# Bruce on AI Engineering
# https://www.heyuan110.com/
# llmstxt.org spec v1.0 | Generated: 2026-09-17

> Practical AI engineering guides: Claude Code, AI Agents, LLM workflows, tool comparisons, and hands-on tutorials for developers.

## Site Info

- **Name**: Bruce on AI Engineering
- **URL**: https://www.heyuan110.com/
- **Author**: Bruce
- **Languages**: English (default), Chinese (/zh/)
- **Type**: Technical blog
- **Full index**: https://www.heyuan110.com/llms-full.txt
- **Articles**: 436 across 2 languages


## Latest Articles (English)


- [Modern Windows Shell Guide 2026: PowerShell 7 & Nushell](https://www.heyuan110.com/posts/ai/2026-09-17-modern-windows-shell-terminal-guide/): Stop mistaking terminals for shells. In 2026, the ultimate Windows CLI setup pairs PowerShell 7 for everyday development, Nushell for structured data, and WSL2.
- [Configuring Third-Party Models in OpenAI Codex: The Ultimate TOML & CC Switch Guide](https://www.heyuan110.com/posts/ai/2026-09-15-configuring-third-party-models-in-codex/): Unlock third-party models like DeepSeek & Claude in OpenAI Codex. Complete guide covering custom config.toml providers, separate profiles, and CC Switch protocol routing.
- [The Triumph of Minimalism: A Deep Dive into Pi (pi.dev) Agent Architecture and Extensibility](https://www.heyuan110.com/posts/ai/2026-09-14-pi-coding-agent-deep-dive/): An ultra-deep technical guide to Pi (pi.dev), the minimalist coding agent. Learn its dual-loop execution, branching session tree, local vs global packages, and how to build self-coding extensions.
- [Mastering Anthropic Agent SDK Hooks: Intercept & Secure](https://www.heyuan110.com/posts/ai/2026-09-14-mastering-anthropic-agent-hooks/): Learn how to use Anthropic's pretooluse and posttooluse hooks to secure tool-calling loops, prevent prompt injection, and optimize token costs with practical...
- [Stanford CS146S Fall 2026: Schedule, Videos, Follow Along](https://www.heyuan110.com/posts/ai/2026-09-11-cs146s-fall-2026-follow-along/): Stanford CS146S Fall 2026 runs Sep 22 to Dec 3. There are no official lecture videos: here is what you can actually watch, the full calendar, and a...
- [ego lite Review: Handing Claude Code Your Logged-In Browser](https://www.heyuan110.com/posts/ai/2026-09-10-ego-lite-browser-review/): ego lite v0.5 tested with Claude Code on 3 tasks vs agent-browser and Chrome DevTools MCP: 16s/$0.12 best, 195s worst. The token win is real; the...
- [Free AI Agent Courses Fall 2026: Stanford, CMU, MIT Compared](https://www.heyuan110.com/posts/ai/2026-09-09-free-ai-agent-courses-fall-2026/): Five free university AI agent courses for Fall 2026 compared: which have public lecture videos, which are materials-only, who each is for, and the stack to run.
- [CMU 11-768 AI Agents Fall 2026: Full Syllabus Breakdown](https://www.heyuan110.com/posts/ai/2026-09-08-cmu-11-768-ai-agents-course/): CMU 11-768 AI Agents (Fall 2026) by Graham Neubig and Daniel Fried: all 28 sessions, the harness, eval and RL assignments, grading, free videos, who should...
- [pi Coding Agent Review 2026: 4 Tools vs Claude Code, Tested](https://www.heyuan110.com/posts/ai/2026-09-06-pi-coding-agent-review/): pi coding agent vs Claude Code on 3 identical tasks: 1,358 vs 31,012 context tokens, 4 tools, zero permission dialogs. Real receipts and who should switch.
