Hallmark is a free design skill for AI coding assistants that refuses to let them ship the look everyone now recognises as machine made: the purple to blue gradient hero, Inter as both display and body face, three equal icon cards, and a call to action bolted on the end. It added roughly 1,485 stars in a single day to reach about 12,275 total, which puts it near the top of GitHub trending. Installing it takes one command and about two minutes, and it works in Claude Code, Cursor and Codex.

  • One line installs it everywhere: npx skills add nutlope/hallmark, or copy the skill folder in by hand.
  • It is a rule set, not a component library. Nothing gets added to your dependency tree and no runtime ships.
  • Four verbs: build by default, plus audit, redesign and study for code you already have.
  • MIT licensed, built by Nutlope and backed by Together AI, with 20 named themes and 21 page macrostructures in the repo.
Default LLM design output versus the Hallmark pipelineA default coding assistant emits the same hero, three feature cards and call to action for every brief. Hallmark inserts theme selection, macrostructure selection, a self critique pass and a 58 gate slop test before the code is returned. DEFAULT ASSISTANT Brief goes in Trained-in defaults fire Hero, 3 cards, CTA, footer Every brief looks the same WITH HALLMARK Same brief goes in Theme + macrostructure Six-axis self critique 58-gate slop test, then emit No dependency, no runtime, just rules the model has to clear genztech.blog
Fig 1 Hallmark does not add a component library. It adds forced decisions and a gate list between the brief and the code your assistant hands back.

What is Hallmark and why is it trending?

Hallmark is a skill file, not a package. The whole thing is a SKILL.md plus a references/ directory of markdown that your coding assistant reads before it writes UI. The repository was created on April 27, 2026, is MIT licensed, and its own description is blunt about the target: an anti AI slop design skill for Claude Code, Cursor and Codex. It is made by Nutlope and powered by Together AI, and the live gallery at usehallmark.com shows pages generated from different briefs that genuinely do not share a layout.

RelatedOpen Interpreter Setup: A Coding Agent for Cheap Models

The reason it is spiking now is that the problem finally became visible to everyone. Coding agents got good enough that people ship the first output, and the first output has a fingerprint. Hallmark names that fingerprint in references/anti-patterns.md: the purple gradient hero, Inter everywhere, the three column feature grid, card in card nesting, and the background-clip: text gradient headline. Once a rule set names a tell, the model can be told to fail on it. That is the entire mechanism, and it is why a pile of markdown outranks most design tooling on trending.

The differentiator the author leans on is structural variety rather than visual variety. A colour swap is not a new design. The repo ships 21 macrostructure files, including bento grid, long document, marquee hero, stat led, workbench, manifesto, quote led, type specimen and map diagram, and the skill picks one per brief and stamps it into a CSS comment so the next build has to pick something different.

How do you install Hallmark on Claude Code, Cursor or Codex?

There is no release binary and no npm dependency to add to your project. The supported path is the skills CLI from Vercel Labs, run through npx:

# install or update the skill, from the repo README
npx skills add nutlope/hallmark

Re-running the same command updates it later. If you would rather not run a third party CLI over your config directory, the README documents the manual route: copy SKILL.md and the references/ folder out of skills/hallmark/ in the repo into the location your tool expects.

# manual install targets, straight from the README
# Claude Code
~/.claude/skills/hallmark/
# Cursor: body of SKILL.md, no frontmatter
.cursor/rules/hallmark.mdc
# Codex: personal, or project-scoped
~/.codex/skills/hallmark/
.codex/skills/hallmark/

Note the Cursor difference. Cursor rules do not take the YAML frontmatter block, so you paste the body only, and you get a single flat rule rather than the progressive reference loading that Claude Code and Codex support. That matters more than it sounds: the skill is designed to pull in references/study.md or references/custom-theme.md on demand, and a flattened rule loses that.

What do the four verbs actually do?

Default behaviour needs no invocation. Ask your assistant to build a landing page and Hallmark takes over: it detects whether the brief carries creative intent, picks one of 20 named catalog themes or branches to a custom one off palette, selects a macrostructure, scores itself, then runs the gate list. Beyond that there are three explicit verbs. hallmark audit <target> reads existing code and returns a ranked punch list without editing anything, which is the safest way to try it on a real repository. hallmark redesign <target> keeps your copy, routes and information architecture and replaces the visual layer. hallmark study <screenshot or URL> extracts the macrostructure, type pairing and colour anchor from a design you admire, and can emit a portable design.md you hand to another tool.

The study verb has the sharpest edges. URL mode reads real HTML and CSS, so it can name exact fonts and exact colour values but cannot judge rhythm. It refuses template marketplace URLs outright, and emitting a design.md from a URL requires you to attest that the source is your own or a public reference for your own brand. If the page is auth walled or a JavaScript only shell, it falls back to asking for a screenshot.

How does it compare with the alternatives?

TraitHallmarkComponent libraryGeneric prompt
What shipsMarkdown rules onlyRuntime dependencyNothing
Structural varietyEnforced per buildSame primitives every timeNot enforced
Works on existing codeYes, via audit and redesignRewrite requiredAd hoc
Failure modeModel ignores a gateEveryone looks identicalReverts to defaults
CostFree, MITVariesFree

What are the gotchas before you rely on it?

First, it is a prompt, and prompts are advisory. Nothing in your build pipeline fails when the model skips gate 47. Hallmark asks the assistant to stamp its self critique scores into a CSS comment at the top of the file, and that stamp is your only cheap audit trail. If it is missing, the skill probably did not run.

RelatedOpenCut Setup: Self-Host the Free CapCut Alternative

Second, the numbers in the documentation already disagree with each other. The README says fifty seven slop test gates. The actual references/slop-test.md header says 58 gates plus the pre-emit self critique. That is a harmless drift, but it tells you the rule set is moving faster than the marketing copy, so pin a version if you need reproducible output. The skill declares version 1.1.0 in its frontmatter.

Third, token cost is real. The skill plus its references is a large body of text your model reads before writing a line of UI, and the skill itself notes that two self critique revision passes are normal. On a metered plan, expect design tasks to cost noticeably more than they did.

Fourth, read the safety rail and mean it. The skill explicitly instructs the assistant never to delete production files, route trees or component directories without an approved file level plan, and to treat README and briefing documents as reference rather than copy to paste into the page. Those lines exist because early users pointed a redesign verb at a live codebase. Start with audit, on a branch.

What to watch · 2026
  • Gate drift. The gate count moved from 57 to 58 between the README and the reference file. Watch whether the project starts versioning the gate list properly.
  • Skills as a distribution channel. Hallmark installs through the Vercel Labs skills CLI, currently at 1.5.19. If design rule sets become a normal install step, that CLI becomes real infrastructure.
  • Copycats. A markdown rule set is trivially forkable. Expect vendor branded variants that quietly steer output toward one component library.

Our take

The interesting thing about Hallmark is not the taste, though the taste is good. It is the admission underneath it, that the fastest way to improve an AI coding assistant right now is not a better model but a better refusal list. Everything Hallmark does could be a paragraph in your project instructions. The reason it has twelve thousand stars is that almost nobody writes that paragraph, and the ones who try write three vague sentences about clean minimal design, which is exactly the phrasing that produces slop.

The honest caveat is that a rule file cannot give you judgement. Hallmark will stop your assistant from shipping a gradient headline. It will not tell you the page is boring. Treat it as a floor rather than a ceiling: install it, run audit on your worst page, fix the punch list, and keep making the actual design decisions yourself.

Primary sources

Original analysis by GenZTech. Facts verified against the Hallmark repository and its shipped reference files.