Dior Candy Persona Site

At 2:17 AM, Will stared at the hero section in Figma that looked like a spilled macaron palette, realizing for the first time the abyss between "beautiful" and "usable." It was the fourth Wednesday of November 2024; Beijing had already turned winter. He had just pushed the first version of the Dior Candy Persona Site to Vercel Edge when his phone started vibrating furiously on WeChat—not congratulations, but screenshots from three early users arriving almost simultaneously: "This pink makes the text unreadable," "My eyes get tired after three seconds of scrolling," "Are you making children's toys?" Will later said this was the moment he understood "visual obesity": each color individually was haute couture candy—#FF7B9C rose frosting, #5CC9A7 mint ice cream, #B48EE0 lavender macaron—but when they lined up side by side like a tri-color cup across a 375px viewport, the information hierarchy collapsed into a pot of sickly sweet porridge.

From "Full-Sugar Disaster" to "Sugar-Control Surgery"

The rollback decision came fast, but the execution was painful. Will didn't simply change the background to pure white or black—that would mean admitting complete experimental failure. Instead, he spent four full nights on the "moodboard stripping method" mentioned in Swarm Engine v2 Birth Story: projecting iOS 26 Liquid Glass references, Xiaohongshu waterfall screenshots, and Dior 2024 Spring/Summer haute couture runway videos side by side on a 27-inch monitor, marking physical sticky notes for "keep this halo" and "kill this gradient." The final Ice Cream dark base wasn't pure black, but #1A1A2E with 4% blue bias—deep enough to make any candy color float, yet "colored" enough to avoid the trap of generic dark mode.

The candy colors' living space was compressed to two precise touchpoints:

• Default aura on persona cards—low saturation, low brightness, a mere 0.8px outline width, like the faint glint of gemstones through velvet jewelry box lining
• Hover halo—a 360° conic gradient radiating from the card center, active-state brightness boosted 340%, but duration strictly controlled under 200ms to avoid visual afterimages

This "sugar-control" strategy drew partial inspiration from Will's concurrent work on the Cattery SNS Operations Timeline project, where he learned to handle pet fur highlights: the most eye-catching elements must be boxed within a "predictable interaction contract," never allowed to invade static reading zones.

The "Darkroom" Moment of Local Agent

A more hidden architectural decision than the visuals was stuffing the Hermes agent into localhost:9119 on a local machine. Will's exact words: "I didn't want six different personas sharing the same cloud 'soul'—it felt like six actors sharing one throat." This decision seemed paranoid at the time—in late 2024, nearly every AI application was preaching serverless, edge inference, global low latency—but Will's preliminary research for OpenClaw Multi-Instance Chronicle revealed that cloud multi-instance cold-start costs weren't measured in milliseconds, but in "cognitive fracture at persona switch time."

For implementation, Will repurposed a spare M2 Mac mini as the "persona darkroom," with six independent system prompt templates:

1. dior-rose — Rose persona, specializes in emotional counseling, color-anchored to #FF7B9C
2. dior-mint — Mint persona, handles data analysis, color-anchored to #5CC9A7
3. dior-lav — Lavender persona, focuses on creative writing, color-anchored to #B48EE0
4. dior-yel — Lemon persona, manages scheduling, color-anchored to #FFCB45
5. dior-ice — Ice Blue persona, dedicated to technical Q&A, color-anchored to #7BD3EA
6. dior-cocoa — Cocoa persona, late-night companion mode, color-anchored to #D4A574

During switching, the frontend sends only two commands: /switch-persona carrying the target ID, and /load-avatar fetching the corresponding Lottie animation resource. The Hermes local instance forwards simultaneously to MiniMax, Kimi, and DeepSeek via OpenAI-compatible endpoints, dynamically routing by response speed. Perceived persona-switch latency compressed from 800ms in the cloud solution to 120ms, approaching the human psychological threshold for "instant feedback." This local-first architecture later directly fed back into the "instance isolation + unified gateway" design paradigm in OpenClaw Multi-Instance Chronicle.

"The most dangerous assumption is that users like color. The truth is, users like the moment they are allowed into color." —Will, Notion memo, early hours of November 27, 2024