Ask a coding agent for a landing page and you often get the same page again: a small label above every heading, one coloured word in the headline, a sign-up box at the bottom. It works, but it looks machine-made. Hallmark is an attempt to fix exactly that habit at the source, inside the agent.
What is Hallmark, and who makes it?
Hallmark is a design skill for coding agents. A skill, in this sense, is a package of instructions and reference files that an agent reads before it does a particular job. The official repository describes Hallmark as a skill for Claude Code, Codex, Cursor and OpenCode that refuses to look AI-generated. It is marked as made by Together AI and is released under the MIT license, so you can read, change and reuse it.
Which problem does it address?
Hallmark builds web pages and then lints them, meaning it checks the result against a list of known patterns. The repository names the tells it looks for: the eyebrow label above every heading, tracked-capital labels, one coloured word in a headline, a page of empty frames or nothing but photographs, and an email sign-up at the end of every page. It also flags a hero section that misses its screen and a form that says thanks without sending anything. That last one is not a style issue at all; it is a broken feature that happens to look finished.
How does it work?
The core idea is the theme. Hallmark ships fifty of them, and each is a “material”: paper, type, colour, surfaces, its own components and texture. A theme does not decide what a page contains or in what order. Each one is a single file under 8 KB, and a build reads exactly one. If you say “custom”, Hallmark drops the themes and designs from your prompt alone, under the same rules. Version 2 also builds from the job itself: a booking page books, a shop sells, a newsletter subscribes. It can take its own screenshots when you point it at a URL, a dev server or an app, and it draws on 755 openly licensed photos, credited on the page.
The checks come in three tiers. The Floor never ships: contrast, focus, reduced motion, invented facts and a broken phone layout. Reflex gates catch the default habits a model falls into; a build may overrule one, but only with a reason that gets recorded. Finish checks cover craft detail and never block a release. The linter, sloplint, checks a rendered page at eight widths, and then a subagent reviews it with fresh eyes. Static checks also run without a browser.
How do you install and use it?
Installation is one command: npx skills add nutlope/hallmark -g. The -g flag installs it for every project; leave it out to install into the current one. You need Node 18 or later. The browser checks additionally need puppeteer-core and Chrome; without them, the static checks still run. In Claude Code the skill lands in ~/.claude/skills/hallmark, and in Codex CLI in ~/.agents/skills/hallmark. For existing pages there are two commands: /hallmark redesign <path> rebuilds a page while keeping its content and wiring, and /hallmark audit <path> only reports. Before a redesign, sloplint can record links, fields, handlers and headings, and the redesign fails if any of them go missing.
What are its limits?
Hallmark does not invent content. Its pre-flight step asks for real material such as photos, screenshots and numbers. It does not use stock sites like Unsplash, Pexels or Pixabay, because their terms bar AI tools from picking their images. And it does not start from an existing design unless you name one. Hallmark itself is a set of instructions; the actual work, and any model usage, still runs through the agent you already use.
Editor’s assessment
The fifty themes will get the attention, but the more useful part is the Reflex tier. It does not simply forbid habits; it lets the agent break a rule and makes it write down why. That turns a silent design choice into a decision a person can review. My practical suggestion: start with /hallmark audit on a page you already have. You get a list of problems without any rewrite, and you decide which fixes are worth handing to the agent. Who takes the next step is then your call, not the model’s default. This is our reading of the documentation, not a test result; how well the gates hold up on your own pages is something only your own audit will show.
