Back to feed
News Story
Hacker News (AI filter)
1 sources

Hallmark: Anti-AI-Slop Design Skill for Claude Code, Cursor, and Codex

Hallmark is a design skill for AI coding assistants like Claude Code, Cursor, and Codex that aims to reduce low-quality AI-generated code ("slop"). It provides a set of design principles and prompts to improve output quality.

SynthePulse Insight · AI deep reading

Hallmark: Using 57 'Anti-AI-Vibe' Gates to Make AI-Generated Designs No Longer All Look the Same

Version 1 · 1 source

As AI coding tools make web design cheap and cookie-cutter, an open-source skill called Hallmark tries to break the templated 'AI-vibe' dilemma with rules and self-reflection.

  • Hallmark is a design skill for Claude Code, Cursor, and Codex, with the core goal of refusing to generate designs that 'look like AI made them'.
  • It includes 57 'slop-test' gates (anti-AI-vibe tests) and a pre-emit self-critique mechanism, forcing output to avoid common distributions in LLM training data.
  • Offers four verbs: default to build new UI, audit to review existing code, redesign to restructure, and study to extract design DNA.
  • Comes with 20 built-in themes, but allows switching to 'Custom' mode when creative intent cannot be matched, designing from scratch with no templates.
  • Each page is self-contained HTML+CSS, with its macro structure marked in CSS comments, ensuring different needs produce different forms.
  • The project is made by Together AI, MIT licensed, and has 18k stars on GitHub.
Open section navigationCore Mechanism: 57 Gates and Self-Critique

Core Mechanism: 57 Gates and Self-Critique

Hallmark's core is a rule set containing 57 'slop-test' gates (anti-AI-vibe tests) and a pre-emit self-critique. According to the project README, these tests are designed to reject common distribution defaults in LLM training data, thereby avoiding generating 'AI-vibe' designs.

The project claims that for two different needs, the two pages generated by Hallmark will feel like different websites, not the same template with swapped colors. This effect is achieved through a process of 'selecting a macro structure for the need, decorating with one of 20 themes, then passing through 57 gates and self-critique'.

Four Verbs: From Building to Auditing to Learning

Hallmark offers four verb modes. The default mode builds new UI, selects a macro structure, and applies the rule set; audit mode scores existing code for anti-patterns, outputting only a list of issues without modification; redesign mode discards the original structure, retains copy, information architecture, and brand, and rebuilds with a different fingerprint; study mode extracts design DNA (macro structure, font pairing, color anchors) from a screenshot or URL, refuses pixel-level cloning and paid templates, and optionally outputs a portable design.md file for use by other AI tools.

These verbs cover the complete design workflow from generation to auditing to learning, making Hallmark not just a generation tool but also a design analysis and refactoring tool.

Theme System and Custom Mode

Hallmark comes with 20 built-in themes, each corresponding to a different visual style. The project shows multiple examples, such as Bubble, Cold Snap, Ferns & Fathom, etc., each generating different page forms for different needs (e.g., a sourdough app, a record label, a tea menu).

When a need includes creative intent and no catalog theme can match it, Hallmark switches to Custom mode, designing the page from scratch: custom palette, fonts, and layout, also passing through the 57 slop-test gates, but with no template. The project emphasizes that Custom mode is a 'quiet branch' and ordinary needs will not trigger it.

Installation and Usage

Hallmark is installed via the command npx skills add nutlope/hallmark, which can be rerun to update. The rule set is located in SKILL.md and the references/ directory, with examples in docs/recipes.md and docs/study-examples.md.

It supports three tools: Claude Code, Cursor, and Codex, each with different configuration file paths. The project is MIT licensed, allowing free use, forking, and distribution.

Credibility boundary

This article's information primarily comes from the GitHub repository README, which is a self-declaration by the project team (Nutlope/Together AI). All functional descriptions and numbers (57 gates, 20 themes, 18k stars) are provided by the project team and have not been independently verified by third parties.

Insight takeaway

Hallmark attempts to solve the homogenization problem of AI-generated designs through rule sets and self-critique mechanisms, but its actual effectiveness and generality still need community practice to verify.

Primary report

Hacker News (AI filter)

Primary source