Weak prompt: “Improve the homepage.” Strong prompt: “Add a trust strip below the hero using existing Reveal and eyebrow styles; copy from marketing.json; no new card chrome.”
I almost always include non-goals. No purple theme. No new dependencies. Do not touch header links. Non-goals stop the model from being helpful in the wrong direction.
One good example in the repo beats three paragraphs of taste. Attach the file. Say “extend this.” Agents imitate structure better than they interpret adjectives.
Acceptance checks keep everyone honest. “npm run build passes.” “Mobile hero still one composition.” “Contact form still posts to /api/contact.” If you cannot name the check, you will not know when to stop iterating.
When the first answer is wrong, I fix the rule or the prompt — not only the code. Otherwise the next run repeats the same mistake with fresh confidence.
I write prompts like I write briefs for clients: outcome, constraints, examples, done criteria. Same discipline. Less romance about AI. More shipped pages.
A pattern I reuse weekly: paste the failing screenshot description or the exact file path that looks wrong, then say what must stay identical. Specificity is kindness to future you.
If a prompt needs more than about eight sentences, I split the job. Long prompts do not make smarter agents. They make fuzzier missions.
Next step
Ready for a site that books patients or wins jobs?
Dental practices and home-service contractors — tell me your services, area, and timeline. I’ll reply with a practical plan.