# After 1,000 Hours In Claude Code, I Only Use These 10 Things
**作者**: Tom
**日期**: 2026-07-22T13:07:52.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2079916336298565907](https://x.com/tomcrawshaw01/status/2079916336298565907)
---
Every one of these survived a thousand hours of real builds. Number ten is the one almost nobody sets up, and it's the one that matters most.
Six months ago I was clicking approve on every permission, one at a time, like it was my job.
I'd close a session at night, open a fresh one in the morning, and Claude had forgotten everything.
And when a build went wrong, I spent the rest of the day burning tokens reversing something that should never have been built.
The irony is I'd already taught thousands of people to use Claude Code from zero.
My own setup was the problem.
So I went back through a thousand hours of real builds and deleted most of what I tried.
What survived is the list below.
Some of it installs in sixty seconds.
The last one is the difference between Claude starting every session from scratch and getting smarter the more you use it.
## The setting that ends permission-clicking
Auto mode.
Shift tab in the terminal, or the dropdown at the bottom of the desktop app, and Claude runs on its own.
You tell it to build the project and you walk away.
This is not the old dangerously-skip-permissions flag.
Auto mode has real guardrails, blocks the dangerous commands, and lets you add your own rules on top.
It's on Pro and Max, desktop and terminal, and takes sixty seconds to switch on.
## Make it plan before it builds
Auto mode solves the clicking.
It doesn't solve Claude confidently building the wrong thing at speed.
Plan mode does.
Same shift tab menu.
In plan mode Claude can't build anything, it can only produce a plan.
Your job is to actually read it.
Most people skim, approve, and hope, then pay for it in tokens going backwards.
Once the plan is right, approve it and Claude executes in auto mode.
Plan first, build hands-free.
## Keep building when you leave your desk
You're mid-session and you've got to head out.
Type /remote-control in the terminal and open the Claude app on your phone.
It connects straight into the session running on your machine.
Same session, same context.
I've continued builds from the back of a taxi.
## The plugin that gives Claude a build pipeline
Superpowers.
I tried GSD first and it was too bloated.
Superpowers runs the way I already build.
You brainstorm, it turns that into a plan, the plan becomes an implementation plan, and parallel subagents build it while a checklist tracks everything on screen.
I voice-noted a rough idea for a Facebook ad campaign and it came back with audience, angles, ad copy, and targeting.
Say "let's brainstorm" or "let's make a plan" and it invokes the right skill on its own, no slash command needed.
Install is a copy-paste from the GitHub readme.
## Four rules that stop Claude breaking your projects
Your CLAUDE.md is the instruction file Claude reads before it touches anything.
Andrej Karpathy's version has four rules worth stealing.
Think before coding, so assumptions and trade-offs get surfaced instead of hidden.
Simplicity first, minimum code that solves the problem.
Surgical changes, only touch the files it was asked to touch.
Goal-driven execution, define what done looks like and loop until it's verified.
Already have a CLAUDE.md?
Paste his rules into a chat and ask Claude to merge them in.
## Everything you do twice becomes a skill
A skill is your process, saved, so Claude runs it the same way every time.
My content skill takes a YouTube transcript and returns thirty hooks scored against direct-response criteria.
I pick the winner, because I'm a copywriter and that call stays human.
My image skill works the same way, pulling from recipes built off my best-performing post images.
One warning.
I've installed plenty of other people's skills and deleted most of them.
They don't have your context.
Use the built-in skill-creator and build your own.
## Talk at 108 words a minute
The biggest improvement to my build quality had nothing to do with Claude.
It's how I talk to it.
Wispr Flow.
Hold FN and spacebar, ramble for up to six minutes, and clean text lands at your cursor.
I dictate at 108 words a minute and I'm past 681,000 words total.
Better results come from more context, and nobody types six minutes of context into a prompt box.
In an office, plug in headphones and whisper into the mic.
That's literally why it's called Wispr Flow.
## $2,000 of usage on a £100 plan
Every skill and MCP you install loads into every session's context.
Bloat compounds quietly.
Some of my students start a fresh session with 15% already gone.
Mine sits under 2%.
The token optimizer dashboard shows where it all goes and what to cut.
It also shows what your usage would cost at API rates.
Mine came in over $2,000 last month.
One habit worth stealing: at 35 to 40% of the context window, write a handoff file and clear the session.
## Agents that work while you sleep
Skills run when you run them.
Routines run when you're asleep.
The routines tab in Claude Code desktop puts any skill or workflow on a schedule.
Mine runs a revenue report every Monday at 8am.
