If AI coding is lowering your code quality, you're not managing quality right

Commenters debate whether falling code quality from AI-assisted coding is a people-and-process problem or an intrinsic tool limitation. Some, like moltar, frame it as caring and craft: developers who guide agents with system-design thinking and strict review produce maintainable results, and managers must enforce quality practices (chadash, yread). Others describe practical harnesses that work: fishfasell, qarl and teliskr report success with heavy planning, subagents, tests and iterative reviews that raised coverage and modernized legacy systems (mark_l_watson gives a retirement-era example). Several emphasize that consistent guardrails, tests and quality-management systems are required to prevent gradual entropy, with concrete workflows - plan aggressively, review every commit, create golden-master tests - claimed to deliver stable outcomes.

Countervoices argue that those prescriptions are unrealistic or unfair. Zug_zug says asking authors to become editors is harder than writing fresh code, and compiler-guy objects to shifting blame onto users instead of fixing model behavior. ThePhysicist warns of recursive agentic loops that probabilistically drift into spaghetti code unless humans intervene every step; bunderbunder says AI can improve code quality but worsen design, returning to waterfall dynamics. Others bluntly blame LLMs for producing unmaintainable output (mococa, VCFundedGenYer), while some taunt that if AI isn’t lowering your quality you either started low or aren’t using it enough. The split is clear: some trust disciplined processes and tooling to tame AI, others see fundamental limits or responsibility failures that make AI a net risk.