auteur gates every site it builds, and asks you to reproduce exactly one number

The project asks you to reproduce one number and refuses the others

There is an explicit disclaimer about numbers, and it is worth reading before anything else. There are no benchmark numbers anywhere in the repository. auteur is described as a design discipline rather than a system with a throughput figure, and the single quantitative claim is the linter result, which the page says you can reproduce in one command.

That framing is the reason to trust the rest of the page. Sites are described with specific, checkable properties instead of adjectives: a 145KB page with zero raster assets, a site holding 131,072 GPU particles on a curl-noise field, another whose measured lightness of 0.740 matches the 0.72 its own commit-sheet promised before the build. The gate results are quoted in the same form every time, as zero fails and zero warns.

So the deliverable is a process you can run, and the process has a stop condition. Whether the stop condition is worth having is a different question, and the rest of this is that.

The portability evidence is one site built by a different model

Nine of the ten showcased sites are presented as Claude builds. The tenth, HORO, was built by Kimi K3, which was handed the same SKILL.md and a brief, saw none of the other nine, and the page reports its gate results as slopscan 0/0/0, 53fps at DPR 2, worst contrast 6.50, and a page lightness of 0.740 against the 0.72 the commit-sheet had promised.

The claim drawn from that is that the skill is the thing that travels, not the model. The supporting detail is that the model named its four house tells in writing before building, in a category whose default reflex is black and gold, and came out lit, serif and green instead. That is a real signal about instruction following rather than about taste.

It is also one data point, from one model, on one brief. The landing page of the documentation site is counted as an eleventh artifact built by the discipline it sells, so the gallery holds eleven sites and one of them is the pitch.

Five stages, and the commit-sheet lands before any markup

The pipeline is presented as a director's sequence, enforced in order. Recon profiles live award-winning sites and pulls a contact sheet. The commit-sheet writes down one committed art direction. Assets are generated. The site is built. And the gate runs last, with nothing shipping until it passes.

The stage that does the work is the second one. The commit-sheet holds a single brand hue, a type system, a motion budget and named anti-references, all written before any markup exists. The stated purpose is to remove let's see how it looks drift: once the decisions are on paper, a later suggestion has to argue against a written position instead of against taste in the moment.

Named anti-references are the part with teeth. They are recorded as choices rather than preferences, so a build that drifts toward the category's default look has violated a line in a document the agent itself wrote.

Recon reads live sites and a numbered contact sheet, not memory

The first stage is where the art direction is supposed to come from. `refscout` profiles sites that have actually won something, and the profile is not a vibe summary: the libraries the page really loads, how many scenes it pins, its scroll budget, and the fonts and palette as painted rather than as declared. `moodboard` pulls a numbered contact sheet from Bing, Pinterest and are.na.

The stated rationale is that art direction gets decided from live material and a dated read of what the category currently reflexively does, rather than from the model's memory of what such pages looked like. That is a specific argument about staleness: a model asked for a cinematic landing page without reference will produce its own average, and its average is dated by its training data.

It also means the pipeline wants network access to run as described, which sits oddly beside the claim that the skill installs with no dependencies, no API keys and no build step. Those two facts are about different things, and the recon stage is where the difference shows.

Assets come from local CLIs, and provenance is part of the output

Imagery, video frames, depth maps and 3D geometry are produced by local tools: Codex and Gemini for image generation, Blender in headless mode, and Depth-Anything for depth. The routing is described as by cost and by what each tool is actually good at, which is a practical instruction rather than a fidelity claim.

Provenance is visible in the results. HALE is a CC0 microscope image taken apart on scroll, described as sourced rather than generated, and it sits under a CC0 HDRI. TRUE NOON computes a year of real sun positions live for your latitude and ships as 145KB with no raster assets at all, which is the extreme case of the same discipline: the page is code and data rather than pictures.

The distinction matters for a skill whose assets are generated by someone else's tool. A page that mixes a generated texture with a sourced photograph says which is which, so the reader of the gallery can tell what was made and what was taken.

The build rules are narrow on purpose

The build stage states its constraints rather than its ambitions. One WebGL context. GSAP with Lenis for scroll. DOM motion restricted to transform and opacity, and nothing else animated through layout properties. The recipes are named as a set: fluid simulation, GPGPU particles, 3D worlds, and scroll-morph state machines, and the video-scrub tier has its own reference file.

Four of the showcased sites are real-time WebGL, two are scroll-scrubbed video with a camera move, and the rest are lighter: a broadcast-glitch type treatment, a product register with five routes kept in step, a clock dial that is CSS and keeps your time.

That constraint list is what makes the gate possible. A page that only ever moves transform and opacity can be measured for frame rate and checked for drift in one WebGL context, which is why a motion check can be a script rather than an opinion.

Eight node scripts are the interface, and Playwright is the only dependency

The repository's own surface is the script list in `package.json`:

"scripts": {
  "slopscan": "node scripts/slopscan.mjs",
  "motionqa": "node scripts/motionqa.mjs",
  "shoot": "node scripts/shoot.mjs",
  "refscout": "node scripts/refscout.mjs",
  "moodboard": "node scripts/moodboard.mjs",
  "source": "node scripts/source.mjs",
  "systemscan": "node scripts/systemscan.mjs",
  "chromadiff": "node scripts/chromadiff.mjs"
}

`slopscan` is described as a zero-dependency linter that fails the build on concrete slop rather than vibes, which matches a script with no imports beyond node itself. The surrounding names describe the rest of the loop: `shoot` captures, `chromadiff` compares captures, `motionqa` is the motion check, and `systemscan` is the system register, the drift check across the five routes of the PROOF site.

The only devDependency is Playwright at 1.62.1, which is what a screenshot and motion check needs. Meanwhile the skill payload is described as about a megabyte with no dependencies, no API keys and no build step, and it is distributed as a SKILL.md with reference recipes, a DESIGN.md, a PRODUCT.md, templates and a Claude plugin manifest.

Three releases in one day, titled as statements

The release history is unusual in a way that says something about the project's voice. Three tagged versions landed on 2026-08-05: 1.2.0 titled the system register, 1.3.0 titled measure honestly, break your own habits, and 1.3.1 titled a state is not a variant. The manifest version is 1.3.1 and the license is MIT.

The last push is dated 2026-08-06, the day after those tags, and the repository is not archived. So the project has been quiet for roughly two months, which is worth keeping in mind before treating any of the eleven showcased sites as a moving target.

Distribution is broader than a single agent. `npx skills add agiwhitelist/auteur` is offered as one command for any agent, project-level or with `-g` for global, across a list of more than seventy-five agents. Claude Code can take it as a plugin that installs and updates in place, OpenClaw has its own install command, and anything that reads a SKILL.md can clone the repository into its skills directory.

Editorial conclusion

auteur is worth reading if you build marketing pages with an assistant and keep ending up with the same look, because the mechanism it enforces is a written art direction committed before code plus a linter that fails the build. Verify three things first: that your agent can load a SKILL.md at all, since that is the only integration surface; that the linter runs in your environment, since the two gates are node scripts and the only dev dependency is Playwright; and that the portability claim means what you think, because the evidence for it is a single site built by one other model, not a matrix. Do not expect a component library or a performance figure. The page itself rules both out, and points at shadcn or originkit.dev if what you want is animated parts you can drop in.

Frequently asked questions

Official sources

1. agiwhitelist/auteur on GitHub
2. License: MIT
3. Project website
4. README
5. Releases