Claude design skills, tested: hallmark vs kill-ai-slop vs visual-skills
"Design skills" for Claude actually means three different jobs: generating a design system, catching AI-sloppy output, and prompting image/video models. We installed the leading skill for each job and ran them, instead of ranking screenshots.
The three jobs, and what we tested for each
| Job | Skill | Stars | Install (measured) |
|---|---|---|---|
| Generate: give the agent taste up front | Nutlope/hallmark | 17,177 as of 2026-07-25 | 2.5s |
| Catch: detect & fix slop in existing code | yetone/kill-ai-slop | 738 as of 2026-07-25 | 2.8s |
| Prompt: brief image/video models properly | smixs/visual-skills | 108 as of 2026-07-25 | 2.5s |
Full test record for kill-ai-slop on its own page; hallmark and visual-skills records are summarized below from the same logs.
hallmark: the generator
Nutlope's anti-slop design skill approaches the problem from the front: it loads opinionated
design-system guidance into the agent before anything is generated, so the first draft
already avoids machine defaults. Instruction-only — SKILL.md plus references, no
runtime dependencies — which explains the clean 2.5s install.
We verified its installation and audit behavior in Claude Code; Cursor and Codex behavior were
not tested. At 17,177 stars (as of
2026-07-25) it is the most adopted design skill in this comparison.
kill-ai-slop: the reviewer
The opposite direction: a catalogue of 33 concrete slop tells, plus a real scanner script that reports file-and-line hits. In our test it caught all seven planted offenses in a deliberately terrible page in 0.06 seconds. Full test record on its dedicated page.
visual-skills: the art director
Two skills (video + image) that teach structured prompting for Nano Banana, GPT Image, Seedance, Kling and Veo — shot structure, emotion, multi-shot briefs. Both documented install paths work (the universal CLI, though the README only mentions manual copy). One honest note for English speakers: the README's usage examples are written in Russian; the skill bodies themselves are usable regardless, since your agent reads them fine. At 108 stars (as of 2026-07-25) it is the smallest project here, but it was pushed the day we tested.
The combination that actually works
These are not competitors — they are stages. Generate with hallmark's rules loaded, review with kill-ai-slop's scanner before shipping, and bring in visual-skills when the deliverable is a generated image or video rather than a UI. We run the first two in that order on this very site: the layout was built under hallmark-style constraints (one accent, no gradients) and then scanned with kill-ai-slop — the only hits were the slop examples we quote on purpose.
FAQ
Do any of these need a paid API?
No. All three are instruction-layer skills; visual-skills assumes you have access to whatever image/video model you are prompting, but the skill itself is free markdown.
Can hallmark and kill-ai-slop conflict?
They agree on the big rules (kill gradients, kill emoji spam, one accent). hallmark is generative preference, kill-ai-slop is detection; when both are installed the agent treats the scanner's output as review input, which is how we use them.