Hallmark

Why it's ranked here

The rules are specific enough to argue with: each named pattern gets a plain reason it reads as machine made and an exact swap, and one rule forces every color and typeface to route through a named token so nothing gets improvised mid-draft. That specificity is the real asset, not the promise of variety. Everything here is prose the host assistant is trusted to obey, not code that enforces it: no test suite, no linter, no build step checks the output against what it describes. The checklist is the strongest part of the package, with the standing caveat that compliance is only as good as the model reading it.

What's good

The strongest parts are the ones that ban invention rather than dictate taste: no fabricated metrics or testimonials, no color or font value outside a named token, and a hard floor for responsive behavior at four specific screen widths with no horizontal scroll allowed. Each anti-pattern entry explains why the pattern reads as machine made before it prescribes the fix, which teaches rather than just filters. The dispatch logic also tracks what it already built for a project so a second page in the same codebase does not repeat the same structural shape as the first, which is a harder problem than picking one good template.

Tradeoffs

Nothing here is enforced by tooling: the whole mechanism is markdown instructions the host assistant is asked to follow, so a model pulled toward its own training defaults could ignore or partly apply the rules with no error surfaced. There is no automated check that verifies an output actually passed the review it claims; the self-scoring step is a comment the model writes about its own work, not an independent gate. Coverage is narrow by design too: it governs visual and structural choices, not backend logic, and accessibility beyond contrast and mobile width, or performance beyond a couple of named rules.

How to use it well

Point it at a written brief with a real audience, use case, and tone; the more specific the brief, the less the assistant has to guess. Run the audit mode on an existing page first if the goal is fixing a live site, since that mode only returns a punch list and never edits on its own. It suits teams shipping several distinct landing pages who want each to feel like its own site, not a product needing many pages to look like one consistent brand: the tool treats a repeated shape across pages as the failure it exists to prevent.

Observed

License

License: MIT (README.md, package.json)

Packaging

Distributed as a skill for three specific AI coding assistants (package.json skill.harnesses: claude-code, cursor, codex), not as a standalone CLI or library

Install surface

Install surface: npx package install, or manual copy of skills/hallmark/SKILL.md and its references directory into a harness-specific skill folder (README.md)

Dependencies

No runtime dependencies declared in package.json; the only script runs a static file server for the example site

Interfaces

Interface is natural-language verbs invoked inside the host assistant (default build, audit, redesign, study), not a programmatic API or CLI binary

Demo gallery

Repository also ships a static demo gallery (site/js/main.js) implemented in plain JavaScript with no framework or build tooling

Design rules

Rule set mandates OKLCH color values and forbids inline color or font-family declarations outside a named CSS custom-property token (skills/hallmark/references/color.md)