English

I never set aside time to study English anymore. Instead, I let an AI quietly correct my English while I’m doing everything else.

The trick is one paragraph of system prompt — wired into every chat tool and CLI agent I use. Once it’s in place, every English message I send comes back with a tiny lesson attached, and then the actual answer I asked for.

The mechanic

Whenever I send a message in English, the model is instructed to do three things, in order:

1. Give brief feedback on my English — one or two sentences, no over-correcting.
2. Rewrite my message in natural, professional English.
3. Then answer my original request as usual.

When I write in Chinese (or any other language), it skips both feedback and rewriting and just answers. That’s the whole idea.

Because the corrections happen on the prompts I’m already typing for work, I’m practising the exact register I need — Slack messages, emails, technical questions — instead of textbook English I’ll never use.

The prompt

Drop this verbatim into the “system prompt” / “custom instructions” / “rules” slot of any tool:

English coaching mode.

If my message is in English:
1. First, give brief feedback on my English (1–2 sentences, max).
   Skip if my English is already clear and natural.
2. Then rewrite my message in natural, professional English.
3. Then answer my original request as usual.

If my message is not in English:
- Skip the language feedback and rewriting.
- Answer directly.

Keep feedback concise and practical. Do not over-correct.

That’s it. Forty lines becomes one prompt becomes a steady, low-friction tutor.

How to wire it into every tool I use

Each tool has its own slot for persistent instructions. Here’s where to put the prompt in each — current as of 2026.

Chat interfaces (web / app)

ChatGPT  ·  Profile → Customize ChatGPT → paste into the second box (“Anything else ChatGPT should know?”). Applies globally to every new chat.

Claude.ai  ·  Settings → Profile → What personal preferences should Claude consider? → paste. Or create a Claude Project and put it in the project instructions if you want it scoped per workspace.

Gemini  ·  Settings → Saved info → add as a free-text fact. For tighter control, create a Gem with the prompt as its system instructions, then chat with that Gem.

Mistral Le Chat / DeepSeek / Qwen Chat  ·  Look for “Custom Instructions,” “System Prompt,” or “Profile” in settings. All major providers now expose this in some form; the prompt above is portable.

Coding agents (CLI / IDE)

Claude Code  ·  Put the prompt in ~/.claude/CLAUDE.md for user-wide effect, or ./CLAUDE.md in a project for repo-scoped behaviour. Both files are auto-loaded as system context every session.

Cursor  ·  Settings → Rules for AI for user-wide, or drop a .cursorrules file in the project root for repo-scoped rules. New “Cursor Rules” (.cursor/rules/*.mdc) also work.

OpenAI Codex CLI  ·  Add AGENTS.md to the project root (or ~/.codex/AGENTS.md for global). Codex CLI reads it as system context, similar to CLAUDE.md.

GitHub Copilot Chat  ·  In VS Code, set github.copilot.chat.codeGeneration.instructions, or drop .github/copilot-instructions.md into the repo. Both feed into Copilot’s system prompt.

Aider  ·  Put the text in .aider.conf.yml under read: (as a file path), or pass it via --message-file on launch. Aider concatenates these as system context.

Continue.dev  ·  Edit ~/.continue/config.json (or .yaml) and set the systemMessage field on the model.

Calling the API directly

If you’re hitting Claude, GPT, or Gemini via SDK, the same paragraph is your system parameter. Cache it (Anthropic’s cache_control: { type: "ephemeral" }, or just keep your tooling reusing the same context). One coach, every script.

What I’ve actually noticed

After about three months of this, the changes are quiet but real:

Aspect	Before	After
Naturalness	Grammatically correct but obviously Chinglish	Sentences flow; collocations feel native
Speed	I’d draft, second-guess, redraft	First-pass English is usable; second pass is polish
Confidence	I’d switch to Chinese the moment things got nuanced	I stay in English even for nuance, because the cost of being wrong is zero

The most surprising effect: the corrections I see most often start to self-correct in my head before I even hit send. The tool slowly makes itself less necessary.

A few notes that matter

• Scope the trigger explicitly — “only when I write in English.” Otherwise the tool starts critiquing your Chinese, which is annoying and rarely useful.
• Don’t over-correct — tell it to skip feedback when your English is already clear. A coach who corrects everything teaches you nothing.
• Keep the answer in the system prompt — some setups will give you the lesson and forget your actual question. Order matters: feedback → rewrite → answer.
• Resist the urge to read every correction — skim. The mistakes you repeat are the ones worth remembering.

If you want to improve the English you actually use at work, try this for a week. It doesn’t add a single new task to your day — it just makes the work you were already going to do quietly teach you English on the side.