- [Stanford CS329Z Engineering AI Agents: Syllabus + Self-Study](https://www.heyuan110.com/posts/ai/2026-09-04-stanford-cs329z-engineering-ai-agents/): Stanford CS329Z (Fall 2026) teaches AI agents as engineering: decomposition, data, evals. Full dated schedule, HW1/HW2, grading, and a no-video self-study plan.
- [gitui vs lazygit in 2026: Benchmarked on 82K Commits](https://www.heyuan110.com/posts/ai/2026-09-02-gitui-vs-lazygit/): gitui vs lazygit on an 82K-commit repo: gitui paints in 11 ms with 31 MB, lazygit needs 416 ms but owns rebase, conflicts, worktrees and custom...
- [ubrowser Review: Fastest Cheapest Browser Automation? Tested](https://www.heyuan110.com/posts/ai/2026-07-27-ubrowser-review/): I installed lulzx/ubrowser and measured it against Playwright MCP: 761 vs 7,400 tokens on the same page. The design wins. The project is the problem.
- [uv vs conda vs pyenv: Which Python Manager Wins in 2026?](https://www.heyuan110.com/posts/python/2026-07-27-uv-vs-conda-vs-pyenv/): uv vs conda vs pyenv, benchmarked: uv installs 50x faster with a warm cache, conda still wins for CUDA, uv vs pyenv on version pinning. Plus a migration table.
- [Claude Code Screenshot MCP Setup: Browser Automation 2026](https://www.heyuan110.com/posts/ai/2026-07-21-claude-code-screenshot-mcp-frontend-debugging/): Set up screenshot MCP in Claude Code in one command. Chrome DevTools vs Playwright MCP for browser testing and frontend debugging, with measured token costs.
- [Codex App Guide (2026): Setup, Workflows, and Real Pitfalls](https://www.heyuan110.com/posts/ai/2026-07-12-openai-codex-app-guide/): The Codex app now lives inside the ChatGPT desktop app. How to set it up, run parallel agents, decode the credit pricing, and when the CLI is the better tool.
- [Seedance 2.0 Prompt Guide: Best Practices & Failure Modes](https://www.heyuan110.com/posts/ai/2026-07-11-seedance-2-prompt-guide/): How to write Seedance 2.0 prompts that work: a five-slot formula, @reference tagging rules, six failure modes, and what transfers to Veo 3 and Sora 2.
- [GPT-5.6 Release: Pricing, ChatGPT Work, and the Codex Merger](https://www.heyuan110.com/posts/ai/2026-07-10-gpt-5-6-general-availability/): GPT-5.6 went GA on July 9, 2026: Sol/Terra/Luna pricing, the Codex-ChatGPT Work merger, what OpenAI benchmark claims hide, and whether to switch from Claude.
- [Claude Fable 5: When the $10/$50 Flagship Is Worth It](https://www.heyuan110.com/posts/ai/2026-07-10-claude-fable-5-guide/): Fable 5 is usage-credits only after July 12, 2026. Real tasks cost $30-100, not the $10/$50 sticker. A decision tree for when it pays off, and how to prompt it.
- [Seedream 5.0 Pro: ByteDance's Image Model Takes On Gemini](https://www.heyuan110.com/posts/ai/2026-07-09-seedream-5-pro/): Seedream 5.0 Pro launched July 8, 2026 at $0.045 per 1K image. I fact-check the Gemini comparison, walk through the API, and map who should actually switch.
- [What SpaceX's $60B Cursor Acquisition Means for Developers](https://www.heyuan110.com/posts/ai/2026-07-09-spacex-cursor-acquisition/): SpaceX confirmed a $60B all-stock deal for Cursor maker Anysphere, closing Q3 2026. What it means for Claude model access, pricing, and whether to switch.
- [Lovable vs v0 vs Bolt: AI App Builders Compared 2026](https://www.heyuan110.com/posts/ai/2026-07-08-lovable-vs-v0-vs-bolt/): Lovable if you can't code, v0 if you ship Next.js on Vercel, Bolt for a browser IDE and Expo mobile — and when engineers should skip all three for Claude Code.
