unslop-ui

Read this first, because the most common misunderstanding sinks the whole thing.

This skill does not give you a good design, and it does not have a preferred look. It does two narrow things. It removes the specific cues that make a site read as AI-generated, and it forces a deliberate, project-specific choice where the model would otherwise reach for its default. Taste, brand, and layout judgment are still yours. A guardrail is not a designer.

The trap to avoid (this is the whole point)

The failure mode of every anti-slop effort is replacing one default look with another. The 2024 tell was the purple-to-blue gradient on a dark hero. The 2026 tell is a warm cream background with a serif display font (Instrument Serif, Fraunces) and a sage or forest green accent, which is the current Claude/Anthropic house style. Swapping the first for the second is not unslopping. It just resets the clock, and people clock it just as fast.

So this skill never prescribes a palette, a font, or a layout. Prescribing one is how you become next year's slop. What it does instead: detect the defaults (old AND new), and require that any replacement be a specific choice the user or project actually made, not the model's next-most-likely guess. The only universal rule here is "make a deliberate choice and be able to say why," which is the one thing a default never is.

If the user genuinely wants purple, neon, a serif, or a cream background as a real brand decision, that is not slop and the skill leaves it alone. A tell is an unspecified default, not a banned color. Honor unslop-ignore (see Audit mode) for anything chosen on purpose.

How this differs from the built-in frontend-design skill

Claude already ships a frontend-design skill, and it is good, but it produces the cream

• serif + green house look that this skill now flags as a tell. unslop-ui is complementary and adds three things that one does not: a deterministic scanner you can run in CI (exit code gates a build), a data-grounded ranking so effort goes where real complaints are, and an explicit check against the new "tasteful default" so you do not launder slop into a different slop. Use both. Let the design skill build; let this one keep it honest.

Mode 1: Build (the important one)

Most "looks AI" outcomes are a specification problem, not a styling problem. An unspecified prompt gets the median of the training data, and everyone's median is identical. So before generating any UI, establish the brief. Either pull it from the user, or if they have not given one, state the choices you are making and why, then proceed. Do not silently fall back to defaults.

Establish, concretely:

• A reference. One real site, screenshot, brand, or product whose design language to follow. This single input does more than every other rule combined. If the user has one, anchor to it. If not, ask for one, or pick a real one from the collection indexed in references/picking-a-reference.md — a specific named direction (editorial, brutalist, utilitarian-dense, warm-consumer, technical-mono, and so on) is the last resort, and "modern and clean" means nothing.
• A color decision. A real brand color or a deliberately chosen one, stated. Not the framework default, and not the cream/sage default either.
• A type decision. A specific typeface or pairing chosen for this project, with a reason. Avoid the top-50 defaults reached for on autopilot (Inter, Geist on the sans side; Instrument Serif, Fraunces, Playfair on the "tasteful" side) unless they are a real choice.
• A layout intent. What the page is actually for and what the user should do first, which determines structure. This is how you avoid the hero + three-card skeleton: the structure follows the goal, not a template.

When the user gives no brief and wants you to just build, do not produce one median result. Produce a deliberate one and say what you chose, or offer two or three genuinely distinct directions. The value is breaking the monoculture, so vary away from the center on purpose.

Then, while building, avoid the specific tells in references/tells.md. references/choosing-a-look.md is a process for making the color/type/layout choices deliberately, written as a method rather than a prescription, precisely so it does not become the next default. When the user cannot name a reference, references/picking-a-reference.md is how to pick a real one from an indexed collection of captured-site guides instead of settling for a genre adjective.

Mode 2: Audit (the guardrail)

When reviewing or cleaning existing code, or when the user says "does this look AI," "de-slop this," "make it not look vibe-coded," run the scanner first, then fix in priority order.

python3 scripts/devibe_scan.py <path>                 # full report + vibe score
python3 scripts/devibe_scan.py <path> --severity high # only the strongest signals
python3 scripts/devibe_scan.py <path> --json          # machine-readable, for CI

It scans .html .css .scss .js .jsx .ts .tsx .vue .svelte .astro, reports each finding with file, line, and fix, and gives a vibe score. Exit code is the high-severity count, so CI can gate on it. The scanner catches the mechanical tells (colors, fonts, gradients, the cream+serif combo). It cannot see layout coherence, spacing consistency, or whether text overflows its container, and those are also what make a site read as AI, so after the scan, check those by eye against the catalog.

Respecting intentional choices. A line with the comment unslop-ignore is skipped. Use it when a flagged value is a real decision, so the audit stays trustworthy and does not nag about a chosen brand color.

Fixing well. Do not fix bg-purple-600 by swapping in bg-emerald-700, which is just a different default. Fix it by applying the project's actual color, or by asking what the color should be. A fix that introduces a new unspecified default is not a fix.

What this deliberately does not flag

Grounded in the data, not vibes. Mesh, aurora, and blob backgrounds barely register as real complaints (mostly a keyword artifact). Bento grids and glassmorphism are low and contested. Dark mode itself is fine; only unprompted glow is a tell. shadcn and Tailwind themselves are fine; their untouched defaults are the tell. Over-flagging trains people to ignore the tool, so the scanner stays narrow.

Reporting an audit

Lead with the verdict and the single highest-impact change. Then findings by priority with file:line and the fix. Close with the vibe score and the top three changes. Plain and specific. The goal is a site that looks like a person made a decision, which is the one thing the scanner cannot do for them.