Opus 5 output is unreadable on long sessions
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
用户反映，Opus 5在处理大型代码库等长时间会话中，其输出变得难以理解。尽管AI的回复在技术上是正确的，但它会采用过于复杂和充满行话的风格，频繁使用“承重”和“爆炸半径”等术语，导致信息无法使用。尝试通过在claude.md中添加“简单技术英语”等指令来缓解此问题，但在长时间会话中无效，因为这些用户消息会被埋没。…
So opus 5 is fine on short sessions but if you let it run on a big repo for a while the output gets honestly unusable. like whatever it tells you is correct but you cant actually use it because it read so much of the codebase that it writes like you already read it too. everything is "load-bearing" and "blast radius" and "surface area" no matter what the context is. got this gem back, actual quote: "This is the one I'd fix first: it's cheap, and it silently degrades everything built on top." like ok thanks
I tried putting "plain technical english, don't reference stuff I haven't seen" in claude.md and it works on small tasks but on long sessions it just stops caring. I think because claude.md is a user message after the system prompt so it gets buried. asking for concise actually backfires because load-bearing is technically fewer words than what it means ended up putting the rules in an output style instead since that goes in the system prompt, plus a hook that blocks on those phrases: https://github.com/ukanwat/claudesplain Its helped me a ton but curious how you guys deal with this.