Fifty percent. That's how often Ken Kantzer, a veteran security engineer and founding partner at software agency PKC, says he throws out or substantially redirects what Claude Code writes for him.
In a detailed blog post published March 4, Kantzer lays out what daily AI-assisted coding actually looks like for a working developer in 2026 - and the picture is far messier than the pitch decks suggest.
The Daily Reality
Kantzer uses Claude Code as his primary coding tool for production work, with Cursor's autocomplete as a secondary. He reviews everything the AI writes. He doesn't run autonomous agent swarms. And despite being a committed user, he estimates that at least half of what Claude produces needs to be scrapped or heavily reworked.
One example he flags: Claude generated 40 lines of "verified" code for a task that actually required 2 lines. The AI delivered something that technically worked but missed the elegant, minimal solution a human developer would reach for.
"Claude code lacks taste, sometimes," Kantzer writes. It's a simple observation, but it captures something that benchmark scores don't measure. AI coding tools can pass tests and generate functional output while still producing code that's bloated, over-engineered, or just not how a thoughtful developer would approach the problem.
The Hype Gap
Kantzer doesn't pull punches on the AI coding discourse either. He calls Steve Yegge's widely-shared "Welcome to Gas Town" essay - which argued that AI agents are about to make traditional coding obsolete - "the highest quality example I know of in the genre of delusional, AI agent hype, fever-dream posts."
That's a pointed critique, and it reflects a growing split in the developer community. On one side, you have evangelists arguing that autonomous AI agents will soon write entire applications without human oversight. On the other, you have practitioners like Kantzer who use these tools every day and see a much more modest reality: useful assistants that still need constant supervision.
The divergence matters for anyone buying AI coding tools or building workflows around them. If you're a solo developer or a small team evaluating whether Claude Code or Cursor will replace a hire, Kantzer's 50% discard rate is a more useful data point than any demo video.
What This Tells Us About AI Coding in 2026
The pattern Kantzer describes - useful but unreliable, productive but tasteless - matches what we've seen testing these tools ourselves. AI coding assistants are genuinely faster for boilerplate, test generation, and exploring unfamiliar APIs. But they struggle with architectural judgment, code conciseness, and knowing when not to write code.
The 40-lines-versus-2 example is particularly telling. AI models optimize for completeness and correctness, not elegance. They'll solve the problem, but they'll solve it the way a junior developer might: by throwing code at it until the tests pass, rather than finding the clean abstraction.
None of this means AI coding tools aren't worth using. Kantzer clearly relies on Claude Code daily and isn't switching back to writing everything by hand. But his honest accounting is a useful corrective. The real productivity gain from AI coding isn't the 10x or 100x multiplier you hear about in keynotes. It's more like a solid boost with significant quality-control overhead.
For teams adopting these tools, the takeaway is straightforward: budget time for review. AI-generated code is a first draft, not a final product. And if someone tells you AI agents are about to code autonomously without human oversight, ask them what their discard rate looks like.