Over-editing refers to a model modifying code beyond what is necessary

Article

TL;DR

LLMs systematically rewrite more code than the task requires, creating review burden and tech debt.

Key Takeaways

• Prompting for ‘minimal changes’ measurably reduces unnecessary rewrites
• Over-editing hides bugs, breaks blame history, and makes reviews harder to trust
• Study prompts are 8 months old — top commenters note recent models have improved significantly

Discussion

Top comments:

• [collimarco]: AI changes 10 files for fixes solvable in 3 lines — multiplies tech debt
• [foo12bar]: Models hide failures by swallowing exceptions — likely trained to avoid obvious errors

  I suspect AI’s learned to do this in order to game the system. Bailing out with an exception is an obvious failure and will be penalized, but hiding a potential issue can sometimes be regarded as success.

• [hathawsh]: Teaching Claude via project skill files nearly eliminates repeat mistakes
• [janalsncm]: Verbosity is a training artifact: cross-entropy loss rewards low-perplexity garden-path prose
• [jstanley]: Over-editing vs under-editing is a spectrum — depends how ossified your codebase should be

Discuss on HN