Prompting the agent

The AI coding agent does its best work when you brief it like you'd brief a capable contractor: clear about the outcome, honest about the constraints, and iterative. This page collects practical guidance that applies whether you're using the in-app agent or an AI editor against a pulled site.

Describe outcomes, not implementations

Say what you want to be true, not how to code it. "Visitors should be able to book a 30-minute call and get a confirmation email" is a better prompt than a list of files to create. The agent knows the platform; let it choose the mechanics.

Give it the real content and context

Vague inputs produce vague sites. The more concrete you are — actual services, real prices, the tone of voice, the audience — the better the result. If you have existing material (copy, a brand, examples of sites you like), share it up front.

Work in focused steps

Big, sweeping requests are harder to get right in one shot. Build in increments: get the structure right, then a feature, then polish. Review the preview after each step so you catch course-corrections early instead of unwinding a large change.

Be specific about data and logic

When a feature involves data — a form, a directory, a members area — say what fields you need, who can see them, and what should happen on submit (email someone? show a thank-you page? require login?). Spelling this out gets you working logic instead of a placeholder.

Correct with specifics

When something isn't right, point at it precisely: "the pricing table should have three tiers, not two" beats "fix the pricing". Reference the page and the exact element. If the agent went the wrong direction, say so directly and describe the target state.

Keep accuracy in mind

The agent will write plausible copy to fill a page if you don't give it real content. For anything factual — hours, prices, contact details, claims about a business — provide the real values, and review generated text before publishing.

When to switch to the CLI

If you find yourself wanting version history, code review, or fine-grained control over files, pull the site with the CLI and use an AI editor — you'll get the same platform knowledge via AGENTS.md, plus your own tooling around it.