Most broken AI code comes from prompts like “make me a todo app”. The model fills every gap with a guess, and you spend the session correcting guesses instead of building.

The one-paragraph contract

Treat your first prompt as a contract. Before any code, it should state:

• What you are building, in one sentence.
• Stack — framework, language, versions if they matter.
• Constraints — no external APIs, mobile-first, keep bundle small.
• Definition of done — two or three checkable criteria.
• A pause clause — ask the model to list assumptions before coding.

That last point is the trick. Forcing the model to surface its assumptions turns silent guesses into a checklist you can correct in seconds.

Why it works

Every missing detail becomes a branch the model picks for you. A spec collapses those branches before they exist. You are not writing documentation — you are removing the space where bad decisions grow.

Common failure

Writing the spec after the first generation, as corrections. That is the expensive order. Corrections fight momentum; specs prevent it.

Build a [what]. Stack: [framework]. Constraints: [limits].
Definition of done:
- [criterion 1]
- [criterion 2]
Before writing code, list your assumptions and ask me about anything unclear.