Chiseling: The Art of Polishing Vibe Code (patrickerichsen.com)

Mitchell Hashimoto’s “anti-slop sessions” inspired a reframing of vibe coding: treat LLM-driven rapid prototyping like sculpture. Use the CLI as a jackhammer to break off big chunks of functionality and validate architectures quickly — the “make it work” phase — but then switch to the IDE for a deliberate “chiseling” phase. At Continue the argument is explicit: AI accelerates iteration, but human-led cleanup (refactoring massive methods, removing duplicated or over‑engineered abstractions, renaming ambiguous APIs, pruning useless tests) is essential to transform a working prototype into maintainable production code. The significance for the AI/ML community is practical and structural. Chiseling forces comprehension—each edit requires understanding what the code actually does, not just that tests pass—so it prevents rapid accumulation of technical debt from thousands of tokens of AI output per hour. Technically, the recommended workflow is iterative: explore with agentic CLI tooling, then refine with IDE features (Chat, autocomplete, focused refactors) and better documentation so future agentic sessions perform better. In short, vibe coding gives velocity; chiseling gives quality—together they produce robust, readable, and maintainable systems.