🤖 HarDojo
Log in Sign up

Writing Great Prompts

The five ingredients of a prompt that gets good code.

🧪 Try it in the Playground 🎯 Practice this path

A good coding prompt gives the model everything it needs to make good decisions. Use this structure:

The five ingredients

  1. Role — who the model should act as.
  2. Context — the files, stack, and constraints.
  3. Task — one clear, specific goal.
  4. Constraints — what NOT to do.
  5. Definition of done — how we'll know it's correct.

Example: weak vs strong

Weak:

"fix the login bug"

Strong:

You are a senior TypeScript engineer. In src/auth/login.ts, the login function fails when the email contains uppercase letters because the lookup is case-sensitive. Task: normalize the email to lowercase before the lookup. Constraints: do not change the password hashing, do not touch the UI. Done when: npm test passes and a new test covers "User@Example.com".

The strong prompt names the file, the cause, the fix, the constraints, and the test.

Rules of thumb

  • Be specific about the "where". "Fix the bug in X" beats "fix the bug".
  • Give the acceptance criteria. "Done when tests pass" is a real requirement.
  • Show, don't describe. Paste a failing test or error message.
  • One task per prompt. Two tasks = two chances to drift.
  • State constraints. What must stay untouched is as important as what must change.

When to be less specific

If you're exploring ("how could I structure this?"), loosen the constraints and ask for options:

"Suggest three ways to structure the caching layer, with trade-offs. Don't write code yet."
The prompt is a spec. Vague specs produce vague code — precise specs produce code you can actually merge.