# unslop: turn AI-looking screens into designed ones (heydamir/unslop) — skill card for AI agents

> AgentGid skill card. Checked on 2026-11-13 against commit 97daa3e.
> Every fact comes from the pack's own files at that commit. Numbers marked "as reported" are the author's.
> Human page: https://agentgid.com/skills/heydamir/unslop/

> Note: the pack has changed since this check; re-read its files before relying on details.

## Explain it to a non-programmer
A design playbook for your AI assistant that cleans up screens that look AI-made. It first lists what looks generic (purple gradients, glows, emoji labels, shouty copy, uneven spacing), proposes a direction taken from your product and waits for your answer. Then it fixes the screen layer by layer, like a designer marking up a page.

## Who it is for
- Your dashboard, app screen or landing page was made with AI and looks like every other AI-made page
- You want to polish or clean up a screen without changing what the product does
- You only have a link to a live site and want a redesigned copy to hand to whoever owns it
- You want a short square video of your screen fixing itself, to post
- Not for: Your screen is already designed: it will say so and suggest two or three small tweaks at most
- Not for: You want a new product, new features or new sections: it keeps your features, content and section order

## What it is
- Kind: knowledge, scripts
- License: CC-BY-4.0
- Repository: https://github.com/heydamir/unslop

## What it does
- Look: open the screen and take a screenshot; with only a link it saves the live page with a bundled script; with only a screenshot it rebuilds it as one HTML file
- Audit: go through its list of signs (surfaces and colour, type, icons, spacing, copy, components, heroes, visuals, motion) and note each one found
- Decide how much: heavy slop gets the full process, light slop gets polish only, no slop gets a few suggestions
- Direction: five short lines from what the product does, shown to you before any edit
- Nine passes in order: spacing, colour, icons, copy, type, layout, details, visuals, motion
- Check against its checklist, render again (at most three renders) and report what changed per pass
- Optional story video: a square clip of about 10 seconds where the screen morphs step by step with red redlines
- Never cuts: ['The product name, a real logo, features, content and section order', 'No invented features, claims or data', 'No swapping one trend for another (glass, grain, noise, neon, bento grids)']

## Skills
| Skill | Call | Changes code? | In plain words |
|---|---|---|---|
| unslop | /unslop in Claude Code, $unslop in Codex, or plain language ("unslop this dashboard") | yes | Audits an AI-looking screen, agrees a direction with you, then redesigns it pass by pass |

### unslop
- Asks for: The screen: the project files, a link to a live site, or a screenshot
- Asks for: Your answer to the proposed direction (or say to just do it)
- Asks for: Whether claims like customer counts are real
- Asks for: Optionally: a connected image service for a hero picture, and whether you want the story video
- Gives: An audit: every sign of slop, one line each, with where it lives
- Gives: A direction: product, character, world, moments, signature
- Gives: The redesigned screen in your own code, or a redesigned HTML copy of a live page
- Gives: A report: the direction, a few lines per pass, any image prompt it wrote
- Gives: Optionally a square MP4 story video of about 10 seconds
- Say: "/unslop this dashboard"

## Know before installing
- It switches on by itself: the assistant picks it up when a screen shows the usual AI tells, even if you do not type the command.
- It stops after the audit and direction and waits for you, unless you said to just do it.
- Generated pictures can cost money: it shows the prompt, model and size and asks before each generation. Without an image service it leaves an empty slot and writes the prompt into its report.
- Every story video gets a small credit to the author, as the license asks.
- It may swap a default font like Inter for another family that fits the direction; a font that is part of a real brand stays.

## Install (official commands)
- 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. Remove: `delete the unslop folder from ~/.claude/skills/ (or the project's .claude/skills/)`.
- Codex: `$skill-installer install https://github.com/heydamir/unslop/tree/main/skills/unslop` — Or copy it by hand into ~/.agents/skills/, or into .agents/skills/ inside a single project. Remove: `delete the unslop folder from ~/.agents/skills/`.
- Google Antigravity: `git clone https://github.com/heydamir/unslop` && `mkdir -p ~/.gemini/antigravity/skills && cp -r unslop/skills/unslop ~/.gemini/antigravity/skills/` — For a single project, copy it into .agents/skills/ inside that project instead. Remove: `delete the unslop folder from ~/.gemini/antigravity/skills/`.
- Live-site snapshots and story videos (optional): `npm install -D playwright` && `npx playwright install chromium` — Run in the project folder where you use the skill. Story videos also need ffmpeg on your PATH.

