Ever asked an AI coding agent to build a landing page and got back the same generic, rounded-corners, centered-layout template that every LLM defaults to? Yeah, me too. I’ve seen more “modern SaaS landing pages” than I can count — and they all blur together after the third one. So when I spotted Hallmark trending with 4.6k stars, a design skill that “refuses to look AI-generated,” I had to put it through its paces.
What Is Hallmark?
The short version: Hallmark is a skill for Claude Code, Cursor, and Codex that picks a macrostructure for your brief, dresses it in one of twenty themes, runs 57 slop-test gates plus a pre-emit self-critique, and hands back something that doesn’t scream “LLM wrote this.” Made by Nutlope (Hassan El Mghari) with Together AI. MIT licensed. Works right now.
The Four Verbs
So I sat down for an afternoon and threw random briefs at it. Here’s what I found:
| Verb | What It Does | My Take |
|---|---|---|
| (default) | Build new UI from a brief — picks macrostructure + theme, runs slop test | Solid for first drafts. Not subtle, but that’s the point. |
hallmark audit <target> |
Score existing code against anti-patterns. Read-only. | Worth the install alone. Found 14 issues in a page I’d been polishing. |
hallmark redesign <target> |
Ditch the structure, keep copy + IA + brand, rebuild | Aggressive. Killed some layout choices I liked. Fair trade-off. |
hallmark study <screenshot|URL> |
Extract design DNA — macrostructure, type pairing, colour anchor | Works best with clean screenshots. Mediocre source = mediocre DNA. |
Putting It to the Test
So I pointed Hallmark at a landing page I’d been working on with Claude Code. Without the skill loaded, Claude defaulted to a predictable hero → three feature cards → CTA layout. But with Hallmark active, the same brief produced an asymmetric grid with stacked typography and a palette I wouldn’t have picked myself — yet it worked. Plus, pressing T lets you cycle through themes quickly — useful when you want to test out alternatives without rebuilding from scratch.
Still, the 57 slop-test gates are where this thing earns its keep. Hallmark checks for overused AI patterns: gradient-heavy backgrounds, centered everything, generic icon grids, predictable card rows. If the output trips too many gates, it refuses and re-rolls. The pre-emit self-critique caught a few layouts I would have accepted at first glance. And honestly, they deserved to be caught.
But it’s not magic. Still, the study verb needs clean source material — messy screenshots produce muddy output. And the gate system can be overly aggressive — I had one brief where it kept rejecting layouts until I trimmed the requirements back. You’ll definitely spend the first hour just learning which themes work for which kind of project.
Installing Hallmark
Installation is trivially simple though:
npx skills add nutlope/hallmark
Re-run that anytime to update. Or drop the files directly into ~/.claude/skills/, .cursor/rules/, or ~/.codex/skills/ depending on your tool.
Who Is It For?
Honestly, if you’re building UI with Claude Code, Cursor, or Codex and you’re tired of the same default layouts, Hallmark is a no-brainer install. Also useful if you’re auditing existing AI-generated pages for quality — the audit verb alone catches things that are easy to miss after staring at a page too long. For designers who want to feed inspiration into their AI tools, the study verb opens up a workflow I haven’t seen elsewhere: extract the DNA from a real page and let Hallmark remix it. If you enjoy profiling what makes your AI agents tick, you’ll find Ditto pairs well with this kind of workflow.
Though if you’re not using AI coding agents to build UI at all, or you’re happy with the default output and just want something up fast, Hallmark adds complexity you don’t need. The learning curve for the four verbs and twenty themes is real — you’ll spend the first session just clicking through themes and reading gate failures.
Bottom line: Hallmark is a genuinely useful design skill for AI agent workflows. It doesn’t just apply a different template — it actively fights the distribution every LLM was trained on. The four verbs give you real control: build new UI, audit existing work, redesign from scratch, and learn from designs you admire. 4.6k stars, still climbing, MIT licensed. I installed it in 30 seconds and I’m keeping it loaded for every Claude Code session from here on.