WorkBuddy Test: A Bare Prompt Spits Out a Flat Card Page; taste-skill Rebuilds It Into a Dark Bento Grid
The cost of a better-looking AI page is now quantified: roughly 3.6x the compute and over 4x the wall-clock time for the same content. Anyone generating personal or product pages from AI tools can see upfront what design polish actually costs and where the time goes.
Prompt one produced card-v1.html in 1:05: 614 lines of HTML/CSS/JS at ◇1.26. A second prompt had WorkBuddy audit its own installed skills; 12 seconds and ◇0.44 later it listed six, including taste-skill, impeccable, and frontend-design. Prompt three called taste-skill by name and asked for Bento Grid, hover micro-interactions, and a click-to-pop WeChat QR modal.
The rebuild delivered card-v2.html in 8:08 at ◇4.59, 658 lines total: dark theme, Bento Grid, a physical-book display card, and the requested popup. Code grew by only 44 lines, so the extra time went to browser-driven multi-width render checks and visual review, not more writing. Deployment advice capped the flow: direct publish inside WorkBuddy for a link in minutes, or Cloudflare Pages and GitHub Pages for custom domains. The whole exercise ran ◇6.44.
Line count barely moved between versions (614 to 658) while compute jumped from ◇1.26 to ◇4.59; the expensive part was visual validation through real browser renders at multiple widths, not code generation.
Skill discovery worked through plain language: instead of browsing menus, the user had WorkBuddy audit its own installation and name the right skill, a pattern worth copying in any agent tool.
The author calls the multi-width render calibration necessary when layout quality matters, effectively pricing design polish at roughly 3.6x the compute of a default generation.