You can spot an AI-generated website at a single glance. It has a symmetrical hero section, a purple gradient, Inter font, rounded cards, and centered layout. Change the color, swap out the copy, and the core structure stays exactly the same. All large language models have been trained on certain "in-distribution defaults", which is why AI-designed interfaces around the world are getting increasingly homogeneous.

Hassan from Together AI built Hallmark, a design skill specifically made to combat this problem. It already has 25.8k stars on GitHub, and its official description is straightforward: A design skill for Claude Code, Cursor, and Codex that refuses to look AI-generated.


Its approach lists all common generic AI design tropes as "don't do this" rules to feed to the model. Its specific mechanism is:

- Select a macrostructure for each design brief instead of applying a one-size-fits-all template

- Wrap the design in one of 21 distinct themes

- Run 57 "slop-test" checks before generation, plus an additional pre-emission self-criticism step

- Explicitly reject all default distribution patterns that LLMs are trained on

How well does it work? Two pages generated from two different design briefs feel like entirely different websites, rather than the same template with a new coat of paint.

Hallmark has four command verbs:

- Default: Build a new UI, complete all slop tests before delivery

- `hallmark audit

- `hallmark redesign

- `hallmark study

Install with one line of command:

```bash

npx skills add nutlope/hallmark

```

Supports Claude Code, Cursor, and Codex. Licensed under MIT.

Hassan demoed this work at the AI Engineer conference:

Hallmark definitely works, but as Baoyu put it, skills alone don't make that much of a difference. What truly determines design quality is:

1. Human aesthetic judgment

2. Model capability (GPT doesn't do well at design, Claude is far stronger, and GLM-5.2 also outperforms GPT)

3. A personalized design system — a complete set of design specs and components, rather than just a list of "don't design this way" rules

LinearUncle added a more practical insight: Spending an extra 10-20% effort on UI is the biggest competitive advantage in the era of generic AI design slop. He summarized several key takeaways from Hassan:

- You have to understand common AI design tropes to help agents avoid them

- Most importantly: Feed agents reference images and screenshots

- Longer, more specific prompts are always better

- Iterate with cheap small models — GLM 5.2 delivers nearly indistinguishable results from Opus, and is much faster

Netizen HSIAO YUAN shared his own experience: When he used Claude to organize content, text-only descriptions often led it to the wrong direction, but adding a screenshot made it understand immediately. When he asked for a specific复盘 table style, it got it wrong three times with text descriptions, but got it right in one try with a reference image. This is a very common experience.


Another netizen called "Uncle Lai" mentioned that the management system he used to work on was always criticized by users for its ugly interface, and his team leader always blamed it on architecture issues and said it needed a full refactor with Vue. Now he can quickly modify and optimize the UI with this skill, which saves a lot of work, and all humans need to contribute is aesthetic judgment. "Professor An" also put it bluntly: Rule out bad designs first, and aesthetic rules are actually more effective than prompts.

But there is one question worth thinking about: When everyone starts using Hallmark, will generic AI slop just turn into generic "Hallmark style"? Or will the inherently subjective nature of aesthetic taste just get absorbed by iterative model improvements over time? No one has an answer right now. But at the very least, this tool has lowered the bar for creating "not ugly" interfaces, and everything else still comes down to human work.