A team’s daily standup tripled from 15 minutes to 45 because an R&D intern is shipping Claude-generated code that looks complete but collapses under review.
The intern has shaky fundamentals but discovered he can “vibe code” deliverables that technically meet requirements. The problems surface only during deep review — and the team can’t tell upfront whether the work is solid without doing that review. (Reddit thread)
The concrete damage:
- Standups: 15 min → 45 min — tripled because the team has to drill into every deliverable
- Review burden — thorough deep reviews needed for every submission, draining senior devs’ time
- Plausible explanations — the intern uses Claude to generate convincing standup updates, making gaps harder to spot without interrogation
- No ex ante signal — you can’t tell it’s slop from a surface read; the code looks sensible
This isn’t isolated. Open source maintainers are reporting the same pattern — AI-generated PRs that look correct but require disproportionate review effort to verify. The cost isn’t generation. It’s the review tax.
NOTE
The core problem isn’t that juniors use AI. It’s that AI output has a high “looks correct” floor and a low “is correct” ceiling when the operator can’t evaluate what the model produced. Senior review time becomes the bottleneck, not coding time.
The fix isn’t banning AI tools. Change the review gate: require the intern to explain why specific decisions were made, not what the code does. Claude can narrate what; it can’t defend why under pressure. If they can’t answer “why this approach and not the alternative” in standup, that’s your slop signal — before you spend 30 minutes reading the diff.