration and velocity.