Prompting Tips

The agent performs best with specific, concrete instructions. Vague prompts lead to generic results.

Be specific about what you want

A prompt like "make the homepage better" gives the agent nothing to work with. Instead, describe the actual change:

Add a hero section with a headline, subtitle, and CTA button.
Below that, add a 3-column feature grid with icons.

Reference existing code

The agent knows your codebase. Point it at specific files, components, or functions:

Refactor the useAuth hook to also return the user's role

The TaskList component re-renders on every keystroke — memoize the filtered list

Break large features into steps

Don't ask for an entire feature in one message. Split it up so you can review each piece:

Create a notifications table with type, message, read status, and user_id

Add an edge function that creates a notification when a task is assigned

Build a notifications dropdown in the header with an unread count badge

Each step builds on the last. You can catch mistakes early instead of untangling a giant diff.

Use plan mode for big changes

For anything that touches more than a few files, the agent shows a plan before writing code. Review it, adjust if needed, then approve. See Plan Mode for details.