Disclaimer:This site is not affiliated with or endorsed by unslop.
YuliusWiki
Published:•Last updated:•yulius
🌐 中文版 (Chinese)

What Is unslop? A De-Slop Skill for Claude Code and Codex: 5,700 Stars in 4 Days, Audit First, Then Nine Passes

Quick Answer

A free CC BY 4.0 skill for Claude Code and Codex that audits AI-looking UI, sets a direction, then fixes it pass by pass.

Vibecoded interfaces are easy to spot: purple gradients, glows everywhere, emoji as labels, shouty copy, uneven spacing. Author Damir Sokolovsky calls that look 'slop', and unslop is a skill built to remove it: it runs inside Claude Code and Codex and turns AI-looking UI into a designed interface with its own character. The repo was created on October 29, 2026 and had about 5,700 stars when checked on November 2, 2026 — roughly 1,430 a day over 4 days — under a CC BY 4.0 license. Three clarifications up front: first, it is not a model or a standalone app, but a set of instructions and a workflow — the agent and model underneath do the actual work; second, it changes interfaces only, not product logic or backends; third, star counts signal heat, not measured quality, and the repo publishes no benchmark numbers, so this article invents no effectiveness data.\n\nDisclaimer: this site is not affiliated with the author, Anthropic, or OpenAI. This is product-technology reporting and analysis, not purchasing, deployment, or legal advice. Claude Code and Codex are their respective products' names. Specs and license terms follow the official repo page and may change; anything unpublished should be checked against the official page.

1. What it is: a de-slop skill for Claude Code and Codex

unslop's author is Damir Sokolovsky (@heydamir), who first ran an Unslopping series on X and then turned that practice into a skill. The repo's one-line pitch says it all: it turns vibecoded, AI-looking UI into a designed interface. Note the two qualifiers in that sentence: the object is the interface, not the product itself; and the goal is not 'a prettier template' but 'a design with its own character'.\n\nIts scope is wide: dashboards, apps, and landing pages; stacks are not picky — Next.js, React, Vue, Svelte, Tailwind, or plain HTML and CSS; and it even works on a live site when all you have is a link. That is exactly the pain point it hits: vibe coding made 'producing an interface' cheap, and the bottleneck moved to 'producing an interface that does not look AI-made'.\n\nThe repo was created on October 29, 2026 and had about 5,700 stars when checked on November 2, 2026 — roughly 1,430 a day over 4 days. That number shows broad resonance, but one caution: stars are a heat signal, not a quality measurement, and the repo publishes no benchmark numbers, so do not read them as proof of results.

2. How it works: audit first, direction second, nine passes in order

unslop's workflow has three stages, each with a clear output. Stage one is the audit: find the slop before touching anything, listing every sign of slop and where it lives — purple gradients, glows, emoji labels, shouty copy, uneven spacing are the typical symptoms named in the README. Without this step, 'prettifying' easily becomes swapping one AI look for another.\n\nStage two sets the direction: before any edit, it proposes a direction — character, world, the one or two moments that carry the color, and a signature detail — all taken from your product, not copied from a template library. The direction comes first so you can steer early, instead of discovering the drift after the edits are done. Stage three is passes in a fixed order: spacing, color, icons, copy, type, layout, details, visuals, and motion — like a designer doing a redline review, one class of problem per pass.\n\nTwo design choices are worth noting. First, if the screen is already good, it says so and suggests a few small refinements instead of redesigning for its own sake. Second, visuals show your product rather than a glossy blob; when a hero needs a scene or an object, it can generate one through an image service like fal, Replicate, OpenAI images, or Higgsfield — if none is connected, it offers to connect one as an MCP server — and it shows the prompt and asks before every paid generation. Dashboards and apps rarely need one, so this step usually never triggers.

3. Install and use: two copy commands, one sentence to start

Installation has two tracks. For Claude Code: clone the repo and copy skills/unslop into ~/.claude/skills; for a single project, copy it into .claude/skills inside that project instead. For Codex: ask it to install the skill straight from the repo ($skill-installer install with the repo skills path), or copy it by hand into ~/.agents/skills, or .agents/skills inside a single project.\n\nUse is a single sentence. In Claude Code type /unslop, in Codex type $unslop, followed by what you want; or just ask in your own words and the agent picks the skill up by itself. The README's examples are blunt: /unslop this dashboard, /unslop the hero of a URL, or add a line asking for the story video and get it done together.\n\nThe story video is a very postable byproduct: a square video under 10 seconds of your screen unslopping itself. The price is environment setup: install Playwright and its Chromium browser, and ffmpeg must be installed and on your PATH for video. The repo also ships two vibecoded example screens — if you do not want to touch your own project first, try it on those.

