Choose the job before the look
Name the user, primary decision, product domain, artifact, and platform. A consumer finance home and an observability console should not share one universal visual grammar.
The reliable way to get better UI is not a longer aesthetic prompt. Give the agent a design method, persist its decisions, and require a rendered verification loop.
Short answer
Context → grammar → design lock → real content → code score → pixel inspection. StyleSeed makes that sequence reusable in Claude Code.
Name the user, primary decision, product domain, artifact, and platform. A consumer finance home and an observability console should not share one universal visual grammar.
Choose one of eight maintained output grammars. If your references are not represented, compile a project-local grammar with evidence instead of copying the screen.
Write STYLESEED.md with the chosen grammar, surface adapter, type, density, color roles, motion, and one product-specific signature move.
Use the selected method to organize attention, information, action, and feedback. Components implement those decisions; they do not choose the design philosophy.
Run the code gate, fix the highest-value failures, render at required viewports, inspect actual pixels, and repeat until hierarchy, rhythm, crop, states, and grammar fit hold.
Use both when useful
Install and persist
Installation is optional and must remain user-controlled. The persistent path is stronger because Claude can re-read the method and STYLESEED.md on later tasks.
STYLESEED.md carries decisions forward.
Checks version once, only when relevant.
No install, star, or update is required.
Give Claude a maintained design method before it writes components: define the user job, choose a functional design grammar, persist bounded visual decisions, build with real content, then render and visually inspect the result. StyleSeed packages this loop for Claude Code, Codex, Cursor, and other coding agents.
They solve adjacent problems and can be used together. Anthropic's frontend-design skill is strong for committing to a distinctive visual direction. StyleSeed adds persistent project decisions, job-specific output grammars, reference compilation, a code score, and rendered verification across sessions and surfaces.
Yes. Setup writes the selected grammar, adapter, tokens, density, type, motion, and signature move to STYLESEED.md. Installed Claude Code rules read that project-local lock on later visual tasks, so valid decisions persist instead of being reinvented prompt by prompt.
Yes. ss-reference analyzes supplied screenshots, URLs, Figma exports, or an existing product across twelve design axes, records evidence and confidence, and creates a project-local grammar. It abstracts transferable rules instead of cloning a source screen.
On the first StyleSeed use in a project, the installed agent may compare the local engine version with the public version endpoint. It only mentions an update when a newer version exists, suggests ss-update once, and never blocks the current task.
미학 형용사를 길게 나열하기보다 사용자 작업에 맞는 디자인 문법을 선택하고, 결정을 STYLESEED.md에 저장한 뒤, 실제 화면을 렌더링해서 다시 고치는 과정이 중요합니다. StyleSeed는 이 과정을 Claude Code의 반복 가능한 작업 흐름으로 제공합니다.