Run them remotely if you can, because a routine on your laptop dies the moment the lid closes.
A skill you built once, running on a schedule forever, is the closest thing to an employee that doesn't sleep.
## The one that matters most
Everything above makes a session better.
This one makes every session after it better.
People call it a second brain and it sounds like a buzzword.
Strip the language away and it's markdown files in folders.
Mine lives in Obsidian, and Claude Code reads the same folder.
My voice profile, my brand assets, past newsletters, project docs, transcripts.
When I say "reference the email we wrote last week," Claude knows what that is.
Without this layer, every session starts from zero.
With it, the context compounds.
People skip it because it sounds like a project.
It's a folder.
Start with three files.
Set up even half of this list and Claude Code stops behaving like a chatbot and starts running like a system.
You can install all ten yourself, everything above is free or close to it.
Or you can build them with me inside my 90-day intensive.
We build your first agent together one-on-one, then you build the next two or three yourself with me getting you unstuck when it breaks.
10 spots, application only, because I read every one myself.
The link is below if you want to see if you're a fit.
https://theaiarchitects.com/x/mentorship/claude-code-10-hacks
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [3.1K](https://x.com/tomcrawshaw01/status/2079916336298565907/analytics)
- [https://theaiarchitects.com/x/mentorship/claude-code-10-hacks](https://theaiarchitects.com/x/mentorship/claude-code-10-hacks)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:07 PM · Jul 22, 2026](https://x.com/tomcrawshaw01/status/2079916336298565907)
- [3,104 Views](https://x.com/tomcrawshaw01/status/2079916336298565907/analytics)
---
*导出时间: 2026/7/23 11:07:19*
---
## 中文翻译
# 在 Claude Code 度过 1,000 小时后，我只用这 10 个功能
**作者**: Tom
**日期**: 2026-07-22T13:07:52.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2079916336298565907](https://x.com/tomcrawshaw01/status/2079916336298565907)
---
下面每一个功能都经受了一千小时真实构建的考验。第十个功能几乎没人设置，但它却是最重要的。
六个月前，我还在像完成工作任务一样，逐个点击批准每一个权限请求。
晚上我结束会话，早上打开一个新的，Claude 就把一切都忘光了。
而当构建出错时，我剩下的一天都在消耗 tokens 去撤销那些本不该构建的东西。
讽刺的是，我已经教过成千上万的人如何从零开始使用 Claude Code。
问题出在我自己的配置上。
于是我回顾了一千小时的真实构建记录，删除了大部分尝试过的东西。
留存下来的就是下面的清单。
有些配置六十秒就能搞定。
最后一个功能决定了 Claude 是每次会话都从零开始，还是随着你的使用变得越来越聪明。
## 告别点击确认权限的设置
自动模式（Auto mode）。
在终端按 Shift tab，或者在桌面应用底部的下拉菜单中选择，Claude 就会自动运行。
你让它构建项目，然后就可以离开去忙别的了。
这不是那个旧的“危险跳过权限”标志。
自动模式有真正的防护栏，会拦截危险命令，并允许你在此基础上添加自己的规则。
它在 Pro 和 Max 版本、桌面端和终端端都可用，开启只需要六十秒。
## 让它在构建之前先规划
自动模式解决了点击确认的问题。
但它解决不了 Claude 自信满满却快速构建出错误东西的问题。
