Editing AI output

The quality of an AI-generated interface is determined by the edit, not the generation

The quality of an AI-generated interface is determined by the edit, not the generation. ## Delete before you add AI generates too much. Too many sections, too many words, too many effects. The first editing pass is always subtraction. For every element on the page, ask: _if I remove this, does the page get worse?_ Not sure? Remove it. You'll know within seconds whether it was load-bearing. Sections AI adds that rarely survive editing: - **"Trusted by" logo strips** with placeholder logos. - **Testimonial carousels** with generic praise. - **Feature grids** listing capabilities nobody asked about. - **Secondary CTAs** that dilute the primary action. - **Decorative dividers** between every section. ## The three-pass edit ### Pass 1: Structure One primary action per page. Eye flows top to bottom without getting stuck. Delete the bottom third. Most AI pages are 30-40% too long. Combine sections that make the same point twice. Promote the most important content to the top. ### Pass 2: Typography and spacing Swap the default font. Tighten heading `line-height`. Open up body `line-height`. Check that heading sizes create clear jumps (not `18→20→22`, but `16→24→36`). Add whitespace between sections. AI underestimates how much breathing room a page needs. ### Pass 3: Details Hover states. Focus styles. Active states. Transitions on interactive elements. `border-radius` consistency. Shadow refinement. Loading states. Empty states. AI skips these because they're invisible in a screenshot. They're not invisible to a user. ## AI-generated CSS to fix on sight | Pattern | Problem | Fix | | ---------------------------------------------- | ------------------------------------------ | ---------------------------------------------- | | `rounded-3xl` or `rounded-full` on cards | Over-rounded, looks inflated | `rounded-lg` or `rounded-xl` max | | `shadow-2xl` | Floating, disconnected from surface | Two-layer shadow: tight contact + wide ambient | | `bg-gradient-to-r from-purple-500 to-blue-500` | Default AI gradient | Solid colour from your palette, or remove | | `text-center` on everything | No hierarchy, no reading direction | Left-align body text | | `gap-4` everywhere | Uniform spacing kills rhythm | Vary spacing by content relationship | | `font-bold` on multiple elements per section | Nothing stands out when everything is bold | One bold element per visual group | ## The editing mindset Sand it, feel the grain, get a splinter, sand again, and repeat until smooth. The question is never "did AI add this?" The question is "does this earn its pixels?" ## Common mistakes - **Editing the prompt instead of the output.** Prompt engineering is useful, but slower than editing HTML. Get the structure from the AI, then fix it by hand. - **Asking AI to "make it more unique."** The model's idea of unique is adding effects. Unique is removing everything that doesn't serve the product. - **Skipping pass 3.** Structure and typography get 80% of the way. The detail pass is the difference between "good" and "someone cared."