kill-ai-slop, tested

An agent skill that scans a web project for the visual and copy tics of AI-generated products — indigo gradients, glowing cards, "not just X — it's Y" copy, invented stat rows — and fixes them. We installed it and fed it a deliberately terrible page.

What it is

kill-ai-slop is two things sharing one taxonomy: a multilingual field guide at killaislop.com cataloguing 33 "tells" of vibe-coded design, and an agent skill that turns the catalogue into detection and fixes. The skill ships a SKILL.md, three reference files (detection, fixes, taxonomy), and — unusually — a real scanner script with its own tests, so detection is code, not vibes.

The install, as it actually happened

npx skills add yetone/kill-ai-slop -a claude-code -y

2.8 seconds, one skill copied into .claude/skills/kill-ai-slop. Cleanest install of our first batch. Manual fallback (copy skill/ yourself) is documented upstream and works too.

The test: scanning a page built to fail

We wrote a small page containing an indigo-to-violet gradient headline, a glassmorphism card, an emoji-led "Not just a tool — it's a revolution" hero line, and a fabricated "10k+ users · 99.9% uptime · 24/7 support" stat row, then ran the bundled scanner:

node .claude/skills/kill-ai-slop/scripts/scan.mjs .

Result, in 0.06 seconds: 7 slop groups, 8 hits — the gradient, the gradient-clipped headline, the AI copywriting voice, emoji-in-copy, glassmorphism/over-rounding, corner nesting, and the invented stat row. Every plant was caught; each hit came with a fix direction and a file:line reference. For an agent skill, this is the right shape: the scanner narrows the search, the agent (armed with references/fixes.md) does the judgment work.

What broke

Nothing broke. Install and the minimal use case both ran clean on first try — rarer than you would think, and worth saying explicitly.

Who it's for

Anyone shipping frontends with an agent in the loop. It is most valuable as a pre-commit habit: scan, review hits, fix with taste rather than mechanically — the SKILL.md itself insists detection is a lead, not a verdict. Pair it with a design-system generator (a tested comparison against Nutlope's hallmark is coming) and you cover both directions: generate clean, and catch regressions to slop.

FAQ

Does it work outside Claude Code?

Yes — the skill format is plain markdown plus scripts; the universal installer targets Cursor, Codex and others with the same command, and the scanner is a standalone Node script.

Does the scanner change my files?

No. It reports hits with locations; fixes are applied by you or your agent, per the references. That separation is deliberate and, in our view, correct.