Writing Prompts That Work

Anatomy of a good prompt

Most working prompts break down into the same four parts. You don't need all four every time, but the more of them you're missing, the more the AI is guessing.

Try it yourself

Draft a prompt for something you actually want to build or fix, then check it against the list below.

• Named the specific file, component, or page
• Described the current (wrong) behavior, if fixing something
• Described the expected outcome, including edge cases
• Said what should NOT change

Putting the four parts together

The anatomy table above is useful as a checklist, but a real prompt reads as one paragraph, not four labelled fields. Here is the search example from earlier, assembled the way you would actually type it.

I have a Next.js blog with posts stored in a posts array insrc/lib/posts.ts. Add a search bar above the post list thatfilters posts by title as the user types. Don't add any newdependencies, use plain React state. Show me the full updatedcomponent.

Notice what is missing: no heading that says “Context”, no bullet points. The four parts are still there, in the order that reads naturally. They have just stopped being labelled.

The cost of skipping a part

Leaving a part out does not usually produce an error. It produces a plausible-looking answer that is wrong in a way you only notice later, when you are three prompts deep trying to explain what you actually meant the first time.

A vague first prompt against a real feature typically costs two or three follow-up messages to converge on what you meant, each one narrowing the guess a little further. A prompt with context, goal, constraints, and format named up front usually gets there in one. The four parts are not extra work — they are work you would do anyway, spread across a longer conversation instead of paid upfront.