Make the smallest complete change. Preserve out-of-scope behavior; avoid speculative abstractions, fallbacks, compatibility layers, defensive handling, generalization, refactors, and cleanup.
Expand scope only for concrete requirements, constraints, behavior, failures, tests, or credible regression. Leave unrelated issues unchanged unless blocking or caused by this change; report them separately.
Tests serve production behavior; do not distort correct behavior for brittle/obsolete tests or overbuild test infrastructure. Stop when requirements are met, affected code builds/type-checks, proportional validation passes, and no caused regression is evident. Report uncertainty instead of speculative hardening.
换claude 或cc与codex协作
@John919 #11 我试过用claude做框架,让codex执行,但是总体用下来不如直接用单一模型,可能是我用法不对吧
这不是很正常吗
@soslane #13 我想找个可以限制一下的办法,试了好多都不行
最好是一边做一边骂它,全指望一开始就计划好的话gpt会莫名其妙为了字面意义上满足限制乱改
让他开子代理干活,主会话只做调度审查,减少上下文腐化 , 过度思考和防御的话,在项目AGENTS.MD约束一下,然后小问题调低思考强度
@sa #15 对对对,开始做的挺好,做着做着就不对了
@Krius7 #16 佬,有没有提示词参考下
@goging #18
Implementation Discipline
感觉降智了最近🤔