AI-Built Interfaces Have a List of Tells Now. Someone Wrote a Skill to Remove Them.

Damir Sokolovsky's unslop reached 11,564 stars in ten days, which is fast for a repository that contains no model, no weights and no new architecture. It contains a checklist.
The skill runs inside Claude Code or Codex. Point it at a screen and it looks for what the README calls slop: purple gradients, glows, emoji labels, shouty copy, uneven spacing. Then it proposes a design direction taken from the product itself, and fixes the screen pass by pass, the way a designer does a redline review.
That is the entire pitch, and the fact that it works says something about the last eighteen months of AI-generated interfaces. The failure mode has become consistent enough to enumerate.
The tells, in the order the skill looks for them
The audit comes first and it is deliberately concrete. A list of symptoms with locations, not an opinion about the vibe. Anyone who has shipped with a coding agent in the last year will recognise most of them, and that recognition is the reason the repository spread.
Purple gradients arrived first because they were the default in early component libraries, then stuck because generated code reuses what it has seen. Glows and glassmorphism came in a wave behind them. Emoji as section labels came from the habit of making plain text feel friendly without designing an icon set. Shouty copy is the all-caps imperative that appears when a model writes marketing text and no editor reads it. Uneven spacing is the one that survives review, because it looks fine in a screenshot and wrong in a browser.
None of these are quality problems in isolation. Together they are a signature, and it reads as "nobody made a decision here."
What the skill does with the list is more interesting than the list. It proposes a direction before editing anything: character, world, the one or two moments that carry the color, a signature detail, all sourced from your product rather than from a mood board. If the screen is already good, it says so and suggests small refinements instead of redesigning.
Passes, in a fixed order
Then it edits, in this sequence: spacing, color, icons, copy, type, layout, details, visuals, motion.
The order is the design argument. Spacing before color, because a layout with the wrong rhythm cannot be rescued by a better palette. Color before icons, because icons inherit weight and contrast from the type scale above them. Copy before type, because text length decides the type. Layout after all four, because the grid is a consequence. Details and motion last, which is where most generated interfaces start.
Anyone who has watched a model "improve" a screen by reaching for the gradient first recognises the inversion immediately.
Install, if you want to try it
For Claude Code:
git clone https://github.com/heydamir/unslop
mkdir -p ~/.claude/skills && cp -r unslop/skills/unslop ~/.claude/skills/
For a single project, copy it into `.claude/skills/` inside that project instead.
For Codex, ask it to install from the repository, or copy the skill into `~/.agents/skills/` by hand:
$skill-installer install https://github.com/heydamir/unslop/tree/main/skills/unslop
Snapshotting a live site, or producing a story video, needs Playwright and Chromium installed in the directory where you run the skill:
npm install -D playwright
npx playwright install chromium
The story video also needs ffmpeg on PATH.
Usage is one line, in the skill's own syntax: `/unslop this dashboard`, `/unslop the hero of https://example.com`, or `/unslop my landing page and make the story video so I can post it`. In Codex it is `$unslop` with the same arguments.
It works on whatever the screen is built with, which is the part that makes it usable rather than a demo: Next.js, React, Vue, Svelte, Tailwind, or plain HTML and CSS. Dashboards, apps, landing pages. It will also work on a live site when all you have is a link, which is a different job from editing source and the README treats it as one.
The paid-generation guardrail
The visuals pass is where a design skill usually loses the plot. A hero needs a scene or an object, the agent generates something glossy, and the page ends up with a beautiful nothing in the middle of it.
unslop's answer has two parts. First, show the product instead of a glossy blob, which is the correct instinct and costs nothing. Second, when a scene genuinely is needed, it can generate one through an image service: fal, Replicate, OpenAI images or Higgsfield. If none is connected, it offers to connect one as an MCP server. Before every paid generation it shows the prompt and asks.
Two details in that design are worth copying into other skills. The gating is on cost, not on capability, so the agent never surprises you with a bill. And the default for dashboards and apps is that no image is needed at all, which is stated explicitly so the agent does not go shopping for one.
The author sells the same service by hand
At the bottom of the README there is a link to heydamir.design/unslop, where the author will unslop one key screen in three days, built in your stack.
That is the most honest part of the project. A skill that removes slop in minutes, from an author who charges for three days of the same work, is telling you where the boundary sits. The skill handles the tells. The three days handle the decisions the tells were standing in for.
It also explains the license. CC BY 4.0, free for commercial use with credit and a link back. That is a permissive choice for a repository that could have gone MIT, and it is the right one for something whose value is partly attribution.
What this says about the interface layer
Two things happened in parallel over the past year. Coding agents got good enough to produce a working screen from a sentence. Design tooling got good enough to produce a styled screen from a sentence. What did not happen is a shared standard for what the result should look like, so the defaults converged on whatever was most common in the training data.
The convergence is now measurable enough that some people wrote a converter. Two repos could be called a trend; the broader signal is that "does this look like AI made it" has become a question with a checklist answer, in the same way "does this read like AI wrote it" did for prose a year earlier.
There is a limit worth naming. Removing the tells gets you to competent. The skill's own audit language points at it: character, world, the moment that carries the color, a signature detail. Those are the parts that require someone to decide what the product is for, and no pass order supplies that. The README's three-day option is there because the author knows it.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.