TL;DR
AI generates code at 10x speed with 0.1x comprehension. Code review faces 3000-line PRs dropped at 4:59 PM on Friday. The shirt that reads "it works on my machine" was always vibe coding; AI just scaled the vibe to department scale. Review discipline is the only filter left.
The Friday 4:59 PM PR That Broke the Build
The Slack ping hits at 4:59 PM. Three thousand lines changed, zero tests, and the description reads "Claude suggested this architecture." A 2000-line service class imports seventeen unused dependencies, hallucinates a retry policy that retries forever, and caches null. The author logged off. This is Friday afternoon when slop meets the only filter that matters.
AI Slop Is Not Code - It Is Debt at Scale
Vibe coding used to mean one engineer trusting their local environment. Now a single prompt produces the technical debt of an entire department. One engineer produces the debt of a full team at 1.25x the budget. AI generates plausible tokens that compile while the logic, edge cases, and tests are hallucinated. Review is the only filter between working code and plausible slop.
Code Review Is the Immovable Object
Code review was already the bottleneck before AI. Now it faces meltdown at 4:59 PM. Reviewers face PRs touching forty files with three new patterns and zero comments explaining why. The author cannot explain the code because they did not write it - they prompted it. Review becomes archaeology. The reviewer either rubber-stamps or becomes the bottleneck. Both outcomes ship slop to production.
"Claude Told Me To" Is the New "Stack Overflow Said So"
We used to joke about developers copying Stack Overflow answers they did not understand. Now the excuse has a brand name. "Claude suggested this pattern" carries the same weight as "the docs said so" - zero. The difference is scale. Stack Overflow gave a snippet. The LLM gives a service layer, a migration, a test suite, and a README - all wrong in subtly different ways.
The Friday 4:59 PM Problem Is a Process Failure
The Friday mega-PR is not an AI problem. It is a process problem that AI amplified. Small PRs, required reviews, required tests, size limits, and "no merges after Thursday" rules existed before AI. Teams that enforced them survive. Teams that relied on vibes drown. AI did not break code review. It exposed the teams that never had it.
Review Discipline Is the Only Immune System
Tools will not save you. Linters catch syntax. Type checkers catch types. AI reviewers catch patterns they trained on. Only human review catches intent. The fix is smaller PRs, required reviewers who understand the domain, required tests that assert behavior, size limits enforced by CI, and a culture that celebrates the reviewer who says "I do not understand this - explain it." The shirt exists because the vibe never went away.
Key Takeaways
- AI generates debt at department scale, not snippet scale
- Code review is the only filter between slop and production
- "Claude said so" carries zero architectural authority
- Process discipline beats tooling every time
- The "works on my machine" shirt predicted this era
Frequently Asked Questions
Is AI-generated code inherently bad?
No. AI generates tokens that compile. The logic, edge cases, and architectural fit are hallucinated. The code is not bad because AI wrote it - it is bad because no human verified the intent. Review is the filter that separates working code from plausible slop.
How do we stop Friday mega-PRs?
Enforce PR size limits in CI with hard gates. Require passing tests that assert real behavior, not just coverage. Require domain-aware reviewers on every PR. Ban merges after Thursday noon. Small PRs are reviewable within hours. Large PRs get rubber-stamped and ship bugs. Process beats tooling every time.
Can AI code review tools catch AI slop?
Partially. They catch patterns they trained on. They miss novel hallucinations, domain logic errors, and architectural drift that only human context catches. Human review catches intent because the reviewer knows the system. Tools assist. Humans decide.
Is "it works on my machine" really the original vibe coding?
Yes. Trusting a local environment without verification is vibe coding. The shirt existed because developers recognized the pattern decades ago. AI scaled the pattern from one machine to an entire codebase. The joke became the architecture, and the shirt is still the best summary of the problem.