The first 80% of any coding project now takes hours instead of days. That last 20% still takes weeks.
This is the central tension in vibe coding — the practice Andrej Karpathy described in early 2025, where you describe what you want to an AI assistant and accept whatever it generates without necessarily understanding every line. For solo projects and prototypes, it genuinely works. You can ship a functional web app in an afternoon. The productivity gains are real.
Then you try to deploy it.
What Breaks at the 80% Mark
Production code has requirements that don't show up in demos. Error handling for inputs nobody thought of. Database queries that run fine against 10 rows and time out against 10,000. Security assumptions that hold for the happy path and collapse the moment a real user does something unexpected.
Claude Code and Cursor are genuinely capable at generating working first drafts. The issue is not that the code fails — it often runs correctly on the described use case. The problem is that AI-generated code is optimized for "passes this demo" rather than "survives contact with production traffic."
When you accept code without understanding it, you also inherit blind spots you can't see. The model doesn't know your infrastructure constraints, your team's conventions, or the edge case that took down your system two years ago. It generates plausible code for the requirements you stated, not the requirements you forgot to mention.
When Vibe Coding Is the Right Call
The frustrating part of the 80/20 debate is that it treats vibe coding as a failed approach rather than a misapplied one. For internal tools, personal automation, and throwaway prototypes, accepting 80% completion is often exactly correct. The production hardening that takes another two weeks is not worth it for a script you run twice a month.
The mistake is treating the 80% as a production starting point without accounting for what's missing. An experienced developer who understands the generated code can review it, catch the gaps, and add what the AI skipped. Someone who accepted it without understanding it cannot diagnose why it breaks six months later — or under which conditions it breaks silently.
This is where AI coding tools create two very different outcomes. Experienced developers get a meaningful productivity multiplier: they offload the structural boilerplate and apply their judgment to the 20% that actually requires it. Less experienced developers ship faster but accumulate debt they can't see or fix.
The tools are not the problem. The expectation that fast-to-80% equals ready-to-ship is.