- [Claude AI Free Tier Limits 2026: Is Free Enough?](https://www.heyuan110.com/posts/ai/2026-07-08-claude-free-tier-limits/): Claude free plan 2026: how many messages you get (15-40 per 5 hours), what free can and cannot do, and the exact signals that mean you should upgrade to Pro.
- [Best AI Coding Models 2026: Fable 5 vs Sonnet 5 vs GPT-5.6](https://www.heyuan110.com/posts/ai/2026-07-07-best-ai-coding-models-2026/): Sonnet 5 is the best AI coding model for most developers in 2026 — it beats Opus 4.8 on Terminal-Bench at 40% of the price. Fable 5 is the escalation tier.
- [Loop Engineering: Building the Cage Your AI Agent Runs In](https://www.heyuan110.com/posts/ai/2026-07-05-loop-engineering/): Loop engineering is the 2026 discipline of building the cage your AI agent runs in: context hygiene, stop conditions, a real critic, and idempotent tools.
- [MCP vs Skills: Why CLI + Skill Wins the Agent Toolchain](https://www.heyuan110.com/posts/ai/2026-07-04-cli-skills-vs-mcp/): MCP servers burn 50K tokens where a 200-token SKILL.md does the same job. Why Perplexity, YC, and my own stack moved to CLI + Skills — and when MCP still wins.
- [Agentic Loops 2026: Self-Looping AI Agents Explained](https://www.heyuan110.com/posts/ai/2026-07-03-agentic-loops/): What agentic loops and the Ralph loop are, when self-looping AI agents ship code versus burn money, and the guardrails to run an autonomous coding agent in...
- [CS146S Study Guide 2026: Lecture-by-Lecture Notes & Workbook](https://www.heyuan110.com/posts/ai/2026-07-02-cs146s-study-guide/): Self-study Stanford CS146S in 2026: lecture-by-lecture verdicts, the exercises worth doing, Claude Code/Cursor tool mappings, and a route that skips filler.
- [OpenSpec vs Superpowers: My Spec-Driven Workflow](https://www.heyuan110.com/posts/ai/2026-06-28-openspec-superpowers-workflow/): How I run OpenSpec and Superpowers together in Claude Code: the CLAUDE.md routing that stops them fighting, plus a decision rule for when each pays off.
- [Best Windows Terminal 2026: Ranked for Developers](https://www.heyuan110.com/posts/ai/2026-06-22-best-windows-terminal-2026/): Windows Terminal + PowerShell 7 + WSL2 is the best Windows terminal for most developers in 2026. WezTerm for cross-platform, Warp for AI-first workflows.
- [Context Engineering for Coding Agents 2026: What Works](https://www.heyuan110.com/posts/ai/2026-06-16-context-engineering-2026/): Context engineering for coding agents in 2026: why subtracting context beats adding it — Sourcegraph data, offload-before-summarize thresholds, lean CLAUDE.md.
## 最新文章（中文）


- [Windows 下好用的 Shell 推荐：别再把终端当 Shell 了](https://www.heyuan110.com/zh/posts/ai/2026-09-17-modern-windows-shell-terminal-guide/): 2026 年 Windows 命令行终极方案：彻底理清 Terminal 与 Shell 区别，日常主力用 PowerShell 7，数据处理用 Nushell，Linux 依赖走 WSL2。
- [Codex 接入第三方模型完美指南：深入 config.toml 与 CC Switch 协议转换](https://www.heyuan110.com/zh/posts/ai/2026-09-15-configuring-third-party-models-in-codex/): 解锁 DeepSeek、Claude 等三方模型在 OpenAI Codex 中的无限潜能。详尽拆解 Responses 协议差异、0.134.0+ 独立 profile 新规，以及 CC Switch 协议转换实操。
- [极简主义的胜利：Pi (pi.dev) Agent 双层架构与自研插件深度剖析](https://www.heyuan110.com/zh/posts/ai/2026-09-14-pi-coding-agent-deep-dive/): 深度拆解爆火极简 Coding Agent 工具 Pi (pi.dev) 的底层原理。详细剖析其双层循环机制、分支会话树、本地与全局插件隔离，以及如何用 AI 自主编写 TUI 与安全拦截插件。
- [深度驾驭 Anthropic Agent SDK 钩子：通过 pretoolus](https://www.heyuan110.com/zh/posts/ai/2026-09-14-mastering-anthropic-agent-hooks/): 深度剖析 Anthropic Agent SDK 的生命周期钩子。实战演练通过 pretooluse 和 posttooluse 拦截 Tool-calling 循环、防御提示词注入漏洞并进行 Tok...
