← ClaudeAtlas

design-brieflisted

Use when a work item has real UI surface and its visual direction isn't settled - a request to make something look good, design a landing page or a screen, pick an aesthetic or a design system, fix a UI that reads as generic AI slop, or turn a vague "looks nice" into visual done-criteria the loop can verify. This is the design front-end `/groundwork` composes to settle foundation tokens and `/plan-grill` composes for a UI feature. Composes the host's design skill and sources one on a gap; hardcodes no aesthetic or stack. Not for auditing existing UI code against accessibility or style guidelines - that is a host design-review skill (`/web-design-guidelines`) or `/review`.
yoelgal/agent-tools · ★ 1 · Web & Frontend · score 67
Install: claude install-skill yoelgal/agent-tools
# Set a design direction before any pixel When a request carries visual intent, the model's default reach is a generic template - and a generic template is exactly what reads as "AI slop." This skill sets the direction and the criteria so the loop has something to build toward and grade against. It does not paint the pixels; it decides what "done" looks like and how you will know. ## 1. Read the brief, state the direction as decisions The brief is the product's audience, its tone, and its constraints - read those before opening a component. On an existing codebase, also read the token source and at least one representative component before stating any direction: a direction stated over an unread codebase is a reinvention, and reinventing what the repo already settles is a finding. Decide the register next: this surface either *is* the product (marketing, landing, portfolio - the design is the message) or *serves* the product (app UI, dashboard, settings - the design is the tool). The registers carry opposite defaults for type, color, and motion, and opposite failure tests - a face surface fails when its look is guessable from the product category alone (the same test `slop-and-checks.md` runs as the guessability test's first altitude - one check, cleared once); a tool surface fails when a user fluent in the category's best tools would pause at an off component. Record the register on the direction card; the tell audit in step 5 filters by it. Then state the direction in