役割と制約
AIの役割と境界を設定し、出力を改善します。
「役割と制約」はCoddyKit上の無料Vibe Codingレッスンです。 これはレッスン1/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはVibe Coding学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 Vibe Codingコースには全4レッスンが含まれています。
このレッスンの一部はまだ翻訳されておらず、英語で表示されています。
Tell the AI Who to Be
You already know how to ask AI for code. Now let's get production-quality output on the first try. The fastest upgrade? Give the AI a role and a set of constraints.
A role tells the AI what kind of expert to channel. Constraints tell it the rules it must not break. Together they turn a vague request into a precise brief — the same way a good client gets better work from a freelancer.
What Is a Role?
A role is a one-line identity you assign at the start of a prompt: "You are a senior React engineer" or "Act as a security-focused backend developer."
This isn't magic — it steers the model toward the patterns, vocabulary and standards that kind of person would use. A "senior engineer" writes cleaner, safer code than an unspecified assistant. Tools like Cursor and Claude Code respond strongly to roles because they shape the whole answer.
You are a senior front-end engineer who writes clean, accessible, well-commented JavaScript. Build a function that validates an email address and returns true or false.Why Roles Improve Output
Without a role, AI gives you the average of everything it has seen — including beginner code from tutorials. With a role, you pull from the expert slice.
- Better defaults — a "senior" picks safer patterns.
- Right vocabulary — it explains things at your level.
- Consistent style — it stays in character across edits.
Think of it as choosing which teammate shows up to help you.
What Is a Constraint?
A constraint is a hard rule the answer must obey. Roles set the vibe; constraints set the boundaries.
Examples of useful constraints:
- "Use only vanilla JavaScript — no libraries."
- "Keep the function under 20 lines."
- "Do not use any external API."
- "Return only the code, no explanation."
Each constraint removes a whole category of unwanted answers.
You are a senior JavaScript developer. Constraints: vanilla JS only, no libraries, under 20 lines, include a one-line comment above the function. Write a function that formats a number as US currency.Roles + Constraints Together
The real power comes from combining them. The role decides how the code is written; the constraints decide what's allowed. Stack them at the top of your prompt so the AI reads the rules before the task.
A reliable template:
- Role: who the AI is
- Constraints: the must/never list
- Task: what to build
ROLE: You are a senior front-end engineer focused on accessibility.
CONSTRAINTS:
- Vanilla JavaScript and HTML only
- No external libraries
- All interactive elements must be keyboard-usable
TASK: Build a button that copies the page URL to the clipboard and shows a 'Copied!' message for 2 seconds.A Constraint in Action
Here's the kind of clean, self-contained JavaScript a well-constrained prompt produces. Notice it's small, readable, and does exactly one job — because the constraints asked for exactly that.
Run it to see the output.
// Format a number as US currency
function toUSD(amount) {
return new Intl.NumberFormat('en-US', {
style: 'currency',
currency: 'USD'
}).format(amount);
}
console.log(toUSD(1999.5));
console.log(toUSD(42));Negative Constraints: The 'Never' List
Some of the most powerful constraints are things you forbid. AI loves to over-deliver — adding libraries, extra features, or rewriting your whole file. A never list reins it in.
- "Never add new dependencies."
- "Never change files I didn't mention."
- "Never use TypeScript — plain JavaScript only."
In agentic tools like Claude Code, a clear never list prevents surprise edits across your project.
You are a careful engineer working in an existing project.
NEVER:
- Add new npm packages
- Edit any file other than cart.js
- Remove existing comments
TASK: In cart.js, add a function getTotal(items) that sums item.price for all items.Scope Constraints
One constraint beginners forget: limiting scope. Without it, ask for one feature and AI may refactor everything. Pin it down:
- "Only modify the function I paste below."
- "Give me just the changed lines."
- "Don't touch the styling."
Tight scope makes AI output easier to review and far less likely to break working code.
Here is my working function. Change ONLY the sort order to descending. Do not rename anything, do not add features, return only the updated function.
function topScores(scores) {
return scores.sort((a, b) => a - b).slice(0, 3);
}Output-Format Constraints
You can also constrain the shape of the reply. This keeps answers easy to paste and use:
- "Return only a single code block, no prose."
- "Reply with the full file, ready to save."
- "List 3 options as a numbered list, code only."
App builders like v0 and Bolt already format output as runnable apps — but inside an editor, saying how you want the answer saves you cleanup time.
You are a JS engineer. Return ONLY the complete function in one code block, with no explanation before or after.
Task: Write debounce(fn, delay) that returns a debounced version of fn.Build a Reusable Role Header
Pros don't retype roles and constraints every time. They save a role header and reuse it. In Cursor you can store it in a .cursorrules file; in Claude Code, a CLAUDE.md; or just keep a note you paste in.
A good standing header sets your stack, your style, and your never list once — so every prompt inherits it automatically.
PROJECT RULES (paste once, applies to all prompts):
Role: Senior JavaScript engineer who writes clean, commented, accessible code.
Stack: Vanilla JS + HTML/CSS, no frameworks.
Never: add dependencies, use TypeScript, or edit unrelated files.
Style: short functions, descriptive names, no clever one-liners.Don't Over-Constrain
Constraints are powerful, but too many can box the AI into bad code. If you forbid everything, it can't find a good solution.
Balance:
- Constrain things you care about (stack, scope, safety).
- Leave room for the AI to choose implementation details.
If output feels stiff or wrong, loosen one constraint and re-run. Aim for a clear brief, not a straitjacket.
Quick Check
You want AI to fix one function without rewriting your whole file. Which addition to your prompt best achieves this?
Recap: Direct the AI
You learned to steer AI like a director:
- Roles pull expert-level output instead of average answers.
- Constraints set hard rules: stack, scope, never-lists and output format.
- Combine them with a Role → Constraints → Task structure.
- Save a reusable role header (.cursorrules / CLAUDE.md).
- Don't over-constrain — leave room for good solutions.
Next up: showing the AI examples so it copies the exact style you want.
よくある質問
「役割と制約」レッスンは無料ですか？
はい。「役割と制約」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、Vibe Codingコースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 Vibe Codingコースには全4レッスンが含まれています。
「役割と制約」で何を学びますか？
AIの役割と境界を設定し、出力を改善します。 ブラウザで直接実行するハンズオンコードでVibe Codingを演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。
Vibe Codingを始めるのに経験は必要ですか？
事前経験は必要ありません。CoddyKitのVibe Codingは初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン1/4です。
「役割と制約」レッスンにはどのくらい時間がかかりますか？
ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。
このVibe Codingレッスンでコードを書いて実行できますか？
はい。すべてのVibe Codingレッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。