---

中文

我已经很久没有专门花时间学英语了。我让 AI 在我做其他事情的同时，悄悄帮我改英文。

整个机制就是一段系统提示——粘贴到我用的每一个聊天工具和 CLI Agent 里。一次配置好之后，每条英文消息都会先附带一点点小课，再回答我真正想问的问题。

核心机制

每次我用英文提问，让模型按顺序做三件事：

1. 简短反馈 —— 1～2 句话，不要过度纠错。
2. 改写我的句子 —— 用自然、专业的英文重写。
3. 回答我原本的问题 —— 像平常一样。

我用中文（或其他语言）时，跳过反馈和改写，直接回答。就这么简单。

因为纠错发生在我本来就要写的工作消息上 —— Slack、邮件、技术问题 —— 我练的是真正用得到的英文，而不是教科书里那些一辈子用不上的句子。

提示词

直接复制粘贴到任何工具的”系统提示 / 自定义指令 / 规则”位置：

English coaching mode.

If my message is in English:
1. First, give brief feedback on my English (1–2 sentences, max).
   Skip if my English is already clear and natural.
2. Then rewrite my message in natural, professional English.
3. Then answer my original request as usual.

If my message is not in English:
- Skip the language feedback and rewriting.
- Answer directly.

Keep feedback concise and practical. Do not over-correct.

就这些。一段话，让英文学习变成”顺手做的事”。

在每个工具里接入

每个工具都有自己存放持久指令的地方。下面是截至 2026 年的接入方式。

网页 / App 对话

ChatGPT  ·  Profile → Customize ChatGPT → 粘贴到第二个框（“Anything else ChatGPT should know?”）。对所有新对话全局生效。

Claude.ai  ·  Settings → Profile → What personal preferences should Claude consider? → 粘贴。如果想按工作区限定，可以创建一个 Claude Project 并放进项目指令里。

Gemini  ·  Settings → Saved info → 作为自由文本添加。要更稳定可以建一个 Gem，把提示词作为它的系统指令，然后只和这个 Gem 对话。

Mistral Le Chat / DeepSeek / Qwen Chat  ·  设置里找 “Custom Instructions”、“System Prompt” 或 “Profile”。主流国内外 chatbot 都已经支持自定义系统提示，上面那段提示词是通用的。

代码 Agent（CLI / IDE）

Claude Code  ·  把提示词放到 ~/.claude/CLAUDE.md（全局）或项目根目录的 ./CLAUDE.md（按项目）。两个文件在每次会话都会作为系统上下文自动加载。

Cursor  ·  Settings → Rules for AI 设置全局，或在项目根目录创建 .cursorrules 文件做项目级控制。新的 Cursor Rules 格式（.cursor/rules/*.mdc）也支持。

OpenAI Codex CLI  ·  在项目根目录添加 AGENTS.md（或 ~/.codex/AGENTS.md 设为全局）。Codex CLI 会把它作为系统上下文读取，机制与 CLAUDE.md 类似。

GitHub Copilot Chat  ·  在 VS Code 里设置 github.copilot.chat.codeGeneration.instructions，或在仓库里放一份 .github/copilot-instructions.md。两种都会进 Copilot 的 system prompt。

Aider  ·  在 .aider.conf.yml 的 read: 字段里指向文件，或启动时用 --message-file 传入。Aider 会拼接进系统上下文。

Continue.dev  ·  编辑 ~/.continue/config.json（或 .yaml），在模型上设置 systemMessage 字段。

直接调用 API

如果你是用 SDK 直接调 Claude / GPT / Gemini，同样这段话就是你的 system 参数。记得开启缓存（Anthropic 的 cache_control: { type: "ephemeral" }，或让你的封装代码复用同一段上下文）。一个 coach，覆盖所有脚本。

真实的变化

大约用了三个月后，变化是缓慢但真实的：

方面	之前	之后
自然度	语法正确但明显是 Chinglish	句子顺畅，词组搭配像母语
速度	草稿、自我怀疑、重写	第一稿就能用，第二稿是润色
沟通勇气	一遇到表达不清就立刻切回中文	涉及微妙含义也敢继续用英文，因为犯错成本是 0

最让我意外的是：反复出现的错误开始在我脑子里自己纠正。这个工具慢慢把自己变得不再必要。

几个关键细节

• 明确触发条件 —— “仅在我用英文时纠正”。否则它会开始评价你的中文，烦且没用。
• 不要过度纠错 —— 告诉它”已经清晰自然的句子就跳过反馈”。每句都纠的 coach 等于没 coach。
• 保留”回答原问题”的部分 —— 有些配置会只给你课、忘了你本来要问什么。顺序很重要：反馈 → 改写 → 回答。
• 不用每条反馈都细读 —— 扫一眼即可。会重复犯的错误，自然会留下印象。

如果你想提升自己实际工作里会用到的英文，试一周这个方法。它不会给你的一天加任何新任务 —— 它只是让你本来就要做的工作，在背后悄悄教你英文。