anti-ai-slop

A portable Claude Code skill that stops Claude's UI and presentation output from defaulting to the recognizable "AI-generated" house style — the purple gradients, Inter-flat type, uniform rounded cards, 3-icon feature grids, emoji icons, kicker-on-every-slide, and benefit-speak copy that make generated work read as machine-made.

It loads automatically whenever you ask Claude to build or restyle any visual output — a website, landing page, web app, component, dashboard, or a slide deck / pitch deck — in any medium (HTML, React, Tailwind, slides).

Why

An LLM emits the most probable token. For an unconstrained visual choice, "most probable" is the statistical average of every tutorial it was trained on — so with no constraint it deterministically produces the same purple-gradient, Inter-font, rounded-card page every time. This skill forces a specific, intentional decision wherever the model would otherwise reach for that average, and it names the second-order traps too (the "tasteful AI" monoculture of warm cream paper + a serif display + a terracotta accent that the first fix tends to converge on).

Install

Per-project — drop it into a project's skills directory:

git clone https://github.com/Vinayak-Shukla-03/anti-ai-slop.git
mkdir -p your-project/.claude/skills
cp -R anti-ai-slop your-project/.claude/skills/anti-ai-slop

Global — make it available in every project:

git clone https://github.com/Vinayak-Shukla-03/anti-ai-slop.git ~/.claude/skills/anti-ai-slop

That's it. The skill's description triggers it on any UI/deck request; no configuration needed. (Keep the references/ folder next to SKILL.md — the skill links to it by relative path.)

What's inside

File	Role
SKILL.md	The core: 9 hard rules, a method (brief → distinctive foundation → real content → audit), a pre-ship self-audit, and a render-it-and-look ship-blocker.
references/web-ui.md	App / dashboard / landing-page tells + fixes, plus a reliability section (a blank page can't be un-slopped).
references/presentations.md	Deck / slide tells + fixes (kicker-every-slide, bullet-rhythm, thank-you-slide, the "deck chrome" kit).
references/toolkit.md	Concrete swap-ins: palettes, font pairings, scales — and a divergence matrix so a set of artifacts doesn't re-converge.

The nine hard rules (the short version)

1. No unbranded purple/indigo and no reflexive gradient.
2. Never the system font or Inter/Geist flat — a real display + body pairing.
3. No uniform-everything — a radius/padding hierarchy; size and space signal priority.
4. No emoji as icons — a real, consistent icon set.
5. No transition: all and no blanket fade-in — purposeful motion only.
6. No benefit-speak ("Elevate", "Unlock", "Seamlessly", "not X — it's Y").
7. No fabricated precision — real numbers with a source, or none.
8. Break the template skeleton — content and audience drive structure.
9. No second-order monoculture — if you make a set, force it to diverge (ground, display face, accent, stance).

Plus a ship-blocker: render it and look before calling any UI done — beautiful code that doesn't mount is worth zero.

How it was built

Empirically, not from vibes. A baseline corpus of 10 artifacts (5 apps + 5 decks) was generated with no design guidance to capture Claude's unprompted defaults, dissected for recurring tells, cross-checked against how designers and dev communities describe "AI slop," distilled into these rules, then used to rebuild the same 10 — and the rebuilds were rendered and scored. A second pass caught a second-order monoculture (the fix's own overused escape hatch) and hardened the skill with the anti-monoculture + divergence rules it now carries.

License

MIT — see LICENSE.