Every large language model trained on the public web learned a visual dialect: the three-card layout, the centered hero with gradient text, the rounded corners on everything, the pastel-and-white color palette. That dialect is now so recognizable that a site built by an AI tool announces itself within seconds. Nutlope/hallmark, a design skill released on GitHub by Together AI, is a deliberate break from that pattern. It refuses to look AI-generated.
Hallmark is not a design system or a component library. It is a skill file that plugs into Claude Code, Cursor, and Codex — the three dominant AI coding tools — and injects a rule set that overrides the models’ default output behavior. The skill picks a macrostructure for a brief, dresses it in one of twenty themes, runs fifty-seven slop-test gates plus a pre-emit self-critique, and rejects the on-distribution defaults every LLM was trained into. Two pages by Hallmark for two different briefs feel like different sites, not color swaps of the same template.
The tool exposes four verbs. The default verb builds new UI: it selects a macrostructure, applies the rule set, and runs the slop test before returning code. hallmark audit <target> scores existing code against a catalog of anti-patterns and returns a punch list without making edits. hallmark redesign <target> discards the existing structure, preserves copy and information architecture and brand, then rebuilds with a different fingerprint. hallmark study <screenshot | URL> extracts the DNA from a design the user admires — macrostructure, type pairing, color anchor — and refuses to produce pixel clones or paid-template derivatives. It can optionally emit a portable design.md for handoff to other AI tools.
The catalog of twenty themes includes names like Bubble, Distil, Cold Snap, Cinder, Ferns & Fathom, and Hollowback Apiary. Each theme is a self-contained HTML and CSS page stamped with its macrostructure in a CSS comment. The full set lives at usehallmark.com. When a brief carries creative intent that no catalog theme fits, Hallmark switches to a Custom mode and designs the page from scratch: a made-to-measure palette, type, and layout, with the same fifty-seven slop-test gates and no template underneath. The protocol for Custom mode lives in custom-theme.md and stays a quiet branch — vanilla briefs never see it.
The fifty-seven slop-test gates are the heart of the project. They encode the specific visual tics that LLMs default to: symmetrical layouts when asymmetry would serve, the same three font stacks, the predictable spacing rhythm, the color palettes that cluster around the same hue ranges. A pre-emit self-critique step forces the model to evaluate its own output against these gates before returning code to the user. If the output fails, the model regenerates. The effect is that Hallmark-generated pages carry structural variety that reads as human-crafted, not as templated.
This matters for a reason that goes beyond aesthetics. The AI coding tools market is consolidating around a narrow set of visual defaults because the models were trained on the same corpus. Every Claude Code or Cursor user who builds a landing page without a guardrail like Hallmark is contributing to the homogenization of the web. The visual dialect of AI-generated design is becoming a pollution problem — not just for designers who can spot it instantly, but for users who internalize it as the normal look of software. Hallmark is a countermeasure that operates at the tool level, not the policy level.
The project is MIT-licensed. Installation is a single command: npx skills add nutlope/hallmark. Re-run to update. The skill file and references directory can be copied into ~/.claude/skills/hallmark/ for Claude Code, .cursor/rules/hallmark.mdc for Cursor, or ~/.codex/skills/hallmark/ for Codex. Worked examples live in docs/recipes.md and docs/study-examples.md.
The implications for AI builders are straightforward. The current generation of coding tools produces output that is statistically average by design — that is what a language model does. Hallmark is a recognition that statistical averageness is a liability for any product that needs to signal quality, taste, or differentiation. A startup using Claude Code to build its landing page without a guardrail is shipping a page that looks like every other AI-generated landing page. A startup using Hallmark is shipping a page that could have been built by a human designer with a point of view.
The harder question is whether this approach scales. Hallmark’s fifty-seven gates are hand-crafted rules written by people who know what AI-generated design looks like. As the models improve, the slop-test gates will need to evolve. The visual tics that Hallmark catches today may shift as training data changes and as more AI-generated content enters the corpus. The project is a snapshot of a moving target.
What is durable is the architectural choice: a skill file that sits between the model and the user, intercepting output before it reaches the screen, and applying a quality gate that the model cannot easily learn to satisfy on its own. That pattern — a lightweight, user-owned filter on model output — has applications far beyond design. Code quality, security, accessibility, brand consistency. Hallmark is a demonstration that the right intervention point is not the model itself but the tool layer around it.
The project ships twenty themes, four verbs, and fifty-seven gates. The most interesting number is the one that is missing: the count of pages that now look different because someone installed a skill file instead of accepting the default. That number is growing.