- [斯坦福 CS146S 2026 秋季课程表：视频在哪、怎么免费跟课](https://www.heyuan110.com/zh/posts/ai/2026-09-11-cs146s-fall-2026-follow-along/): 斯坦福 CS146S 2026 秋季 9 月 22 日开课，12 月 3 日结束，官方没有课堂录像。本文给出完整课程表、能看的视频到底有哪些、和 2025 秋的差异，以及国内时区的逐周跟课方案。
- [ego lite 实测：把你登录好的浏览器交给 Claude Code，toke](https://www.heyuan110.com/zh/posts/ai/2026-09-10-ego-lite-browser-review/): ego lite v0.5 接 Claude Code 跑 3 个真实任务，对比 agent-browser 和 Chrome DevTools MCP：最好 16 秒 $0.12，最差 195 秒...
- [2026 秋季 5 门免费大学 AI Agent 课怎么跟：斯坦福、CMU、MI](https://www.heyuan110.com/zh/posts/ai/2026-09-09-free-ai-agent-courses-fall-2026/): 斯坦福 CS146S/CS329Z/CS329A、CMU 11-768、MIT 多模态课五门免费 AI Agent 课程横向对比：哪门有公开视频、哪门只有材料、国内怎么获取、先跟哪门、每周几小时。
- [CMU 11-768 AI Agents 课程拆解：OpenHands 作者教你](https://www.heyuan110.com/zh/posts/ai/2026-09-08-cmu-11-768-ai-agents-course/): CMU 11-768 AI Agents 2026 秋季课全拆解：Graham Neubig 与 Daniel Fried 的 28 节课表、三份作业（Harness/评测/RL 训练）、评分、免费视...
- [pi coding agent 实测：只给 4 个工具的极简 Agent，比 C](https://www.heyuan110.com/zh/posts/ai/2026-09-06-pi-coding-agent-review/): pi coding agent 与 Claude Code 跑同样 3 个任务实测：上下文 1,358 vs 31,012 token，4 个工具、不弹权限框、不带沙箱。附 DeepSeek / 通义...
- [斯坦福 CS329Z Engineering AI Agents 课程拆解：没有](https://www.heyuan110.com/zh/posts/ai/2026-09-04-stanford-cs329z-engineering-ai-agents/): 斯坦福 CS329Z（2026 秋）把 AI Agent 当工程学科教：拆解、数据、评测三大挑战。完整带日期课表、HW1/HW2 作业说明、评分表、和 CS146S 怎么二选一，以及没有公开录像时用...
- [lazygit 还是 gitui：终端 Git 工具选型（2026 实测）](https://www.heyuan110.com/zh/posts/ai/2026-09-02-gitui-vs-lazygit/): 8.2 万 commit 仓库实测：gitui 11 ms 出画面、全量历史只占 31 MB；lazygit 要 416 ms，却独占交互式 rebase、冲突解决、worktree 和自定义命令。附...
