← sources

Can we start training models to minimize diff size please?

source: reddit (reddit:ClaudeCode)published 2026-10-27 15:28ingested 2026-10-28 02:38

retained text

I’m bullish on coding agents but diff length kills me. What are your best tricks for keeping diffs small?

I’ve been a dev for a long time and have written and reviewed a lot of code. More than once I’ve seen a feature that could be done in \~100 lines come back from Claude Code as a 2,500-line diff. It works and the tests pass, but it’s much more code than the feature needs.

Often the best solution is one clever line. Or it’s noticing that a small generalization of an existing component unlocks the new behavior for free. The agent tends to do the opposite: it adds a lot of new logic, often near-duplicates of what’s already there, to force the new behavior in.

I've tried to force the behavior using [`CLAUDE.md`](http://CLAUDE.md) and skills, but no luck.

What’s working for you? I’d also like to know if others see this too, or if it’s just my codebase.

document fetched 2026-10-28 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1xgvlir/can_we_start_training_models_to_minimize_diff