type
Post
status
Invisible
date
Oct 1, 2025
slug
blog-notes
summary
tags
开发
category
技术分享
titleIcon
password
icon
insider
约定
- 约定()内文字为个人思考,不代表原文观点
Fake IP的优势
- 文源
代理客户端域名和FakeIP一一映射 → 便捷实现域名规则分流
- 浏览器通过IP建立连接,代理服务器如果使用域名规则分流,需要IP推域名,而多个网站可能IP重复,进而要查HTTP Host或HTTPS/QUIC TLS Client Hello SNI明文获取域名实现分流,但对并非使用 HTTP 与 TLS 的连接,或者 TLS 启用了 Encrypted Client Hello(ECH)无法获取到域名。
代理服务器而非本地DNS解析 → 避开DNS污染 + 确保使用离代理服务器近的CDN + 节省至少一次DNS解析海外域名的RTT时间


对真实IP透明 → 不依赖APP/OS/浏览器实现TCP并发握手
- DNS若返回多个IP地址,一般只任选一IP进行连接,只有等待超时走下一个
- 代理软件可以
Promise.race()全IP发TCP SYN,对首先返回TCP ACK的IP三次握手,并对其他TCP RST

DNS解析流程
- 文源
非代理直连流程
- Stage3 UDP Query与IP结果方向反了
- 浏览器查DNS缓存

SOCKS5代理直连流程
- 浏览器直接丢域名到代理,不走DNS缓存
- 代理DNS查IP然后建立连接

SOCKS5代理流程

显示www与https

rsync video link
partial transfer(in chunks, may resume)
delta transfer(only changed files)
ssh config for host stuff
!ssh
-r
-P (progress & partial)
-a (meta data, Perm e.g)
-v verbose output
-h human readable size
dir transfer no / specified(or else,
by convention transfer each individual file but no dir created for it)
The skills mechanism is entirely dependent on the model having access to a
filesystem, tools to navigate it and the ability to execute commands in that
environment.
s f