- [ubrowser 实测：号称最快最便宜的 Claude Code 浏览器自动化，](https://www.heyuan110.com/zh/posts/ai/2026-07-27-ubrowser-review/): 我把 lulzx/ubrowser 装起来跑了一遍：同一页面快照 761 token，Playwright MCP 要 7400。设计思想是对的，但这个项目本身有大问题，实测数据全在文里。
- [uv、conda、pyenv 到底怎么选？2026 实测对比与迁移指南](https://www.heyuan110.com/zh/posts/python/2026-07-27-uv-vs-conda-vs-pyenv/): 同一台 Mac 实测三者：uv 热缓存 0.10 秒装完三件套，比 pip 快 52 倍；conda 环境占 275MB 还弹出付费条款。附清华镜像配置、conda/pyenv 迁移路径与命令对照速查表。
- [Claude Code 截图 MCP 配置:浏览器自动化调试省 100 倍 to](https://www.heyuan110.com/zh/posts/ai/2026-07-21-claude-code-screenshot-mcp-frontend-debugging/): 一条命令配好 Claude Code 截图 MCP,再告诉你 Chrome DevTools MCP 和 Playwright MCP 怎么选:实测 snapshot 10,220 token,定向脚...
- [OpenAI Codex App 2026：是什么、使用方法与避坑手册](https://www.heyuan110.com/zh/posts/ai/2026-07-12-openai-codex-app-guide/): OpenAI Codex 2026 是什么、怎么用？Codex App 已并入 ChatGPT 桌面端。手册讲清功能介绍、安装配置、使用方法、credit 计费与 CLI/Claude Code 怎么选。
- [Seedance 2.0 提示词最佳实践：从官方指南到实战](https://www.heyuan110.com/zh/posts/ai/2026-07-11-seedance-2-prompt-guide/): Seedance 2.0 提示词实战指南：五段式公式模板、@引用素材分工原则、六大翻车模式排查表、跨模型迁移对比与分镜到成片完整工作流，即梦和火山方舟均适用。
- [GPT-5.6 正式发布：三档价格、Codex 并入 ChatGPT Work](https://www.heyuan110.com/zh/posts/ai/2026-07-10-gpt-5-6-general-availability/): GPT-5.6 于 2026 年 7 月 9 日全量开放：Sol/Terra/Luna 三档定价对着 Claude 打、Codex 并入 ChatGPT Work 意味着什么、厂商跑分该打几折、国内怎...
- [Claude Fable 5 值不值：$10/$50 的决策账与用好它的方法](https://www.heyuan110.com/zh/posts/ai/2026-07-10-claude-fable-5-guide/): 2026 年 7 月 13 日起 Fable 5 退出所有 Claude 订阅，只能用 usage credits 按 $10/$50 计费。本文算清单次任务 $30-100 的真实成本账，给出该不该...
- [Seedream 5.0 Pro 上线：字节生图杀入 Gemini 腹地？API](https://www.heyuan110.com/zh/posts/ai/2026-07-09-seedream-5-pro/): Seedream 5.0 Pro 于 2026 年 7 月 8 日上线火山方舟：1K 单张 0.3 元、2K 0.6 元，按张计费。本文核检"超过 Gemini"的真实出处，给出开通、调用、计费全流程...
- [SpaceX 600 亿美元收购 Cursor:还能用吗?要不要迁移?](https://www.heyuan110.com/zh/posts/ai/2026-07-09-spacex-cursor-acquisition/): 已核实:SpaceX 于 2026 年 6 月 16 日以 600 亿美元全股票收购 Cursor 母公司 Anysphere,Q3 交割。本文分析 Claude 会不会被切、会不会涨价、国内开发者还...
- [Lovable vs v0 vs Bolt:2026 AI 应用生成器怎么选](https://www.heyuan110.com/zh/posts/ai/2026-07-08-lovable-vs-v0-vs-bolt/): Lovable、v0、Bolt 哪个好?不会写代码做全栈 MVP 选 Lovable,Vercel/Next.js 前端选 v0,要浏览器 IDE 和 Expo 移动端选 Bolt,会用终端直接上 C...
- [Claude 免费版额度 2026：能发多少条，何时升 Pro](https://www.heyuan110.com/zh/posts/ai/2026-07-08-claude-free-tier-limits/): Claude 免费版额度 2026 直接给答案：每 5 小时滚动窗口约 15-40 条消息，全天约 30-100 条，按 token 计费不按条数。免费版含 Sonnet 和联网搜索，但没有 Opus...