4. Who it fits, and where it stops: what it does not do

Three groups fit best. First, developers and indie makers whose vibe-coded first version looks, even to themselves, too AI-made. Second, people who want one consistent character for their product rather than another template. Third, people who want to post the redesign process — the story video is built for exactly that.\n\nThe boundaries should be stated up front. It does not change product logic or write backends; it only touches interfaces. It needs a screen to start from — code or a link — it cannot conjure a product from nothing. Image generation depends on third-party services and may cost money, though it asks before any paid run. When a screen is already good, it will not force a redesign. And the final result depends on the agent and model underneath: it is not a magic button, and you still need to review the result yourself.\n\nOn cost and license, the skill itself is free: CC BY 4.0, including commercial use, on the condition that you credit author Damir Sokolovsky and link back to the repo. The author also offers a paid service — one key screen in 3 days (heydamir.design/unslop) — which is a separate human service, not to be confused with the free skill.

5. Why it blew up, and sources

Why did this one blow up, rather than yet another 'AI prettify' prompt? One explanation: it writes a designer's workflow into a form an agent can execute. Audit, direction, passes in order — those three steps are how designers revise drafts in the first place. Fixed in place, the agent stops editing everything at once like a headless fly, and stops equating 'good-looking' with 'a new template'. Add the spread of vibe coding — everyone can generate an interface in one sentence, and what is missing is exactly the 'de-slop' step — and broad resonance follows naturally.\n\nBut be honest about what is verified: heat (about 5,700 stars in 4 days) and mechanism (the workflow written in the README). There are no third-party measurements and no before/after comparison data. The best test of whether it is worth using is the repo's own two example screens, or one run of /unslop on your own page — check whether the problems its audit lists actually hit the mark.\n\nSources (verified November 2, 2026, Beijing time):\n1) https://github.com/heydamir/unslop — official repo README: features, install, use, nine-pass order, image generation and story video, CC BY 4.0 license, author's paid service;\n2) GitHub API — repo created 2026-10-29T15:41:21Z, 5,707 stars at check time, about 1,430 a day (recomputed: 5,707 / about 3.98 days ≈ 1,434).\n\nDisclaimer: this site is not affiliated with the author, Anthropic, or OpenAI. This is product-technology reporting and analysis, not purchasing, deployment, or legal advice. Specs and license terms follow the official repo page and may change; anything unpublished should be checked against the official page.

unslop vs. manual redesign vs. asking AI to 'make it prettier': three differences

DimensionunslopDesigner manual passAsking AI to 'make it prettier'
Before editingAudits first: lists every sign of slop and where it livesRelies on the designer's eye, item by itemNo audit, edits straight away
Edit orderFixed nine passes: spacing, color, icons, copy, type, layout, details, visuals, motionOrder by experienceEverything at once
Direction sourceCharacter and signature details taken from your productSet by the designerA style guessed by the model
When the screen is already goodSays so plainly, offers small refinementsSameTends to redesign anyway, swapping in a new template
CostFree (CC BY 4.0); third-party image generation pay-as-you-go, asked before each paid runHuman labor costToken cost

Frequently Asked Questions

Is unslop a model?

No. It is a skill — a set of instructions and a workflow — that runs inside Claude Code or Codex. The agent and model underneath do the actual work; unslop defines the order and the standard for the edits.

Do I have to install the skill? Can't I just ask AI to make it prettier?

You can, but the difference is process: unslop audits first, sets a direction, then edits in nine ordered passes. A bare 'make it prettier' usually skips the audit and edits everything at once, which often swaps one AI-looking template for another. If you already get good results reliably, there is no need to switch; if the result still smells AI-made, it is worth one run to see whether its audit names the real problems.

Does image generation cost money?

The skill itself is free (CC BY 4.0). Only when a hero needs a generated scene or object and you have connected a third-party service like fal, Replicate, OpenAI images, or Higgsfield can a fee arise — and the prompt is shown and approval asked before every paid generation. Dashboards and apps rarely need image generation.

Can I use the result commercially? Is attribution required?

Yes, commercial use is allowed. The CC BY 4.0 license requires crediting author Damir Sokolovsky and linking back to the heydamir/unslop repo.

© 2026 YuliusWiki. All rights reserved.