计划模式（Plan mode）可以。
同样是在 Shift tab 菜单里。
在计划模式下，Claude 无法构建任何东西，它只能生成一份计划。
你的工作就是真正地去阅读它。
大多数人只是略读、批准，然后抱有侥幸心理，最后却要花费 tokens 去返工。
一旦计划确认无误，批准它，Claude 就会在自动模式下执行。
先规划，后免提构建。
## 离开工位时继续构建
你正处在会话中途，却得外出。
在终端输入 /remote-control，然后在手机上打开 Claude 应用。
它会直接连接到你机器上运行的会话。
同一个会话，同一个上下文。
我曾在出租车的后座继续进行构建。
## 赋予 Claude 构建流水线的插件
Superpowers。
我先试了 GSD，但它太臃肿了。
Superpowers 的运行方式符合我已有的构建习惯。
你头脑风暴，它将其转化为计划，计划变成实施方案，并行子代理进行构建，同时清单会在屏幕上追踪所有进度。
我对 Facebook 广告活动口述了一个粗略的想法，它返回了受众、切入点、广告文案和定向建议。
只需说“让我们头脑风暴”或“让我们制定计划”，它就会自动调用相应的技能，无需斜杠命令。
安装只需复制粘贴 GitHub readme 中的内容。
## 防止 Claude 破坏项目的四条规则
你的 CLAUDE.md 是 Claude 在触碰任何东西之前读取的指令文件。
Andrej Karpathy 的版本有四条值得借鉴的规则。
先思考后编码，让假设和权衡浮出水面，而不是被隐藏。
简单优先，用最少的代码解决问题。
精准修改，只触及被要求修改的文件。
目标驱动执行，定义“完成”的标准，循环验证直到达成。
已经有一个 CLAUDE.md 了吗？
把他的规则粘贴到聊天中，让 Claude 合并进去。
## 你做两次的任何事情都会变成技能
技能就是保存下来的流程，让 Claude 每次都以同样的方式运行。
我的内容技能接收一份 YouTube 转录稿，返回根据直复营销标准评分的 30 个钩子。
我来挑选获胜者，因为我是文案，这个决定权留给人。
我的图片技能也是如此，它根据我表现最好的帖子图片提取出的配方来工作。
有一个警告。
我安装了很多别人的技能，但删掉了大部分。
它们不具备你的上下文。
使用内置的技能创建器，构建你自己的技能。
## 以每分钟 108 个字的速度对话
对我构建质量的最大提升与 Claude 无关。
而在于我与它的对话方式。
Wispr Flow。
按住 FN 和空格键，连续唠叨长达六分钟，整洁的文本就会出现在光标处。
我以每分钟 108 个字的速度口述，总共已经超过了 681,000 字。
更好的结果源于更多的上下文，但没人会在提示框里手动输入六分钟的上下文。
在办公室里，插上耳机，对着麦克风低语。
这就是它为什么叫 Wispr Flow 的原因。
## 100 英镑计划享受 2,000 美元的用量
你安装的每一个技能和 MCP 都会加载到每个会话的上下文中。
臃肿感会悄然累积。
我的一些学生开启新会话时，上下文窗口就已经被占用了 15%。
我的则保持在 2% 以下。
Token 优化器仪表盘显示了资源去向以及该削减什么。
它还显示按 API 费率计算你的用量需要多少钱。
上个月我的用量折算下来超过了 2,000 美元。
有一个值得借鉴的习惯：当上下文窗口使用率达到 35% 到 40% 时，编写一个交接文件并清理会话。
## 在你睡觉时工作的代理
技能在你运行时才工作。
例行程序（Routines）则在你睡觉时运行。
Claude Code 桌面应用中的例行程序选项卡可以将任何技能或工作流排入计划。
我的会在每周一上午 8 点运行收入报告。
如果可能的话远程运行它们，因为笔记本电脑上的例行程序在合盖的那一刻就会停止运行。
一次构建、永远按计划运行的技能，就是最接近那种不睡觉员工的东西。
## 最重要的一点
以上所有内容都能让一次会话变得更好。
这一点则让之后的每一次会话都变得更好。
人们称之为“第二大脑”，听起来像是个流行词。
剥去外壳，它其实就是文件夹里的 Markdown 文件。
我的存放在 Obsidian 里，Claude Code 读取同一个文件夹。
我的语音资料、品牌资产、过往的新闻通讯、项目文档、转录稿。
当我说“参考我们上周写的那封邮件”时，Claude 知道那是什么。
没有这一层，每次会话都从零开始。
有了它，上下文就会不断累积。
人们跳过它是因为听起来像个大工程。
其实它就是一个文件夹。
从三个文件开始。
即使只设置了清单的一半，Claude Code 也不再表现得像个聊天机器人，而是开始像一个系统一样运行。
你可以自己安装全部十个，上面的所有内容都是免费或接近免费的。
或者你可以在我为期 90 天的高强度训练营里和我一起构建它们。
我们一对一构建你的第一个代理，然后你自己构建接下来的两三个，当出现问题时我会帮你解开症结。
10 个名额，仅限申请，因为我会亲自阅读每一份申请。
如果你想看看自己是否合适，链接在下面。
https://theaiarchitects.com/x/mentorship/claude-code-10-hacks
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [3.1K](https://x.com/tomcrawshaw01/status/2079916336298565907/analytics)
- [https://theaiarchitects.com/x/mentorship/claude-code-10-hacks](https://theaiarchitects.com/x/mentorship/claude-code-10-hacks)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:07 PM · Jul 22, 2026](https://x.com/tomcrawshaw01/status/2079916336298565907)
- [3,104 Views](https://x.com/tomcrawshaw01/status/2079916336298565907/analytics)
---
*导出时间: 2026/7/23 11:07:19*