Matt Van Horn 的 22 个 Agentic Engineering Hacks

来源:Matt Van Horn @mvanhorn 2026-06-03 长文(82万 浏览 / 2,768 喜欢 / 8,511 书签,2026-06-07 抓取) 原文:https://x.com/mvanhorn/status/2061877533885473181 关联:[[claude-code-dynamic-workflows]] · [[Claude-Code架构深度解读-Agent系统的真正护城河不在模型-而在-Harness]] · [[Codex才是最适合普通人的顶级牛马-Agent]] · [[every-agentic-engineering-hack-2026-06-digest|速读摘要]]


一句话主张

No IDE. Just plan.md files and voice.

Agentic Engineering 的全部工作流是:语音输入 → plan.md 沉淀 → agent 执行。3 月那篇 91.3万 浏览的 “Every Claude Code Hack” 是 1.0 版,这是 6 月 2.0 版。


5 大核心原则

1. Plan 是缰绳,不是读物

2. 思考进计划,执行机械化

3. Research → Plan → Build 标准循环

4. 多 session 并发 + Human Signal

5. 现实警告:AI Psychosis


22 个 Hack 速查表

# Hack 工具 / 命令 一句话
1 CE plan.md 第一时间 /ce-plan /ce-work /ce-brainstorm 想到 idea 立刻出 plan,80% 计划 + 20% 编码
2 Don’t Read the plan.md inline “TLDR?” / “eli5” 计划是给 agent 的缰绳,别坐下来读 300 行
3 Plan for the Plan /ce-plan make a plan for the plan 防止 LLM 偷懒的万能招:先 plan 再 plan
4 Get Voice-Pilled Monologue / Wispr Flow / Apple dictation 语音转 LLM 不需要完美转写
5 Lots of Tabs in cmux cmux 4-6 tab 并行,每个跑不同任务
6 Terminal 默认进 Claude Ghostty config + launcher.sh 新 tab 直接进 Claude Code
7 Remote Control + Email Claude mobile + AgentMail Claude 可以收邮件触发新 session
8 Skip Permissions YOLO skipDangerousModePermissionPrompt: true + 声音 hook It’s my computer. YOLO.
9 Codex 不开 CLI Codex IDE 扩展 + /ce-work --codex Claude plans, Codex builds
10 last30days 跑研究 /last30days <topic> 决策前先跑多源新鲜研究(26K stars)
11 Granola Raw Transcript Granola + Claude 整段生 transcript 丢给 LLM,不先总结
12 Human Signal 你的 judgment 你是品味、方向、react-and-redirect loop
13 HyperFrames 做视频 HyperFrames + agent 视频 HTML 化,agent 写 script.md 渲染 MP4
14 Notes as Agent KB Bear / Obsidian / gbrain / supermemory 选有 CLI/API 的笔记,让 agent 能读
15 Mac mini 远程工作 Mosh + Tmux + Hermes + OpenClaw Mosh 远端不卡,Tmux 飞机不掉
16 Proof 给人看 plan Proof plan.md 丢进 Proof,非终端人也能评论
17 Write Your Own Skills “look at CE skill and help me make one like this” 做超过 2 次的事做成 skill
18 Contribute to OSS /ce-plan + /ce-work 循环 用同一套循环提 PR 给 Python/Go/Vercel Agent Browser
19 M5 Max + 64GB M5 Max 笔记本 6 session 起步配置,Anker 充电宝救命
20 Printing Press + Agent Cookie printingpress.dev 把现实服务包成 agent-native CLI,自动登录
21 AI Psychosis mindset 陷阱是沉迷 build 失去人,不是空 launch
22 本文写作方式 cmux + Monologue + Proof No IDE. No typing code. Talk, plan, build.

关键配置代码(开箱即用)

~/.claude/settings.json — YOLO + 声音 hook

1
2
3
4
5
6
7
8
9
{
  "permissions": {
    "allow": ["WebSearch", "WebFetch", "Bash", "Read", "Write", "Edit", "Glob", "Grep", "Task", "TodoWrite"],
    "deny": [],
    "defaultMode": "bypassPermissions"
  },
  "skipDangerousModePermissionPrompt": true,
  "remoteControlAtStartup": true
}
1
2
3
4
5
{
  "hooks": {
    "Stop": [{ "hooks": [{ "type": "command", "command": "afplay /System/Library/Sounds/Blow.aiff" }] }]
  }
}

~/.codex/config.toml — Codex YOLO

1
2
approval_policy = "never"
sandbox_mode = "danger-full-access"

Ghostty + Claude launcher

1
2
3
# ~/.local/bin/claude-launcher.sh
claude --dangerously-skip-permissions
# when Claude exits: print short note, drop into login zsh

工具栈生态速查

工具 角色 关键能力 Stars
Compound Engineering 工作流核心 /ce-plan /ce-work /ce-brainstorm
last30days 研究前置 多源并行抓 30 天新鲜内容 27K
Printing Press CLI 工厂 任意服务 → agent-native CLI 3.7K
Agent Cookie 浏览器 session 同步 免密码登录,CLIs 直接以你身份行事
AgentMail Claude 收件箱 email → 新 session(WebSocket + DKIM/SPF)
cmux 终端多 tab 4-6 个独立 session
Monologue / Wispr Flow 语音转 LLM Mac 端语音输入任意 app
Proof 计划评审 plan.md → 可分享可评论的文档
HyperFrames 视频生成 视频 HTML 化,agent 渲染 MP4
Hermes / OpenClaw 自主 agent 生态 远程自主工作
Mosh + Tmux 远程连接 飞机/差旅不掉 session

反直觉的金句

“It’s also the single best trick I know for making an LLM not lazy. Ask for the deliverable directly and it cuts corners. Ask it to first plan how it will produce the deliverable, then execute that plan, and it does the deep version every time.”

—— 防止 LLM 偷懒的万能招:Plan for the Plan

“The reason a plan gets better every time is that Claude has access to every prior plan I’ve written. Compounding context.”

—— 计划 + 笔记 = Compounding context

“Not ‘AI writes my code.’ Agentic Engineering does the errands, watches the game, warms the car, and books the trip, while I’m doing something else.”

—— AI Coding 的下一站:AI Errands

“The trap isn’t the empty launch, it’s vanishing into the build and losing the people around you.”

—— AI Psychosis 的真正危险


关联阅读


标签

#主题/AI-Coding #主题/AI-Agent #主题/效率 #主题/APP研发 #场景/技术博客 #场景/落地案例 #手法/案例驱动 #手法/权威背书 #手法/对比冲突 #手法/焦虑共鸣