Hallmark is an open-source design skill for AI coding assistants that sets itself one blunt goal: to stop AI-generated interfaces from looking AI-generated. Made by Together AI and released in April 2026, the MIT-licensed project plugs into Claude Code, Cursor, and Codex, and it has climbed past 12,000 GitHub stars on the strength of a problem every developer using these tools has felt — the sameness of the default, on-distribution layouts that large models reach for. Rather than another component library, Hallmark is a rule-set and process that pushes the assistant off those defaults toward pages that feel deliberately designed. ## The Problem It Targets Language models are trained on the average of the web, so when asked to build a UI they tend to converge on the most probable design: the same hero, the same card grid, the same spacing and gradients. The industry nickname for this is "AI slop." Hallmark's premise is that avoiding it is not a matter of taste alone but of process — you can encode the anti-patterns, test against them, and force a different result. Its tagline, a design skill that "refuses to look AI-generated," captures the intent: two pages built from two different briefs should feel like different sites, not color-swaps of one template. ## How It Works Hallmark runs a structured pipeline when it builds UI. It first picks a macrostructure appropriate to the brief — the overall page architecture — then dresses it in one of twenty themes. Before handing anything back, it runs what the project describes as fifty-seven "slop-test" gates plus a pre-emit self-critique, deliberately rejecting the defaults a model was trained into. Each page it produces is self-contained HTML and CSS, stamped in a CSS comment with the macrostructure it used, so the reasoning behind a layout is legible after the fact. Because the structure is chosen per brief rather than pulled from a fixed template, the output varies in shape, not just in palette. ## Four Verbs The skill exposes four modes of operation. The default verb builds new UI, applying the rule-set and running the slop test before returning. `hallmark audit <target>` scores existing code against the anti-patterns and returns a punch list without editing anything. `hallmark redesign <target>` discards the existing structure while keeping the copy, information architecture, and brand, then rebuilds with a different visual fingerprint. And `hallmark study <screenshot or URL>` extracts the "DNA" of a design you admire — its macrostructure, type pairing, and color anchor — while explicitly refusing to produce pixel-clones or copy paid templates, optionally emitting a portable design.md for handoff to other AI tools. Together these cover creating, evaluating, rebuilding, and learning from designs. ## Usability Hallmark installs as a skill into the AI coding tools developers already use, so there is no separate app to run — the design process happens inside the assistant. A live demo at usehallmark.com showcases the range, letting visitors cycle themes with a keypress and browse a gallery of example pages generated from distinct briefs, each a standalone HTML file. That self-contained output is a practical strength: the pages have no framework dependency and can be dropped into a project or studied directly. Being from Together AI lends the project credibility and suggests ongoing maintenance rather than a one-off experiment. ## Pros and Cons The strengths are a sharp, well-defined purpose, a genuinely different approach — encoding and testing against anti-patterns rather than offering more components — broad compatibility with popular AI coding assistants, framework-free HTML/CSS output, and a permissive MIT license. The limitations are worth naming. Design quality is subjective, so the slop-test gates encode one point of view; users who disagree with its aesthetic priorities may find its refusals opinionated. It targets front-end HTML and CSS work specifically, not general application logic, so its scope is narrow by design. And because it deliberately steers away from conventional layouts, its choices can occasionally trade familiarity for distinctiveness in ways that not every project wants. ## Outlook Hallmark arrives at a moment when AI can produce working UI in seconds but rarely produces memorable UI, and the gap between "functional" and "intentional" has become a common complaint. Tooling that treats design distinctiveness as a testable property, rather than hoping the model gets it right, is a credible response. As AI-assisted front-end work becomes routine, approaches that push against homogeneity may become a standard layer in the workflow. Whether Hallmark's specific rule-set becomes a reference point or simply one voice among many, its framing of the problem is timely. ## Who Should Use This Hallmark suits developers and designers who build front-end interfaces with Claude Code, Cursor, or Codex and want output that avoids the generic AI look — especially for landing pages, marketing sites, and product pages where distinctiveness matters. Teams that need conventional, highly familiar layouts, or whose work is primarily back-end and application logic, will get less from it. But for anyone frustrated by templated-looking AI designs, it offers a structured way to demand something better.