Hallmark, an open-source "design skill" for AI coding agents, gained about 12,000 GitHub stars in the past week and is near the top of GitHub trending with about 46,200 stars and 2,400 forks. It was created by GitHub user Nutlope and its README says it is "Made by Together AI". It is released under the MIT licence. The project is not new: AI Heartland counted 14,933 stars on 22 July 2026 and 28,173 on 7 September.
The problem it targets: pages generated by coding agents tend to look alike. Its docs list typical "tells" such as small eyebrow labels above every heading, identical card grids, heavy gradients and an email sign-up at the end of every page. Hallmark is not a library or a component: it is a set of rule files (SKILL.md plus references) that the agent loads before it designs.
How it works: it asks for real material (photos, real numbers, screenshots) instead of inventing it, picks one of 50 themes or designs a custom page, and lays out each page from scratch so sections differ between builds. It then lints the result in a real browser at eight widths with its sloplint tool and has a fresh-context sub-agent review it. Checks come in three tiers: Floor (never ships), Reflex (can be overruled with a reason) and Finish. Photos come from openly licensed Wikimedia Commons images, credited on the page. /hallmark redesign and /hallmark audit work on existing sites.
Numbers vary between versions. Articles from July describe 57 "slop test" gates and 20 themes; AI Heartland could only confirm gates up to 53 in the file. The current README and website describe 50 themes and the three-tier checks. The website lists Claude Code, Codex and OpenCode; the README also names Cursor.
It is part of a wider wave. In the same week, Impeccable by Paul Bakaus (Apache 2.0, about 100,700 stars) gained about 12,000 stars and Taste Skill (MIT, about 108,900 stars) about 5,700. Hallmark is also widely covered in Japanese, including on Zenn. Reviewers note limits: results still depend on the model and the brief, the rules add overhead for quick fixes, and automated gates cannot replace a human design review.
