Drag to rotate · More specific prompts assemble more complete, better-structured pages
What happens after you hit send
- Planning: the agent breaks your prompt into a build plan — layout, sections, styles, scripts.
- Generation: it writes HTML/CSS/JS, usually in one file for simple pages.
- Self-checking: good agents render or lint their own output and fix errors before showing you.
- Iteration: your follow-up messages ("make the header sticky") patch the existing code rather than starting over.
The "walk away" part is real: agents run autonomously for minutes. But autonomy amplifies the prompt — a vague brief autonomously builds the wrong thing faster.
The 80/20 of quality
In practice, results improve most when you specify: audience, sections, one example of tone, and what "done" means. Everything else the agent guesses well.
A prompt template that ships
Build a single-file HTML landing page for [WHO/WHAT].
Audience: [who reads it]
Sections: hero with headline "[...]",
3 feature cards, pricing table
(3 tiers: $X/$Y/$Z), FAQ (5 questions), footer.
Style: light background, [color] accent,
rounded cards, mobile-first at 375px.
Content: write real copy, no lorem ipsum.
Done means: valid HTML, no console errors,
loads with zero external dependencies.
Why each line earns its place
- "Single-file" avoids broken import paths.
- Named sections stop the agent from inventing scope.
- Real numbers ($X tiers, 5 FAQs) prevent placeholder sprawl.
- "Done means" gives the agent a checkable finish line — the single biggest upgrade over vague prompts.
Honest limits
- Great for: landing pages, prototypes, calculators, internal tools, one-page sites — anything self-contained.
- Shakier for: multi-page apps with auth, payments, or databases — those need review, tests, and a real deployment story.
- Always check: claimed features actually work (click every button), the page is responsive, and no secrets or fake testimonials were invented.
Treat the agent like a fast junior developer: brilliant output speed, needs a clear brief and a code review.