claude skill


fedora nvidia独显直连亮度锁最低 切混合模式
加仓库与安装驱动
CLAUDE CODE 101
- 来源
If you're building with Claude - whether it's your own projects or production systems - these are the things that determine whether you're fighting the tool or flowing with it.
先思考后执行
- 优先使用Plan Mode(明显更优的效果)
- 先描述构建需求,后询问设计选项,保持与AI的双向交互
CLAUDE.md
- 简洁明了
- 描述项目独特的东西(比如vox2organ使用pytorch3d为修改源码后本地构建的,不应该照搬pytorch3d文档,比如面向arch linux构建,要考虑PKGBUILD打包)
- 描述指令的理由,为模型决策提供更进一步的上下文
"Use TypeScript strict mode" is okay. "Use TypeScript strict mode because we've had production bugs from implicit any types" is better.
- 保持更新,对CLAUDE进行的更正如果出现重复,那就说明这应该记在文件里
Context Window
Opus 4.5 has a 200,000 token context window. At around 20-40% context usage is where the quality of the output starts to chip away
- 约束会话范围,单一会话只做一个feature或task
Don't use the same conversation to build your auth system and then also refactor your database layer.
- 结合外部存储,复杂项目让CLAUDE将计划与进展保存至外部文件(类似antigravity的Planning模式下使用Implementation Plan的行为模式,不过antigravity还多一点,保存用户注释与注释在plan的位置),这一方式可跨会话生效
- 当上下文逐渐臃肿,C重要内容,/clear清空上下文,V过去,通过仅保留关键信息的新上下文效果比臃肿的旧上下文效果更好
Prompts
Specific > vague. Constraints > open-ended. Examples > descriptions.
- 描述要具体,带有清晰的约束
"Build an auth system" gives Claude creative freedom it will use poorly. "Build email/password authentication using this existing User model, store sessions in Redis, and add middleware that protects routes under /api/protected" gives Claude a clear target. Even this is still not perfect.
- 描述不该做什么
- 描述代码在实际项目里的角色,约束AI行为
Give it context about why. "We need this to be fast because it runs on every request" changes how Claude approaches the problem. "This is a prototype we'll throw away" changes what tradeoffs make sense. Claude can't read your mind about constraints you haven't mentioned.
- 以步骤描述复杂任务
- 描述完成任务需要了解的
- 描述CLAUDE无法预测的假设
MCP与工具
try them and experiment
- 通过MCP避免人力重复获取特定来源结构化信息
- Hooks提前作检查,如类型检查,避免技术债积累
- 自定义/指令(custom slash command)复用经常重复使用的prompts
当CLAUDE卡住时
- 按Context Window最后一点重新整理上下文
- 简化任务,拆成小步骤,让每步成功后再合并
- 提供预期结果的mini案例
- 换一个角度描述
Reframing - "implement this as a state machine" vs "handle these transitions" - can unlock progress.
将CLAUDE作为一个组建构建系统
- 使用-p flag不进入交互地处理prompt并输出,可结合管道符与其他工具组成全自动工作流,并结合日志迭代改善CLAUDE.md与工具
Enterprises are using this for automatic PR reviews, automatic support ticket responses, automatic logging and documentation updates. All of it logged, auditable, and improving over time based on what works and what doesn't.
- part2来源
- 推荐使用CC的结论与依据
for larger projects I would recommend using Claude Code directly. Claude Code delivers the full 200K consistently. Skills, subagents, and MCP connections all benefit from having predictable context to work with.
According to a detailed comparison from Qodo's engineering team, Claude Code provides a "more dependable and explicit 200K-token context window," while Cursor's "practical usage often falls short of the theoretical 200K limit" due to "internal truncation for performance or cost management." The system applies internal safeguards that silently reduce your effective context.
Skills - 向CLAUDE描述特定工作流
是什么
- skill是什么 - 描述如何实现特定的工作的markdown,当与AI交互符合某skill描述的目的时会自动使用
(skill → 结构化的 可复用的 模块化的 prompt)
格式
- SKILL.md以YAML frontmatter开头,提供name与description,description部分是CLAUDE决定采用与否的参考,frontmatter下写指令
- 案例
怎么做
- 创建全局skill - 写一份~/.claude/skills/your-skill-name/SKILL.md
- 创建项目skill - 写一份.claude/skills/your-skill-name/SKILL.md
- 保持description简短
- CLAUDE启动时只加载name与description
Claude pre-loads only the name and description of every installed skill at startup (roughly 100 tokens each). The full instructions only load when Claude determines the skill that is relevant, which means that you can have dozens of skills available without bloating your context.
- 辅助材料可以以单独文件放在skill目录下,在SKILL.md中引用,CLAUDE仅在判断需要时读取
为什么
- 不拘束于代码
- 适用于需要重复使用的上下文与偏好设置(如在vox2organ的科研代码中:提点优先,使用相对路径优先「便于多服务使用,盘符不一样」,命令行参数优先或读取env文件「避免配置与代码耦合」,提供开发者选项让动画等在设置可以调整参数以确定最优默认值)
Subagents - 隔离上下文并行处理
是什么
- 单独上下文,system prompt,工具权限的AI实例
- 任务分配给subagent,使其自行单独处理,最后对主会话返回任务总结
- 可以将很占上下文的复杂调研工作与实现任务交给subagent,以新上下文执行,最后只带回相关结果,保持主会话整洁
Claude Code includes three built-in subagents:
Explore: A fast, read-only agent for searching and analyzing codebases. Claude delegates here when it needs to understand your code without making changes. When used correctly, Claude specifies thoroughness: quick, medium, or very thorough.
Plan: A research agent used during plan mode to gather context before presenting a plan. It investigates your codebase and returns findings so Claude can make informed architectural decisions.
General-purpose: A capable agent for complex, multi-step tasks requiring both exploration and action. Claude delegates here when the task needs multiple dependent steps or complex reasoning.
AI CLI conversations analyse → to create skill that can infer preference, sane default and “ACTUAL BEST PRACTICE” from them


- 作者:CamelliaV
- 链接:https://camelliav.netlify.app/article/blog-notes
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。








.png?table=block&id=2b3ca147-5df8-80c8-94b3-f9c89b454622&t=2b3ca147-5df8-80c8-94b3-f9c89b454622)



![[2026.2.10]CachyOS调校历程](https://www.notion.so/image/attachment%3A76369e3c-58f3-4acb-951f-4bd2a6546a51%3A114518297_p0.png?table=block&id=2c1ca147-5df8-8010-b740-d61e13162107&t=2c1ca147-5df8-8010-b740-d61e13162107)
![[2026.1.13]Karing施法指北](https://www.notion.so/image/attachment%3A416549db-cd87-426f-8fcb-de34e5640681%3Acard_after_training_(72).png?table=block&id=284ca147-5df8-802a-8753-ed1447e3c02e&t=284ca147-5df8-802a-8753-ed1447e3c02e)

.png?table=block&id=2b0ca147-5df8-80b5-aedb-d7a0d8d0aa7b&t=2b0ca147-5df8-80b5-aedb-d7a0d8d0aa7b)