← Back to blog
AiAbout 6 min read

A Claude skill that audits your interface for slop, then fixes it in nine passes

Published Nov 21, 2026
A Claude skill that audits your interface for slop, then fixes it in nine passes

The repo went up on 29 October 2026 and had 19,413 stars by 21 November. It is called unslop, by Damir Sokolovsky, and it does one thing: it takes a screen that an AI built and makes it look like a person designed it.

The problem it targets is easy to recognise and hard to name. You ask a model for a dashboard. You get a layout with a violet-to-indigo gradient on the hero, a glow behind the primary button, three emoji in the section labels, copy that shouts ("Supercharge your workflow!"), padding that is 24px in one card and 31px in the next, and a card grid that is almost aligned. Every element is defensible. The screen as a whole reads as machine output.

unslop starts from a list of those tells and works outward from there.

Install, and what you get

For Claude Code, one command puts it in every project:

claude plugin install unslop --marketplace heydamir/unslop

Older builds without --marketplace need two steps: add the marketplace, then install unslop@heydamir. Add --scope project to install it for everyone who works in the repo, and commit .claude/settings.json. Then start a new session, or run /reload-plugins in one that is already open.

Codex users install from the repo directly:

$skill-installer install https://github.com/heydamir/unslop/tree/main/skills/unslop

Or clone it and copy skills/unslop into ~/.agents/skills/, which also works for a single project if you drop it in .agents/skills/. Antigravity takes the same folder at ~/.gemini/config/skills/. Nothing here is exotic, which is the point: a skill is a folder of instructions, and this one wants to be one keystroke away.

Two optional dependencies matter. Snapshotting a live site or rendering the story video needs Playwright and its Chromium build (npm install -D playwright, npx playwright install chromium), and the video also needs ffmpeg on PATH. If you skip both, you still get the audit and the edits.

The audit comes first

You type /unslop this dashboard, or /unslop the hero of https://example.com, or hand it a live URL when the only thing you have is a link. The skill works in whatever the screen is built with: Next.js, React, Vue, Svelte, Tailwind, plain HTML and CSS.

Before it changes a pixel, it lists every sign of slop and where it lives. That ordering is deliberate. A design pass that starts with edits has already committed to a direction; an audit that names the problems can be argued with. If the screen is fine, it says so and suggests a few small refinements instead of redesigning something that works.

Then it proposes a direction drawn from your product: character, world, the one or two moments that carry the colour, and a signature detail. This is the step most AI design tools skip, and it is the reason their output converges. A gradient is not a direction. "The only saturated colour in the whole product appears on the number that matters" is a direction.

Nine passes, in a fixed order

The edits run in this sequence: spacing, colour, icons, copy, type, layout, details, visuals, motion.

The order is doing work. Spacing first because an uneven rhythm makes every later decision harder to judge. Colour second because it is the cheapest way to make a screen feel intentional. Copy sits in the middle and usually changes the most: an audit tends to find labels written for a demo ("Unleash insights"), and the fix is to name the thing. Motion last because it is the pass most likely to be overdone, and doing it after layout means you are animating a structure that has already settled.

Each pass is small enough to review. That matters when the thing being edited is a real product rather than a mockup.

The part with a budget

Visuals get a rule that most design agents do not have: show the product instead of a glossy blob. When a hero genuinely needs a scene or an object, the skill can generate one through fal, Replicate, OpenAI images or Higgsfield, and it shows the prompt and asks before every paid generation. If none of those services is connected, it offers to wire one in as an MCP server.

Dashboards and apps rarely need a generated image, and the skill says so. That single sentence prevents a common failure mode, where an agent fills empty space with a synthetic photograph because a hero section "should" have one.

There is also a story video: a square clip under ten seconds of your screen unslopping itself, ready to post. It needs Playwright and ffmpeg, and it exists because the before-and-after is the argument. You can watch the violet gradient drain out of the hero in real time.

What it cannot do

Three honest limits.

It is instructions, not a linter. Nothing enforces the passes. A model can declare a screen fixed while the 31px padding is still there, and the only defence is reading the diff. Skills that touch design inherit the same weakness as skills that touch code: the model grades its own homework.

Taste is borrowed. The direction comes from your product, but the vocabulary of what counts as slop is one designer's. Purple gradients, glows, emoji labels, shouty copy and uneven spacing are a decent list and also a list of things that are sometimes right. A brand built on violet gradients will get an audit telling it to stop.

The author sells the manual version. Damir unslops one key screen in three days, in your stack, for a fee. That is a fair business and also a signal about where the ceiling is.

Why this one spread

Design skills are not new. There are skill repos for landing pages, for architecture diagrams, for pixel art. What unslop got right is the diagnosis. "AI-looking" is not a mystery, it is a short list of defaults that models reach for when they have no direction: the gradient, the glow, the emoji, the exclamation mark, the 8px misalignment that nobody fixed because nobody looked.

Naming the list turns a vague complaint into a checklist, and a checklist is something an agent can execute. The nine passes then give it an order, which is the difference between a redesign and a review.

If you have a vibecoded screen sitting in a repo, the cheapest test is the audit. Run it, read the list, and decide whether you agree with the diagnosis. That is worth more than the edits, because the diagnosis is the part you will reuse on the next screen.

The repo is CC BY 4.0: use it, change it, ship it commercially, credit the author. Two example screens are included, both vibecoded, both deliberately ugly, and both fixed by the same nine passes.

Related articles