怎么写代码提示词：让AI写出能跑、能维护的代码（2026）

我前阵子用 AI 写一个小脚本，第一版看着挺美，跑起来直接报错，改了三回还是没跑通。后来才想明白：问题不在模型够不够聪明，是我给的指令太"礼貌"——只说"帮我写个爬虫"，没说语言、没说目标网站结构、没说异常处理。2026 年的编程助手（Cursor、Copilot、通义灵码、Claude Code 这些）已经很强，但它们不会替你想清楚需求。这篇文章把写代码提示词的几招一次讲透：一个三段式框架 + 约束前置 + 给输出形状 + 贴现有代码 + 报错怎么贴，最后告诉你怎么用工具一键出结构化代码指令。

⚠️ 文中方法来自 2026 年公开实测与多家权威来源（DAIR.AI Prompt Engineering Guide、Anthropic 官方 prompt 工程文档、各 AI 编程助手官方最佳实践），多家结论一致。模型迭代快，具体话术以你用的工具当日表现为准。

一、先想清楚：把"上下文"喂给 AI（Context→Constraint→Task）

写代码提示词最容易被忽略的就是上下文。人知道项目用什么框架、跑在什么环境、旁边有哪些文件；AI 不知道。2026 年多个权威来源都把"上下文"放在第一位。一个能用的代码提示词基本是这三段：

【上下文 Context】语言/版本、框架、目标文件、上下游、运行环境（如 Python 3.11 + FastAPI）
【约束 Constraints】不准怎么、必须兼容什么、风格要求（如"不引入新依赖""保持函数签名"）
【任务 Task】用明确动词（实现 / 修复 / 重构 / 加测试），并说清期望的输出形态

一句话自检：AI 拿到这段指令，能不能不追问就动手？不能，就补上下文。

二、约束前置：把"不准怎么"写在前

很多人写"帮我写个函数"，AI 自由发挥，结果用了你项目里压根没装的库。2026 年实测下来最有效的一招是：约束前置——把"不要做什么""必须兼容什么"放在最前面，比事后纠正省事得多。比如："不要使用 pandarallel；必须兼容 pandas 2.0；保持原函数的入参顺序。"约束写清楚，AI 就不会自作主张。

三、给输出形状：Show, don't just tell

你说"返回一个好用的结果"，AI 不知道你心里的"好用"长什么样。与其描述，不如直接给形状：要 JSON 就贴个字段结构，要函数就说明入参出参。多家来源都强调"给示例输出（show the output shape）"比形容词管用。事实类/格式特殊的任务，给一个"照这个写"的小样例，模型从例子学到的远多于从形容词里学到的。

四、贴现有代码：@file 引用 vs 直接粘贴

Cursor、Claude Code 这类支持把相关文件作为上下文（@file 或直接把文件拖进来）。实测经验：改 bug 一定把出问题的文件/函数贴进去，只贴报错信息 AI 多半猜错。但别把整个仓库糊进去——贴"相关的最小片段"比贴全量更准，也省 token。引用现有代码时顺带说清"这个函数的约定是什么"，AI 才接得上你的风格。

五、把需求拆小 + 让 AI 先出方案

一次要 AI 写"一个完整的用户系统"，十个里有八个会跑偏。2026 年的成熟用法是拆小：先把需求切成"建表 → 注册接口 → 登录校验"这种小步，一步步来。更稳的是让 AI 先出方案再写代码——加一句"先列出实现思路，确认后我再让你写"。它列错了你立刻纠，比写完再返工便宜太多。这招对复杂任务几乎是必选项。

六、报错怎么贴才有效

贴报错最常见的错误是只丢一行 traceback。有效的贴法是四件套：完整报错 + 触发它的命令 + 相关代码片段 + 你已经试过的办法。再加一句"只返回最小修复，不要重写整个文件"，AI 给你的补丁才小、才好 review。我踩过坑：贴一行报错让 AI 猜，它给我重写了一整页，反而更乱。

七、三个真实改写案例（弱 → 强）

案例 A · 写个数据清洗函数

❌ 弱：「帮我写个处理 csv 的脚本」

✅ 强：「Python 3.11 + pandas 2.0 环境。写一个函数 clean(df)，清理包含字段 name/age/city 的 DataFrame：去掉 age 非数字的行、city 去前后空格并统一小写、缺失 name 填 'unknown'。不引入新依赖，保留逐行注释，返回清洗后的 df。输出用函数定义形式，不要示例调用。」

案例 B · 修一个报错

❌ 弱：「这个报错怎么修」

✅ 强：「环境 Python 3.10。运行 pytest tests/test_api.py 报 KeyError: 'user_id'。相关函数是 api/handler.py 的 get_user()。请只返回最小 Git diff 补丁并指出原因，不要重写整个文件。约束：保持原函数签名、兼容 Python 3.10。」

案例 C · 加一个功能

❌ 弱：「给网站加个搜索」

✅ 强：「技术栈：Next.js 14 + PostgreSQL。先列出给文章表加全文搜索的实现思路（含需要的索引和接口签名），确认后我再让你写代码。约束：不引入额外搜索服务、返回 JSON 且字段含 id/title/snippet。」