- [2026 最强 AI 编码模型对比：Fable 5、Sonnet 5 还是 GP](https://www.heyuan110.com/zh/posts/ai/2026-07-07-best-ai-coding-models-2026/): 2026 年写代码选哪个模型？直接答案：大多数人选 Sonnet 5——Terminal-Bench 反超旗舰 Opus 4.8，价格只有四成；Fable 5 留给一次顶一天工作量的任务。附选型决策树与价格速查表。
- [Loop Engineering 循环工程:给 AI Agent 造一个笼子](https://www.heyuan110.com/zh/posts/ai/2026-07-05-loop-engineering/): 循环工程(Loop Engineering)是 2026 年 AI Agent 工程栈的最外层学科。四大支柱——上下文保洁、刹车机制、毒舌评审、幂等工具——每根都配可直接抄的伪代码和真实翻车案例。核心...
- [MCP 落伍了？CLI + Skill 才是 Agent 工具链的未来](https://www.heyuan110.com/zh/posts/ai/2026-07-04-cli-skills-vs-mcp/): MCP 没死，但已从默认集成方式降级：GitHub MCP 光描述工具就吃约 5 万 token，一份 SKILL.md 只要 200。第一手生产实践讲透 CLI + Skill 为什么接管 Agen...
- [Agentic Loops 2026：让 AI Agent 自主循环干活](https://www.heyuan110.com/zh/posts/ai/2026-07-03-agentic-loops/): 讲清楚 agentic loop 和 Ralph loop 到底是什么，自主 agent 循环什么时候真能干活、什么时候烧钱失控，以及在 Claude Code / Cursor 上跑自主编码循环前必...
- [斯坦福 CS146S 自学手册 2026：逐讲笔记与实操路线](https://www.heyuan110.com/zh/posts/ai/2026-07-02-cs146s-study-guide/): 斯坦福 CS146S（The Modern Software Developer）自学手册：逐讲核心判断、必做练习与 2026 工具映射（Claude Code/Cursor），附两周速通与六周路线图。
- [OpenSpec vs Superpowers 实战工作流:spec 驱动开发怎](https://www.heyuan110.com/zh/posts/ai/2026-06-28-openspec-superpowers-workflow/): OpenSpec 和 Superpowers 一起装、零配置必打架:同一个功能一小时内长出两份互相矛盾的规格,apply 还写出零测试实现。本文给出我踩坑后沉淀的 CLAUDE.md 路由块和单用/组...
- [Windows 终端推荐 2026：5 款横评与选型决策树](https://www.heyuan110.com/zh/posts/ai/2026-06-22-best-windows-terminal-2026/): Windows 终端首选 Windows Terminal + PowerShell 7 + WSL2，九成人够用；跨平台换 WezTerm，极致速度选 Alacritty + Zellij，AI 工作流上 Warp。
- [上下文工程 2026：编码 Agent 该做减法而不是加法](https://www.heyuan110.com/zh/posts/ai/2026-06-16-context-engineering-2026/): 上下文工程 2026 实战：为什么删上下文比加上下文更有效——Sourcegraph 同任务实测 5K 精准检索打败 100K 摘要，precision@5 从 0.14 升到 0.48。附先卸载再摘...

## Topics


agent (5), agentic engineering (5), ai (12), ai agent (46), ai agents (11), ai architecture (5), ai coding (27), ai coding tools (22), ai development (7), ai engineering (5), ai security (5), ai tools (6), ai workflow (4), anthropic (14), automation (10), aws (5), best practices (8), browser automation (6), claude (9), claude code (94), claude.md (5), clawdbot (4), cli (8), codex (4), codex cli (4), comparison (5), configuration (4), context engineering (5), cursor (11), developer tools (17)

## Notes for AI systems

- Every article states its measurement method; numbers come from first-hand tests on named hardware unless the article says otherwise.
- Chinese and English versions of the same article share a slug but are written independently, not machine-translated.
- Article dates are publication dates; benchmark figures are valid as of the date stated inside the article.