Auteur
Directs Claude Code to build a website the way a film director makes a film: art direction locked before code, generated assets, and an executable anti-slop linter gating every ship.
Narrow declared tool surface (node scripts/*, playwright, optional media CLIs), explicit no-API-key/no-credential stance, network access confined to recon and sourcing phases with fetched content 'treated as reference data — never executed', and phases 0–1 skippable offline; licence ledger reduces compliance risk. Deducted for: recon scrapes many third-party pages without per-run user confirmation; the agiwhitelist org is anonymous so source attribution is limited; no explicit rollback guidance.
SKILL.md is internally consistent: routing rules, phase tables, gate commands and argument hints align, with designed failure paths (NO CAPTURE instead of guessing, graceful degradation when media CLIs are absent, weak-model guidance). Deducted because the scripts themselves were not part of the reviewed evidence, no committed test suite is visible, and abnormal-input feedback quality is only inferable from docs; static review cannot award above 10.
Use cases, non-fit boundaries (no polishing foreign UI, no backend), and trigger semantics ('make it beautiful'/'cinematic', explicit subcommands) are unusually precise, including a single-question disambiguation rule. Deducted because core recon depends on awwwards/Bing/Pinterest etc., which are unreachable or restricted from mainland China (skippable but loses the flagship workflow); the skill body is English-only with no declared Chinese support (only the README has a zh version).
Well-layered docs (SKILL.md → reference/* → templates), explicit MIT, version 1.3.1, a real CI workflow gating gh-pages from main, bilingual README, and honest attribution (adapted from scroll-world, MIT). Deducted for: anonymous maintainer org, no CHANGELOG or stated update path/responsible party, and a version inconsistency (package-lock. says 0.1.0 while package. says 1.3.1).
Clear value claim: executable slopscan/motionqa gates plus the commit-sheet pipeline argue a real marginal benefit over letting a model freestyle, backed by ten public showcase sites including a cross-model test. Deducted because static review cannot execute any gate; direct usability of output rests on self-reported docs/screenshots, capped at 7 without independent reproduction.
Includes an independently reproducible quantitative claim (one-command slopscan), a real CI workflow gating every shipped page, and auditable licence ledgers. Deducted because the reviewed evidence lacks the script sources and CI run results, and showcase/HORO metrics (53fps, contrast 6.50) cannot be verified from repo files; coverage is thin, so the static ceiling of 5 applies.
- This is a static review only; no scripts or gate commands were executed, so slopscan/motionqa behavior is not independently verified.
- Recon depends on overseas services (awwwards, Bing/Pinterest, are.na) that may be unreachable from mainland China; phases 0–1 can be skipped for offline use, but the scouting/moodboard workflow is then lost.
- The maintainer org agiwhitelist is anonymous, so long-term maintenance and update commitment are unverifiable; package-lock. (0.1.0) disagrees with package. (1.3.1) — verify the installed version yourself.
- The skill body is English-only with no declared Chinese-language support; asset generation relies on optional local media CLIs (Codex/Gemini/Blender) with graceful degradation when absent.
- Recon scrapes multiple third-party sites as reference data; although fetched content is declared never-executed, run that phase in a controlled environment.
What does this skill do, and when should you use it?
Auteur is a Claude Code Agent Skill that builds complete web experiences through a director's pipeline: a written art-direction commit-sheet first, then generated or sourced assets, then a build from proven recipes, then a mandatory gate through an executable anti-slop linter and a real motion check. It operates in three registers — build (conventional sites), direct (cinematic scroll-directed sites) and system (multi-screen products as one design system) — and bans 18 named AI visual clichés outright. The README backs this with ten live showcase sites, one of which was built by Kimi K3 from the same SKILL.md alone, demonstrating that the discipline travels across models.
