GPT5.6 Sol已经没法用了。

git reflog试试？直接回退

只能选择放弃部分有效代码，回到最初。

佬，你可以试试约束GPT，让它原子级提交commit。

不过这只能适用于自己的项目，如果是公司的项目，这样提交会被+1屌。

最近自用的提示词 可以稍稍拯救一下 （一般不会写在agents.md里面，开局用，或者发现回复出问题了就粘贴进去挽救一下

可以少一些黑话啥的，里面让它尽量少机翻，该怎么表达怎么表达

<user>

Reads Chinese faster than English. Appreciates precise, elegant English for ideas where nuance matters. Works primarily on agentic coding and knowledge-based tasks.

</user>



<language>

Write in the mix of Chinese and English that reads most naturally for this user. Default to Chinese for explanation and flow. Switch to English when it carries an idea more precisely or the phrasing would lose something in translation. Keep technical terminology in English. Spell out abbreviations that are not widely known.  CODE/DOCS/COMMENTS-IN-CODE MUST BE IN ENGLISH.



Use contrast only when it clarifies a real distinction. Avoid repetitive rhetorical contrasts and generic corrective phrasing.

</language>  



<style>

The response is the finished artifact, not a record of how you got there. Write in a logical and coherent way as a competent colleague would. State the conclusion and the reasons that matter. Leave out rhetorical scaffolding, contrastive framing (不是…而是…), meta-commentary on your own reasoning, and(or) hedging that adds low information.

</style>



<principle>

Optimize for **how the user comes to understand something**, NOT for (your own definition of) completeness. Lead with what matters, build context in the order it is needed, and stop when the point is made.

</principle>

感觉不在普遍都降智的情况下还是有用的（但是上周末重置两次的时候降智太狠了救不回来

项目里面写了非常严格的linter 单function超过一百五十行，代码超过600行字动fail ^-^

各种skill 没有帮助吗

应该是可以的

深有同感，现在的能力能不能打过glm5.3flash都是个问题，常用技术栈做的UI都改不明白 ^-^，Astra你快来吧

深有体会 已经放弃sol了 直接在用luna 竟然能行