八、不想手搓？用工具一键出结构化代码指令

每次都现想这套结构也累。用本站 提示词生成器 选「代码」场景，填语言、任务类型、约束，一键生成结构化的中英双语代码提示词——正好对应大家搜的「代码提示词生成器」。你写大白话，工具帮你转成"上下文 + 约束 + 任务 + 输出形状"齐活的指令，丢给 Cursor、通义灵码这类编程助手就能直接用。

How to Write Code Prompts: Get AI to Write Code That Runs and Maintains (2026)

A while back I asked an AI to write a small script. The first version looked clean, then crashed on run, and three rewrites later it still didn't work. I finally realized the problem wasn't the model — my instruction was too "polite": I said "write me a scraper" without the language, the target site's structure, or error handling. In 2026 the coding assistants (Cursor, Copilot, Tongyi Lingma, Claude Code, and friends) are strong, but they won't think your requirements through for you. This guide lays out the whole method: a three-part framework + constraint-first + show the output shape + paste the existing code + how to paste errors, and ends by showing how to generate a structured code prompt with a tool.

⚠️ Methods come from 2026 public testing and authoritative sources (DAIR.AI Prompt Engineering Guide, Anthropic's official prompt-engineering docs, official best practices from AI coding assistants) — consistent across vendors. Models move fast; exact phrasing depends on the tool you use that day.

1. First, Feed the AI Context (Context→Constraint→Task)

The most overlooked part of a code prompt is context. You know the framework, the runtime, the neighboring files; the AI doesn't. In 2026 multiple authoritative sources put context first. A usable code prompt is basically these three parts:

[Context] language/version, framework, target file, upstream/downstream, runtime (e.g. Python 3.11 + FastAPI)
[Constraints] what NOT to do, what it must stay compatible with, style (e.g. "no new dependencies", "keep the function signature")
[Task] a clear verb (implement / fix / refactor / add tests), plus the expected output shape

Self-check: could the AI start without asking a follow-up? If not, add context.

2. Constraint-First: Write the "Don'ts" Up Front

People say "write me a function" and the AI freelances — then uses a library your project doesn't even have. The most effective trick from 2026 testing: constraint-first — put "don't do X" and "must stay compatible with Y" at the very top. Cheaper than fixing after. Example: "Do not use pandarallel; must stay compatible with pandas 2.0; keep the original parameter order." State constraints clearly and the AI won't improvise.

3. Show the Output Shape: Show, Don't Just Tell

You say "return a useful result" and the AI has no idea what "useful" looks like to you. Instead of describing, give the shape: for JSON, paste the field structure; for a function, state inputs and outputs. Multiple sources stress "show the output shape" beats adjectives. For factual or format-specific tasks, give one "write like this" sample — the model learns more from examples than from adjectives.

4. Paste the Existing Code: @file vs. Inline Paste

Cursor and Claude Code let you pull related files as context (@file, or drag the file in). Real-world lesson: when fixing a bug, always paste the offending file/function. A bare error message usually leads the AI to guess wrong. But don't paste the whole repo — the minimal relevant snippet is more accurate and saves tokens. When referencing existing code, also state the function's conventions so the AI matches your style.

5. Cut the Request Small + Make the AI Plan First

Asking for "a complete user system" goes off the rails eight times out of ten. The mature 2026 usage is small steps: split the work into "schema → register endpoint → login check" and go one at a time. Even steadier: have the AI plan before coding — add "first list the approach; I'll confirm before you write." If its plan is wrong you correct early, far cheaper than redoing after. For complex tasks this is almost mandatory.

6. How to Paste an Error Effectively

The most common mistake is pasting one line of traceback. The effective paste is a four-piece set: full error + the command that triggered it + the relevant code snippet + what you've already tried. Add "return only the minimal fix, don't rewrite the whole file" so the patch is small and easy to review. I learned the hard way: a one-line error made the AI rewrite a whole page, which only made it messier.

7. Three Real Rewrite Examples (weak → strong)

Example A · A data-cleaning function

❌ Weak: "write me a script to process a csv"

✅ Strong: "Python 3.11 + pandas 2.0. Write a function clean(df) that cleans a DataFrame with columns name/age/city: drop rows where age is non-numeric, strip and lowercase city, fill missing name with 'unknown'. No new dependencies, keep line comments, return the cleaned df. Output as a function definition only, no example call."

Example B · Fix an error

❌ Weak: "how do I fix this error"

✅ Strong: "Python 3.10. Running pytest tests/test_api.py throws KeyError: 'user_id'. The relevant function is get_user() in api/handler.py. Return only a minimal Git-diff patch and the reason; don't rewrite the whole file. Constraints: keep the original signature, compatible with Python 3.10."

Example C · Add a feature

❌ Weak: "add search to the site"

✅ Strong: "Stack: Next.js 14 + PostgreSQL. First list the approach to add full-text search on the articles table (including the needed index and the interface signature); I'll confirm before you write code. Constraints: no extra search service, return JSON with fields id/title/snippet."

8. Don't Hand-Roll? Generate a Structured Code Prompt With a Tool

Inventing this structure every time is tiring. Use our Prompt Generator — pick the "Code" scenario, fill in the language, task type, and constraints, and get a structured bilingual code prompt in one click — exactly the "code prompt generator" people search for. You write plain language; the tool turns it into a ready-to-use instruction with Context + Constraints + Task + output shape, which you can drop straight into Cursor or Tongyi Lingma.