Zenn guide: vague prompts cause AI-ish landing pages, 9-step fix

The guide's starting point is a familiar frustration: a request for a "stylish landing page" yields the same blue gradient, rounded buttons, and emoji icons every time. It attributes this not to a failure of taste but to a failure of specification — AI fills unspecified parts with the average, and the average is what readers recognize as generic. The guide's remedy is to make the implicit explicit, replacing adjectives with numbers and concrete prohibitions.

The nine steps move from preparation to generation to finishing. The first three — deciding a tone system, choosing three reference sites by role, and taking screenshots — are framed as the stage whose omission causes "revision hell" later. The middle steps have ChatGPT or Claude translate screenshots into a design spec with estimated HEX values and px sizes, then recombine those specs by role so the result does not copy any single site. The system prompt then carries Japanese-specific instructions and a list of bans — no gradient backgrounds, no emoji icons, no glass-style cards, no strong drop shadows.

The guide is explicit about the limits of the approach. It says AI alone is unlikely to reach publishable quality, and that final adjustments to line breaks, images, and fabricated numbers should be done by hand. It also warns that having AI write the copy invites stock phrases, and that AI may fill empty slots with invented testimonials or figures — a risk it links to Japan's Act against Unjustifiable Premiums and Misleading Representations if published as-is. The tone throughout is procedural rather than promotional; the guide offers a workflow, not a promise of conversion gains.