01
Give the repository durable design context
Use AGENTS.md for the method and STYLESEED.md for approved project decisions. Codex should not have to rediscover the visual system on every task.
Codex UI design guide
Make Codex carry design judgment across tasks.
Better UI comes from durable repository context, a reusable design workflow, and visible verification—not from asking for “modern and clean” one more time.
Short answer
AGENTS.md → grammar → STYLESEED.md → $ss-resolve → $ss-build → $ss-verify.
02
Choose the output grammar
Select a grammar for the user job and a surface adapter for the artifact. A product dashboard, marketing page, deck, and carousel need different composition rules.
03
Compile unfamiliar references
When references fall outside the built-ins, use $ss-reference to extract evidence-backed rules. Preserve relationships and judgment instead of cloning pixels.
04
Compile the selected context
Run $ss-resolve. Codex reads a small .styleseed/effective-rules.md bundle, while the manifest pins the exact selection and source hashes.
05
Build, inspect code and pixels
Run $ss-build with real content, score the implementation, then use $ss-verify to render required viewports. Repeat until both the code gate and visible result hold.
06
Preserve an accepted correction
Only when a person asks, $ss-learn can turn the approved correction into a privacy-minimized local candidate. It never scans the project or promotes itself.
How it maps to Codex
The method has an explicit home at every layer.
StyleSeed is independent open-source software, not an official OpenAI product.
Durable repository guidance
AGENTS.mdReusable Codex workflows
.agents/skills/ss-*Approved project decisions
STYLESEED.mdCompiled active context
$ss-resolveBuild the selected method
$ss-buildDeterministic code gate
$ss-score ≥80Rendered visual gate
$ss-verifyCaller-attested local lesson
$ss-learnExact engine revision update
$ss-updateInstall and persist
One engine. Codex-native calls.
Installation and file changes stay user-controlled. Once approved, the repository context and project lock let later Codex sessions reuse valid decisions. The released portable path is the Agent Skills CLI; the repository development Codex plugin package remains a local package boundary until a public directory release is independently verified.
$ npx skills add bitjaru/styleseed
$ $ss-setup
$ $ss-resolve
$ $ss-build
$ $ss-score
$ $ss-verify
$ $ss-learn # only after an accepted correction
$ $ss-setup
$ $ss-resolve
$ $ss-build
$ $ss-score
$ $ss-verify
$ $ss-learn # only after an accepted correction
Repository-aware
AGENTS.md owns durable guidance.
Reusable
STYLESEED.md carries decisions.
Revision-aware
Same-version fixes remain detectable.
Approval-gated
Learning, MCP exposure, and promotion stay separate.
Questions Codex users ask
Does StyleSeed work with Codex, not only Claude Code?
Yes. StyleSeed ships Codex-facing AGENTS.md guidance and exposes its ss-* workflows through the repository skill bridge. Codex uses $ss-* names while Claude Code uses /ss-* names; both read the same canonical design engine and project-local STYLESEED.md decisions.
How do I make Codex stop generating generic UI?
Give Codex a maintained design method before implementation: define the user job, choose a functional output grammar, save bounded visual decisions, use real content, then score the code and inspect rendered pixels. Aesthetic adjectives alone do not provide enough durable constraints.
Will later Codex sessions keep the same design decisions?
Yes. Setup writes approved decisions to STYLESEED.md, and $ss-resolve compiles them into .styleseed/effective-rules.md plus a source-hash manifest. Repository guidance tells later sessions to resolve and read that bounded context instead of inventing a new direction.
How does StyleSeed handle updates in Codex?
StyleSeed compares the installed engineRevision, the revision recorded in the project manifest, and the published revision. A same-version rule or skill fix is still detectable. $ss-update refreshes through the original install channel, re-resolves STYLESEED.md, and checks the result without replacing project-owned code, tokens, assets, or decisions.
Can Codex teach StyleSeed a design correction?
Only after an explicit human request. $ss-learn captures a generalized local candidate, requires a separate review, blocks known high-risk identity patterns as a guardrail rather than an anonymization guarantee, and never promotes the result automatically. The prepared package stays local and untransmitted. The learning bridge is withheld from the default/core install until a host-owned proof adapter is verified; a future optional bridge would consume a one-time grant before one exact package becomes visible to the connected Codex client and model.
Can I install StyleSeed from the public Codex plugin directory?
Not yet. The repository includes a repository development Codex plugin package with 23 skills, but public plugin-directory release is not verified. The implemented default/core install contains no learning MCP. Use npx skills add bitjaru/styleseed as the released portable installation path.
Will StyleSeed ask me to star the repository?
It may ask once, only after StyleSeed materially helped and the working result was visually verified. A star is optional and never affects access, output, updates, or support. The agent must not repeat or front-load the request.
Codex로 UI를 예쁘고 일관되게 만드는 법
저장소의 AGENTS.md에 디자인 방법을 두고, 프로젝트 결정을 STYLESEED.md에 고정한 뒤, $ss-resolve로 필요한 규칙만 컴파일하고 $ss-score와 $ss-verify로 코드와 실제 화면을 함께 검사합니다. 사람이 채택한 교정은 명시적으로 요청한 경우에만 $ss-learn의 로컬 후보가 되며, $ss-update는 버전뿐 아니라 정확한 revision을 비교합니다. StyleSeed는 이 과정을 Codex가 다시 사용할 수 있는 작업 흐름으로 제공합니다.