惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

N
Netflix TechBlog - Medium
G
Google Developers Blog
H
Hackread – Cybersecurity News, Data Breaches, AI and More
T
The Blog of Author Tim Ferriss
Microsoft Azure Blog
Microsoft Azure Blog
GbyAI
GbyAI
L
LangChain Blog
云风的 BLOG
云风的 BLOG
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
aimingoo的专栏
aimingoo的专栏
P
Proofpoint News Feed
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
小众软件
小众软件
WordPress大学
WordPress大学
A
About on SuperTechFans
大猫的无限游戏
大猫的无限游戏
C
Check Point Blog
月光博客
月光博客
Stack Overflow Blog
Stack Overflow Blog
美团技术团队
Jina AI
Jina AI
T
Tailwind CSS Blog
Google DeepMind News
Google DeepMind News
D
Docker

LINUX DO - 最新话题

谷歌云盘下载700g数据集,求方法 OpenAI推出了100美元的Pro订阅后,plus的Codex 5小时限额大幅缩水 之前买的super grok居然还没掉 关于CPA认证文件周限 佬们,默认CDK的要求是什么等级啊? 最新版本的微信群聊机器人方案 有没有人知道如何free号没有封,那么是否可以循环使用,因为我看主要是周限 L站改版了?吓我一跳,我以为我浏览器崩了 淘宝这种宽带可信吗,500兆移动宽带月费8元到2099年 docker内部应用访问宿主机mysql和redis时被拒绝connection refuse Erp全栈想转行做Ai有什么推荐的吗 boost有bug 佬们,有没有靠谱点的 Plus 购买渠道 大妈,狗妈用的 lg 服务有源头开源项目吗? 有人有能过验证码打码的嘛 上次帖里好像发过通过大模型来打码的 gpt plus 封号似乎也太快了点,一天就给封号了 按流量/token收费的国产官方AI推荐 我算是知道了为什么Oracle总是ABC了 佬友们帮我分析一下 ChatGPT Team账号只有一个人使用和4个席位邀请满了使用的总额度是一样的吗? gpt-free 10个带rt CPA反代claude是默认1m吗? 我终于敢说我做出来windows上tmux的替代了,目标windows/全平台最强的终端Ai编程工具 claude pro升级max,除了原来的$20,好像还能再领一次$100 关于AI agent的知识框架 独乐乐不如众乐乐,分享一下我的的AI对话程序 佬们自建网站支付问题是怎么解决的 怎么能让gpt模仿claude风格输出 codex free已经死了,下一个会是plus或者team吗 请问chatgpt pro里的fast模式,速度快了,降智吗
昨天CC中Claude模型犯病,写个东西中途老是乱编,有同遇到的佬吗
hushaoyong (shaoyong777) · 2026-06-25 · via LINUX DO - 最新话题

昨天CC中Claude模型犯病,写个东西又慢,中途还老是乱编,是我的配置有问题还是服务器问题呀,有佬也遇到了吗?

zdf73 (DenisZheng) 2

加个提示词试下: Zero-Guessing Principle: Never Encode Uncertainty Into Code [Core Guardrail]

When working with code paths, data structures, field ownership, call chains, and business logic flow, you must first establish facts using tools, and only then write logic. Using fallback code to paper over unresolved questions is forbidden.

  • Facts before implementation: Do not speculate about system flow, field origin, or object ownership until the actual path has been confirmed via source code search, AST scanning, file reading, type definitions, call-site tracing, tests, or terminal execution.
  • Single source of truth: When facing questions like “which object does this value actually belong to,” “where does this node actually come from,” or “who actually produces this field,” you must trace the source code to find the one definitive source. If it can’t be found, keep searching or stop and explain the blocker — do not write logic first and figure it out later.
  • No speculative fallback code: Do not write a.xxx || b.xxx, a?.xxx, a && a.xxx, multi-layer fallbacks, fuzzy field-name compatibility shims, or candidate-path polling to sidestep uncertainty — unless the source type definitions, protocol spec, or business requirements explicitly state that the field is genuinely multi-source/optional.
  • No placeholder naming that then propagates: Do not create temporary abstractions like target, payload, data, or node for node names, field names, event names, or config keys when their real identity is still unconfirmed, and then keep building on top of them. Confirm the real structure first, then name and implement.
  • Exceptions must be made explicit: If the factual chain cannot be confirmed, stop the current implementation and report what you’ve verified so far, what evidence is missing, and what needs user confirmation. Do not “advance by covering multiple possible cases” instead.
  • Permitted exception for multi-branch logic: Multi-way conditional logic is only allowed when type definitions, protocol docs, existing code patterns, or business requirements explicitly indicate the target is genuinely multi-branch, multi-version, multi-source, or has an optional field — and in that case, the rationale must be documented in the code or change description.

hushaoyong (shaoyong777) 4

感谢佬 不过今天好像就正常了 目前没那么卡也不乱回复了