# Zhaphar > Aklman 的博客 A personal editorial site on writing, engineering, agents, and durable growth. This is a source index, not a full-text archive. Published Chinese and English posts, sessions, and reviews have individual `/llms.md` versions; Canon has a collection version. Works link to HTML pages and have no Markdown twin. About, Now, Consult, and other static pages have no Markdown twin. Chinese lives at the root; English lives under `/en`. The Kazakh interface lives under `/kk`; it does not currently expose Markdown twins. Read individual sources as needed for summaries, Q&A, and reasonable quotations, retaining author attribution and the original link. Quotation is distinct from full republication. For full republication, bulk collection, or model training, contact the author for prior agreement; machine-readable access does not grant those permissions. ## Posts - [Agent 工程学习指南(2026–2027)](https://aklman.com/posts/ai/agent-engineering-guide-2026/llms.md): 从「会用 Claude Code」到系统性 Agent 工程的六阶段路径:吃透循环、工具、上下文、评测这些不变量,跟紧协议与编排范式这些变量,用自建评测集校准一切结论。 - [把智能体接进现有系统](https://aklman.com/posts/ai/agents-in-existing-systems/llms.md): 智能体不是新系统,它是现有系统的一个新客户端。用对待客户端的方式对待它。 - [评测先行的工程节奏](https://aklman.com/posts/ai/evaluation-first/llms.md): 没有评测集的时候,每一次改动都是凭感觉。有了它,才谈得上迭代。 - [上下文比提示词更重要](https://aklman.com/posts/ai/context-over-prompts/llms.md): 调提示词的收益很快见顶。真正决定输出质量的,是你喂进去的那些材料。 - [链表:绕过,而不是抹除](https://aklman.com/posts/data-structures/linked-list/llms.md): 把数组的账反过来:增删是几根指针的改写,随机访问没了。头插、尾插、中插、删除、反转,全部画进可步进的分镜。 - [数组与切片:连续内存的红利](https://aklman.com/posts/data-structures/arrays-and-slices/llms.md): 连续内存买到两样东西:O(1) 随机访问和 cache 友好;切片是数组的租约——长度、容量、扩容摊还,一次讲清。 - [指针与结构体:地址就是数据](https://aklman.com/posts/data-structures/pointers-and-structs/llms.md): 系列的地基:内存是一排编了号的格子,指针是存着编号的格子;&、*、nil 三个符号,加「Go 只有值拷贝」这一句。 - [复杂度:数步骤,不背表格](https://aklman.com/posts/data-structures/complexity/llms.md): Big O 不是背表格,是数步骤:同一段代码,数出它随规模长大的方式;摊还与常数把工程判断拉回测量。 - [打开引擎盖:这个系列怎么读](https://aklman.com/posts/data-structures/opening/llms.md): 系列开篇立记法:内存表、记录体、可步进分镜、Go 示例——每一篇都回答同三个问题:数据在哪,指针指谁,代价花在哪。 - [让模型做它擅长的事](https://aklman.com/posts/ai/let-the-model-do-what-it-is-good-at/llms.md): 模型擅长把模糊的东西变具体,不擅长保证一件事一定发生。工程的边界就在这里。 - [Go 1.15 有什么新变化?](https://aklman.com/posts/repoers-and-reviews/go1.15-what-is-new/llms.md): 一篇关于 Go 1.15 的版本笔记,重点放在链接器、运行时成本和标准库的小而实用的变化。 ## Works - [Seeker 探索者](https://aklman.com/works/seeker): 一个装 AI 的「壳」,壳里是可自由开关的小应用。对话就是入口:说出需求,Agent 自己判断该查数据、出卡片,还是跑你沉淀好的技能——它想成为的不是「又一个聊天框」,而是长在你自己数据上的个人工作台。 - [Library](https://aklman.com/works/library): 一个独立的 AI 电子书库:每本书只讲透一份你已经在付费的 AI 订阅或一套工作流——该学什么、能跳过什么、什么时候换工具。免费在线读,无注册墙。 ## Sessions - [如何把 Agent 放进真实工作流](https://aklman.com/sessions/build-with-agents/llms.md): 一次面向社群的短分享:把 Agent 从工具列表重新表征为上下文、角色、交付物和反馈组成的工作系统。 - [Codex App:从任务到可验证变更](https://aklman.com/sessions/codex-app-demo/llms.md): 一个产品演示脚本,展示如何把目标、上下文、执行和验证放进同一条开发流。 - [AI 时代如何重新设计学习](https://aklman.com/sessions/learning-with-ai-video-script/llms.md): 一份长视频脚本草案:从信息获取、理解、拓展、纠错到应用,重新组织个人学习系统。 ## Canon - [Canon — collection Markdown](https://aklman.com/canon/llms.md): Curated entries with attribution and original sources. - [学习即压缩:表征 / 图式 / 心智模型 / 解释框架](https://aklman.com/canon#canon-learning-is-compression): 重写了我对「学习」的定义——从此读任何东西都问「能拿走什么结构」。 · 原文: https://www.dedao.cn/course/article?id=qavDm012GolV7OYAjxVxVjEy8zdk73 - [The AI-Native Software Engineer](https://aklman.com/canon#canon-ai-native-software-engineer): 把「AI 时代的工程师」讲清楚了——和我现在做的这套东西同源。 · 原文: https://addyo.substack.com/p/the-ai-native-software-engineer - [How to Write a Good Spec for AI Agents](https://aklman.com/canon#canon-good-spec-for-ai-agents): 改变了我给 agent 写规格的方式。 · 原文: https://addyo.substack.com/p/how-to-write-a-good-spec-for-ai-agents - [Building an AI-Native Engineering Team](https://aklman.com/canon#canon-build-ai-native-engineering-team): AI-native 团队怎么组织的参照。 · 原文: https://developers.openai.com/codex/guides/build-ai-native-engineering-team - [System Design Primer](https://aklman.com/canon#canon-system-design-primer): 系统设计的入门正典。 · 原文: https://github.com/donnemartin/system-design-primer - [OSSU — Computer Science](https://aklman.com/canon#canon-ossu-computer-science): 自学计算机的那条路。 · 原文: https://github.com/ossu/computer-science-cn - [Refactoring.Guru — Design Patterns](https://aklman.com/canon#canon-refactoring-guru-design-patterns): 第一次把设计模式讲得我真看懂了。 · 原文: https://refactoring.guru/design-patterns/ - [Addy Osmani](https://aklman.com/canon#canon-addy-osmani): 我追的 AI 时代工程实践第一信源。 · 原文: https://addyosmani.com/ - [Dave Cheney](https://aklman.com/canon#canon-dave-cheney): 学 Go 时最信得过的那个人。 · 原文: https://dave.cheney.net/ - [The Go Blog](https://aklman.com/canon#canon-the-go-blog): Go 语言的官方权威源。 · 原文: https://go.dev/blog/ ## Reviews - [三体](https://aklman.com/reviews/%E4%B8%89%E4%BD%93/llms.md): 刘慈欣把宇宙写成一场黑暗森林里的生存博弈。 - [艾尔登法环](https://aklman.com/reviews/%E8%89%BE%E5%B0%94%E7%99%BB%E6%B3%95%E7%8E%AF/llms.md): 魂系十年功力熔成的开放世界——只要肯跨过前期门槛,交界地值得每个人专程走一趟。殿堂级,逢人就推。 - [英雄联盟手游](https://aklman.com/reviews/%E8%8B%B1%E9%9B%84%E8%81%94%E7%9B%9F%E6%89%8B%E6%B8%B8/llms.md): 给想在手机上打真联盟的人:操作简单、还原度第一梯队,随时开一局都值——唯一要忍的是排位匹配磨人。 - [塞尔达传说 旷野之息](https://aklman.com/reviews/%E5%A1%9E%E5%B0%94%E8%BE%BE%E4%BC%A0%E8%AF%B4-%E6%97%B7%E9%87%8E%E4%B9%8B%E6%81%AF/llms.md): 任天堂把开放世界还给了好奇心。 - [IntelliJ IDEA](https://aklman.com/reviews/intellij-idea/llms.md): 功能和软件工程上没有替代品。 - [Google Chrome](https://aklman.com/reviews/google-chrome/llms.md): 浏览器母体,功能迭代适配有些慢。 - [Typora](https://aklman.com/reviews/typora/llms.md): Markdown 编辑器的角度非常好,但做笔记软件还差一点意思。 - [HapiGo](https://aklman.com/reviews/hapigo/llms.md): Raycast 的替代品,90% 的功能非常有用,我每天用十次以上。 - [带触控 ID 和数字小键盘的妙控键盘(黑色)](https://aklman.com/reviews/magic-keyboard/llms.md): 方便,但体验谈不上多好;触控 ID 在输入密码验证上带来一定便利。 - [MacBook Pro 16 英寸(M4 Pro)](https://aklman.com/reviews/macbook-pro/llms.md): 性能过剩,但续航和屏幕值回票价。 - [Lex Fridman Podcast](https://aklman.com/reviews/lex-fridman-podcast/llms.md): 三小时长谈把嘉宾聊到卸下防备。 - [ChatGPT](https://aklman.com/reviews/chatgpt/llms.md): 会一本正经地编造引用——它生成的是「最像回答的文本」,不是核验过的事实;流畅不等于真实,当检索器用之前先想清楚这一点。 - [Codex App](https://aklman.com/reviews/codex-app/llms.md): 值得看的不是写得快,而是把目标、上下文、执行、验证放进同一条开发流——「交付可信」这件事,它演示得最清楚。 - [Claude Code](https://aklman.com/reviews/claude-code/llms.md): 给用惯 IDE 的人换一条主干道——把目标说清楚它就能自己跑完。 - [Claude Desktop](https://aklman.com/reviews/claude-desktop/llms.md): 日常问答和 MCP 的家。 - [Warp](https://aklman.com/reviews/warp/llms.md): 终端里最好用的 AI 补全和块状输出。 - [Cursor](https://aklman.com/reviews/cursor/llms.md): AI 编辑器的标杆但我已少用。 - [Dia](https://aklman.com/reviews/dia/llms.md): 深入使用中。 - [Kimi Code](https://aklman.com/reviews/kimi-code/llms.md): 深入使用中。 - [Kimi Desktop](https://aklman.com/reviews/kimi-desktop/llms.md): 深入使用中。 - [Gemini CLI](https://aklman.com/reviews/gemini-cli/llms.md): 单纯工具层面做得不错,但结合自家模型拉胯。 - [Antigravity](https://aklman.com/reviews/antigravity/llms.md): CLI 用得很少,桌面端还可以、速度很快,但功能和体验上离 OpenAI 桌面端差距明显。 - [OpenCode](https://aklman.com/reviews/opencode/llms.md): 体验还是可以,Claude Code 和 Codex 的备选方案,但窗口管理不太习惯。 - [GitHub Copilot](https://aklman.com/reviews/github-copilot/llms.md): 辅助编程的祖师爷,体验没得说,只是额度太少了。 - [Monica](https://aklman.com/reviews/monica/llms.md): 功能和体验一绝,只是模型更新上架太慢了;用了三年,单就体验说我还是非常喜欢。 - [Manus](https://aklman.com/reviews/manus/llms.md): 功能和体验一绝,额度太少不划算。 - [OpenClaw](https://aklman.com/reviews/openclaw/llms.md): 暂还没深入使用。 - [Hermes Agent](https://aklman.com/reviews/hermes-agent/llms.md): 暂还没深入使用。 - [Kiro](https://aklman.com/reviews/kiro/llms.md): 暂还没深入使用。 - [Replit](https://aklman.com/reviews/replit/llms.md): 云上 vibe coding 体验非常好,但额度给得太少,还额外扣了 500 多元,一点都不划算。 - [Z.ai(Zcode)](https://aklman.com/reviews/z-ai-zcode/llms.md): 简单用了一下,总体设计还可以,但没有他们家的订阅就没有尝试的必要。 - [Claude](https://aklman.com/reviews/claude/llms.md): 截至 Fable 5——编码和长任务的天花板。 - [GPT](https://aklman.com/reviews/gpt/llms.md): 截至 5.6 Sol——编码和长任务的天花板仅次于 Fable,前端设计不如 Fable 5 和 K3。 - [Kimi](https://aklman.com/reviews/kimi/llms.md): 截至 K3——编码和长任务完全能胜任,前端天花板。 - [DeepSeek](https://aklman.com/reviews/deepseek/llms.md): 截至 V4 Flash——编码和长任务完全能胜任,欠缺多模态能力,前端设计一般。 - [Gemini](https://aklman.com/reviews/gemini/llms.md): 截至 3.6 Flash——编码和长任务的可用水平,前端设计还行。 - [GLM](https://aklman.com/reviews/glm/llms.md): 截至 GLM-5.2——编码和长任务完全能胜任,欠缺多模态能力,前端设计一般。 ## English - [Wiring an Agent into Your System](https://aklman.com/en/posts/ai/agents-in-existing-systems/llms.md): An agent is not a new system. It is a new client of the system you already run — treat it like one. - [Evaluation First, Then Everything Else](https://aklman.com/en/posts/ai/evaluation-first/llms.md): Without an eval set, every change is a hunch. With one, iteration finally means something. - [Context Matters More Than Prompts](https://aklman.com/en/posts/ai/context-over-prompts/llms.md): Prompt tuning hits a ceiling fast. What actually decides output quality is the material you feed in. - [Let the Model Do What It Is Good At](https://aklman.com/en/posts/ai/let-the-model-do-what-it-is-good-at/llms.md): Models are good at making vague things concrete. They are bad at guaranteeing that something happens. That boundary is where engineering lives. - [What is new in Go 1.15?](https://aklman.com/en/posts/repoers-and-reviews/go1.15-what-is-new/llms.md): A field note on Go 1.15, with editorial components for code, references, tabs, steps, and callouts. - [Seeker](https://aklman.com/en/works/seeker): A shell that holds AI, filled with mini-apps you can switch on and off. Conversation is the entry: say what you need, and the agent decides whether to query your data, produce a card, or run a skill you've saved — not another chat box, but a personal workbench that grows on your own data. - [Library](https://aklman.com/en/works/library): An independent AI e-book library: each book covers exactly one AI subscription you already pay for, or one workflow — what to learn, what to skip, when to switch tools. Free to read online, no signup wall. - [Putting Agents Into Real Workflows](https://aklman.com/en/sessions/build-with-agents/llms.md): A community session on reframing agents from tool lists into work systems made of context, roles, artifacts, and feedback. - [Codex App: From Task to Verified Change](https://aklman.com/en/sessions/codex-app-demo/llms.md): A product demo script for showing how goal, context, execution, and verification fit into one development flow. - [Redesigning Learning for the AI Age](https://aklman.com/en/sessions/learning-with-ai-video-script/llms.md): A long-form video script on reorganizing personal learning from acquisition to understanding, expansion, correction, and application. - [Canon — collection Markdown](https://aklman.com/en/canon/llms.md): Curated entries with attribution and original sources. - [The Three-Body Problem](https://aklman.com/en/reviews/%E4%B8%89%E4%BD%93/llms.md): 刘慈欣把宇宙写成一场黑暗森林里的生存博弈。 - [艾尔登法环](https://aklman.com/en/reviews/%E8%89%BE%E5%B0%94%E7%99%BB%E6%B3%95%E7%8E%AF/llms.md): 魂系十年功力熔成的开放世界——只要肯跨过前期门槛,交界地值得每个人专程走一趟。殿堂级,逢人就推。 - [英雄联盟手游](https://aklman.com/en/reviews/%E8%8B%B1%E9%9B%84%E8%81%94%E7%9B%9F%E6%89%8B%E6%B8%B8/llms.md): 给想在手机上打真联盟的人:操作简单、还原度第一梯队,随时开一局都值——唯一要忍的是排位匹配磨人。 - [塞尔达传说 旷野之息](https://aklman.com/en/reviews/%E5%A1%9E%E5%B0%94%E8%BE%BE%E4%BC%A0%E8%AF%B4-%E6%97%B7%E9%87%8E%E4%B9%8B%E6%81%AF/llms.md): 任天堂把开放世界还给了好奇心。 - [IntelliJ IDEA](https://aklman.com/en/reviews/intellij-idea/llms.md): 功能和软件工程上没有替代品。 - [Google Chrome](https://aklman.com/en/reviews/google-chrome/llms.md): 浏览器母体,功能迭代适配有些慢。 - [Typora](https://aklman.com/en/reviews/typora/llms.md): Markdown 编辑器的角度非常好,但做笔记软件还差一点意思。 - [HapiGo](https://aklman.com/en/reviews/hapigo/llms.md): Raycast 的替代品,90% 的功能非常有用,我每天用十次以上。 - [Magic Keyboard with Touch ID and Numeric Keypad](https://aklman.com/en/reviews/magic-keyboard/llms.md): 方便,但体验谈不上多好;触控 ID 在输入密码验证上带来一定便利。 - [MacBook Pro 16-inch (M4 Pro)](https://aklman.com/en/reviews/macbook-pro/llms.md): 性能过剩,但续航和屏幕值回票价。 - [Lex Fridman Podcast](https://aklman.com/en/reviews/lex-fridman-podcast/llms.md): 三小时长谈把嘉宾聊到卸下防备。 - [ChatGPT](https://aklman.com/en/reviews/chatgpt/llms.md): 会一本正经地编造引用——它生成的是「最像回答的文本」,不是核验过的事实;流畅不等于真实,当检索器用之前先想清楚这一点。 - [Codex App](https://aklman.com/en/reviews/codex-app/llms.md): 值得看的不是写得快,而是把目标、上下文、执行、验证放进同一条开发流——「交付可信」这件事,它演示得最清楚。 - [Claude Code](https://aklman.com/en/reviews/claude-code/llms.md): 给用惯 IDE 的人换一条主干道——把目标说清楚它就能自己跑完。 - [Claude Desktop](https://aklman.com/en/reviews/claude-desktop/llms.md): 日常问答和 MCP 的家。 - [Warp](https://aklman.com/en/reviews/warp/llms.md): 终端里最好用的 AI 补全和块状输出。 - [Cursor](https://aklman.com/en/reviews/cursor/llms.md): AI 编辑器的标杆但我已少用。 - [Dia](https://aklman.com/en/reviews/dia/llms.md): 深入使用中。 - [Kimi Code](https://aklman.com/en/reviews/kimi-code/llms.md): 深入使用中。 - [Kimi Desktop](https://aklman.com/en/reviews/kimi-desktop/llms.md): 深入使用中。 - [Gemini CLI](https://aklman.com/en/reviews/gemini-cli/llms.md): 单纯工具层面做得不错,但结合自家模型拉胯。 - [Antigravity](https://aklman.com/en/reviews/antigravity/llms.md): CLI 用得很少,桌面端还可以、速度很快,但功能和体验上离 OpenAI 桌面端差距明显。 - [OpenCode](https://aklman.com/en/reviews/opencode/llms.md): 体验还是可以,Claude Code 和 Codex 的备选方案,但窗口管理不太习惯。 - [GitHub Copilot](https://aklman.com/en/reviews/github-copilot/llms.md): 辅助编程的祖师爷,体验没得说,只是额度太少了。 - [Monica](https://aklman.com/en/reviews/monica/llms.md): 功能和体验一绝,只是模型更新上架太慢了;用了三年,单就体验说我还是非常喜欢。 - [Manus](https://aklman.com/en/reviews/manus/llms.md): 功能和体验一绝,额度太少不划算。 - [OpenClaw](https://aklman.com/en/reviews/openclaw/llms.md): 暂还没深入使用。 - [Hermes Agent](https://aklman.com/en/reviews/hermes-agent/llms.md): 暂还没深入使用。 - [Kiro](https://aklman.com/en/reviews/kiro/llms.md): 暂还没深入使用。 - [Replit](https://aklman.com/en/reviews/replit/llms.md): 云上 vibe coding 体验非常好,但额度给得太少,还额外扣了 500 多元,一点都不划算。 - [Z.ai(Zcode)](https://aklman.com/en/reviews/z-ai-zcode/llms.md): 简单用了一下,总体设计还可以,但没有他们家的订阅就没有尝试的必要。 - [Claude](https://aklman.com/en/reviews/claude/llms.md): 截至 Fable 5——编码和长任务的天花板。 - [GPT](https://aklman.com/en/reviews/gpt/llms.md): 截至 5.6 Sol——编码和长任务的天花板仅次于 Fable,前端设计不如 Fable 5 和 K3。 - [Kimi](https://aklman.com/en/reviews/kimi/llms.md): 截至 K3——编码和长任务完全能胜任,前端天花板。 - [DeepSeek](https://aklman.com/en/reviews/deepseek/llms.md): 截至 V4 Flash——编码和长任务完全能胜任,欠缺多模态能力,前端设计一般。 - [Gemini](https://aklman.com/en/reviews/gemini/llms.md): 截至 3.6 Flash——编码和长任务的可用水平,前端设计还行。 - [GLM](https://aklman.com/en/reviews/glm/llms.md): 截至 GLM-5.2——编码和长任务完全能胜任,欠缺多模态能力,前端设计一般。 - [学习即压缩:表征 / 图式 / 心智模型 / 解释框架](https://aklman.com/en/canon#canon-learning-is-compression): 重写了我对「学习」的定义——从此读任何东西都问「能拿走什么结构」。 · source: https://www.dedao.cn/course/article?id=qavDm012GolV7OYAjxVxVjEy8zdk73 - [The AI-Native Software Engineer](https://aklman.com/en/canon#canon-ai-native-software-engineer): 把「AI 时代的工程师」讲清楚了——和我现在做的这套东西同源。 · source: https://addyo.substack.com/p/the-ai-native-software-engineer - [How to Write a Good Spec for AI Agents](https://aklman.com/en/canon#canon-good-spec-for-ai-agents): 改变了我给 agent 写规格的方式。 · source: https://addyo.substack.com/p/how-to-write-a-good-spec-for-ai-agents - [Building an AI-Native Engineering Team](https://aklman.com/en/canon#canon-build-ai-native-engineering-team): AI-native 团队怎么组织的参照。 · source: https://developers.openai.com/codex/guides/build-ai-native-engineering-team - [System Design Primer](https://aklman.com/en/canon#canon-system-design-primer): 系统设计的入门正典。 · source: https://github.com/donnemartin/system-design-primer - [OSSU — Computer Science](https://aklman.com/en/canon#canon-ossu-computer-science): 自学计算机的那条路。 · source: https://github.com/ossu/computer-science-cn - [Refactoring.Guru — Design Patterns](https://aklman.com/en/canon#canon-refactoring-guru-design-patterns): 第一次把设计模式讲得我真看懂了。 · source: https://refactoring.guru/design-patterns/ - [Addy Osmani](https://aklman.com/en/canon#canon-addy-osmani): 我追的 AI 时代工程实践第一信源。 · source: https://addyosmani.com/ - [Dave Cheney](https://aklman.com/en/canon#canon-dave-cheney): 学 Go 时最信得过的那个人。 · source: https://dave.cheney.net/ - [The Go Blog](https://aklman.com/en/canon#canon-the-go-blog): Go 语言的官方权威源。 · source: https://go.dev/blog/