Skip to main content
Vibe Code Forum

Article: Prompting that works: small steps, clear checks

Good prompts for building aren't clever. They are specific, small and checkable.

The shape of a good request

  1. Context: where we are. "We're building the booking app in PLAN.md. Sign-in works."
  2. One task: what to do now. "Add a page listing the signed-in user's bookings, newest first."
  3. Details that matter: "Show date, time and status. Empty state: 'No bookings yet' with a button to make one."
  4. Limits: what not to touch. "Don't change the sign-in code or the database tables."
  5. Done means: how you'll check. "Done when I can sign in as a test user and see only their bookings."

You don't need all five every time, but the more the change matters, the more of them you include.

Habits that help

  • One feature per request. Big requests produce big, tangled changes that are hard to check and undo.
  • Ask for a plan before big changes: "Before writing code, tell me how you'll do this and which files you'll change." Fix the plan, then say go.
  • Show, don't just tell. Paste the exact error, a screenshot, or an example of the output you want.
  • Say what you see, not what you think the cause is. "When I press Save, nothing happens and the console shows this error" beats "the database is broken".
  • Ask it to explain. "Explain what you changed and why, in plain English." You learn, and you notice when something is off.
  • Reuse what works. Keep your best prompts and project rules in a rules file or notes file.

Prompts worth saving

  • "Read the project and explain how it fits together before changing anything."
  • "What could go wrong with this approach? What would a senior developer criticise?"
  • "Is there anything insecure here, especially keys, sign-in and who can see whose data?"
  • "Make the smallest change that fixes this. Don't refactor anything else."
  • "Write a short test that proves this works."

When the answers get worse

If the AI starts going in circles, contradicting itself or forgetting rules, the conversation is probably too long. Start a new one with a short summary. See Tokens, context windows and why the AI forgets.

When it's broken and you're stuck, switch to the debugging routine.

More in Build guides