The stack, and why

Each option is built around a technique taken from real sites in the library. The builder reads that reference's signature moves and component notes, and watches its motion clip before claiming it saw the motion. The doctrine is to reproduce the craft layer (grid, spacing rhythm, type scale, motion, signature moves) faithfully, since no one owns that, while protecting the identity layer (a brand's palette, wordmark, fonts, photos and copy) absolutely.

A hook records every library read and whether it succeeded. An option cannot be marked ready unless it names a reference the project actually read, or says plainly that it is ungrounded because the library could not be reached. Verification runs the checks as commands and stores their output; a scope that has to be a command cannot be passed by declaring it. A check that could not run is named as such, never counted as a pass.

An independent critic, in a separate agent, reviews the picked direction before it spreads across the site and the finished site before handover, and a separate builder makes the corrections, within four rounds in total. Then the site is compared with exemplar sites from the library, each comparison run twice with the images swapped, in a fresh agent. It passes only when it reads as comparable or better. A worse reading goes back to the critic loop.

Before verification, every factual claim the built pages make (years, counts, ratings, prices, specifications, guarantees, testimonials) is listed with its page and the client source it rests on. Verification fails a claim the source does not support, a quote that is not actually on the page, or two pages that give different values for the same thing. We added this after a build invented a founding year, a specification and testimonials that passed every other check.

The starter ships a sitemap, a robots file that blocks crawling on previews and allows every crawler in production, AI crawlers included, canonical links, social cards, structured data and an llms.txt built from the site's own facts and pages. The SEO check runs over the built site, and a partial result fails rather than passing with unknowns.

Each option is built around an interaction, not decorated with fades afterwards. The builder uses an animation library when the choreography needs one and browser primitives when they are enough. Touch and keyboard get the same information as a pointer, reduced motion gets a considered alternative, and with JavaScript off the content and navigation still work.

If Higgsfield is installed, Palate asks once before using it. Stills can appear while you compare options, at most two per option, and a scroll film can follow once you pick, for the opening, one section or the whole page when the design calls for one. It checks the price before each job and never spends past your cap. Generated images are never presented as the client's real premises, people or products. Without Higgsfield you are never asked.

Yes, at the edges: the host, whether there is a CMS, the animation library, whether generated media is used. What does not change is the spine: working options before a full build, one design system per site, judges that are not the builder, and a release only from a verified build. Those are load-bearing, and pulling one out tends to pull the value out with it.

The process is fixed, because that is what makes the result repeatable: grounded in the reference library, built on one system, checked by judges that are not the builder. The output is not locked in. It is an ordinary Astro project in a folder you own.

Because taste is a system, not a coat of paint, and a system you re-decide every project is not a system. Every choice here trades some flexibility for work of a known quality. Each decision is written down with its reason and its trade-off, so it can be argued with rather than taken on faith. This page is that argument, in the open.