What do you do when the AI writes code that runs, but you wouldn't accept it in a code review?
That's the "slop" problem. In AI-generated code, slop isn't a crash or a wrong answer - it's output that's technically functional but mediocre: over-engineered for simple tasks, missing edge case handling, or generic enough that editing it takes longer than writing it from scratch would have. Hallucinations get most of the attention because they're easy to catch (the code breaks). Slop ships.
Engineering teams using AI coding tools like Claude Code, Cursor, and Aider daily are building explicit processes to filter slop out before it enters the codebase. The core insight: the problem isn't the model, it's the feedback loop. Most AI coding sessions follow the same pattern - write a prompt, accept the output, discover problems during review or in production. A slop-proof workflow reverses that order: define what good looks like before generating, verify against that definition immediately, and reject outputs that don't meet the standard.
What Actually Filters Slop
Specificity upfront. Vague prompts produce vague code. Teams that consistently get usable output write prompts that include the constraints ("handle empty arrays without throwing"), the performance context ("this runs 50,000 calls per second, optimize for speed not readability"), and the expected behavior - effectively a mini-spec. AI tools are far better at hitting a precise target than inferring what you need from a broad description.
Treating AI output like human code. Some teams run AI-generated code through the same automated review pipeline as anything a human wrote - linting, static analysis, the same checks - before it enters the codebase. If the output fails, it gets regenerated or manually corrected. This catches the most common slop patterns: unused imports, functions that do too many things, missing error handling on external calls.
Scoped context. AI coding tools produce better output with focused context. Feeding a large codebase and asking for a feature often produces code that's consistent with the files the AI can see but structurally wrong for the system it can't. Scoping context to the specific files and functions relevant to the task narrows that gap significantly.
The pattern across teams reporting real productivity gains is counterintuitive: they're doing more upfront work, not less. Clear requirements, well-structured codebases, comprehensive test coverage - AI rewards all of these and amplifies problems in teams that skip them. If your workflow was sloppy before AI tools, the output will be sloppy too.