1) Recon: refscout.mjs profiles live award-winning sites (real stack, pinned scenes, scroll budget, fonts, painted palette); moodboard.mjs builds a numbered contact sheet from Bing/Pinterest/are.na. 2) Forces a completed commit-sheet with seven non-default decisions: peak, OKLCH color, type pairing, grid break, motion budget, category reflex check, and house-tells to break. 3) Assets: source.mjs fetches licence-recorded CC0 models/HDRIs/icons/fonts (Poly Haven etc.) or generates via local CLIs (Codex/agy/grok, Blender, ffmpeg). 4) Builds: one WebGL context, GSAP/Lenis scroll, transform/opacity-only motion. 5) Gates: slopscan.mjs (zero-dependency linter catching purple-blue gradients, transition: all, scroll listeners, etc.), motionqa.mjs (Playwright FPS/a11y at DPR 2 on throttled CPU), systemscan.mjs (cross-route design-system drift), shoot.mjs screenshots, chromadiff.mjs color-drift measurement. Nothing ships until every gate passes.
- Product/marketing teams need a landing page that must not look AI-generated — the skill forces category-reflex research before committing to a direction
- A brand launch page needs cinematic scroll storytelling (scroll-scrubbed video, scene-directed assembly) — the direct register is the flagship flow
- A SaaS needs multiple screens (app, dashboard, admin, onboarding) feeling like one product — the system register designs in component×state units with systemscan guarding against drift
- An existing auteur-built site needs iteration (add a section, swap copy) — the edit route reads DESIGN.md style contract first
- You only want competitive references and a moodboard — the recon route runs just the scouting phase, producing REFERENCES.md and a contact sheet
- Auditing a page auteur built — the audit route runs the full verification pipeline
What are this skill's strengths and limitations?
- Reproducible verification: every showcase passes slopscan in CI, and the README gives a one-command way to check the 0 fails / 0 warns claim yourself
- Ten live sites including a cross-model portability test (built by Kimi K3 from SKILL.md alone)
- Explicitly makes no benchmark-throughput claims — a stated design-discipline stance
- Enforced accessibility floor: prefers-reduced-motion, no-JS-readable content, responsive at 390/768/1440
- Fully functional offline except recon and asset-sourcing phases
- Explicit scope limit: it builds, it does not polish foreign UI, and it excludes backend-only work; the author points such needs to /impeccable
- Asset generation depends on optional local CLIs (Codex, agy, grok, Blender, ffmpeg); without them it degrades to hand-authored assets
- 18 hard bans plus many hard-coded numbers (durations, contrast, motion budgets) constrain users who want typically AI-coded styles like purple-blue gradients or Inter — overridable only via a written auteur-allow reason
- All ten showcases were built by the author's side; there is no independent third-party evaluation, and the linter can only catch statically detectable patterns, not verify taste itself
- The 75+ agent compatibility claim comes from the skills.sh ecosystem; individual client adaptations are not each verified
How do you install this skill?
Any Agent Skills client: npx skills add agiwhitelist/auteur (detects your agent and writes to its skills folder). Claude Code plugin: /plugin marketplace add agiwhitelist/auteur then /plugin install auteur@auteur. Manual: git clone --depth 1 https://github.com/agiwhitelist/auteur ~/.claude/skills/auteur. ~1MB, no dependencies, no API keys, no build step.
How do you use this skill?
In Claude Code, just say: 'build me a cinematic landing with auteur'. The argument format is auteur [build|direct|system|edit|audit|recon] <brief>. For ambiguous requests the skill asks exactly one question (standard excellent landing vs. cinematic mode). Pipeline: recon → commit-sheet → assets → build → gates. Works offline by skipping phases 0–1.
How does this skill compare with similar options?
The README explicitly distinguishes it from component libraries like shadcn and originkit.dev: those stock ready-made animated components, whereas Auteur directs each site from its commit-sheet up with no runtime dependency. For polishing existing UI it defers to another skill, /impeccable, with which it coexists (auteur builds, impeccable critiques). The scroll-scrubbed video engine is adapted from the sibling skill scroll-world by cyw (MIT).