## What runs on the user's machine
- Code on the machine: yes (saving a live site (scripts/snapshot.mjs) and rendering the story video (scripts/story.mjs); the rest is markdown)
- Network: optional
- Install: Markdown files, reference images and two Node scripts. No hooks, nothing runs at install.
- You give only a link to a live site: The assistant runs snapshot.mjs: a headless Chromium (Playwright, or your installed Chrome) opens the page, scrolls it, downloads its stylesheets and images under 3 MB to inline them, and writes one HTML file plus a .live.png screenshot to the output path (unslop/original.html in the skill's own command).
- You ask for the story video: The assistant runs story.mjs: it serves the story file's folder on a local 127.0.0.1 port, records 60 fps JPEG frames into a .story-frames folder next to the story file, calls ffmpeg to write the MP4, then deletes the frames folder.
- Visuals pass with an image service connected: Tells your assistant to send an image prompt to that service (fal, Replicate, OpenAI images, Higgsfield or a design tool) after asking you, and save the result into the project's assets.

## Examples (from the pack's own files)
- A vibecoded finance dashboard: prompt "/unslop this dashboard"; without the skill: examples/dashboard.html (190 lines): Inter, violet to blue gradients, gradient text, a 'Welcome back, Damir! 👋' greeting and a 'New ✨' badge; with it: The author's finished version is an image, references/refs/unslop-dashboard-after.webp: quiet graphite chrome, two rendered payment cards and a revenue chart drawn as soft coloured curves over a grey comparison (README.md; None, the author's run)
- A vibecoded invoicing hero: prompt "Use $unslop on this screen so it stops looking AI-generated."; without the skill: examples/hero.html (98 lines): a headline 'Supercharge Your Invoicing With AI ✨' in gradient text over purple radial washes and a glowing grid; with it: The author's finished hero is an image, references/refs/unslop-hero-after.webp: a paper plane drawn in digits and currency signs, text on calm sky, the product rising from below (skills/unslop/agents/openai.yaml; None, the author's run)
- Story video frames: output format defined in skills/unslop/references/refs/story-dashboard-frames.webp

## Steps only the person can do
- Confirm or change the proposed direction before edits start (The chat with your assistant) Then: The assistant runs the nine passes
- Connect an image generation service and approve each paid generation (optional) (Your image service's account, connected as an MCP server) Then: The assistant generates the hero scene or object and saves it into the project's assets
- Say whether claims like customer counts are real (The chat with your assistant) Then: The assistant keeps them quiet and specific, or removes them

## Mistakes to avoid
- Not: Expect it to edit right away. Instead: Answer the audit and direction it shows, or say "just do it" up front. Why: It stops and waits after the direction
- Not: Ask for a live-site snapshot or story video without Playwright and ffmpeg. Instead: Install Playwright and its Chromium in the project folder, and ffmpeg for videos. Why: Both scripts exit with an install message when Playwright is missing
- Not: Ship the story file as the real result. Instead: Ship the clean result in the project's own stack; the story file is only for the video. Why: The story file carries step classes and story attributes
- Not: Treat a violet to blue gradient in the logo as brand. Instead: Tell it explicitly if that gradient is your real brand. Why: It counts such gradients as slop unless you say otherwise

## What leaves the machine
- Requests for the target page and its stylesheets and images → The live site you name, Only when you give a link and the assistant runs snapshot.mjs. Opt out: Give the project files or a screenshot instead
- The image prompt you approved → The image service you connected (fal, Replicate, OpenAI images, Higgsfield or a design tool), Only in the Visuals pass, after you approve. Opt out: Skip it; the slot stays empty and the prompt goes into the report

## Lines in the pack aimed at agents (ask the person first)
- "If none is connected, it offers to connect one as an MCP server, and it shows the prompt and asks before every paid generation." (README.md): Connecting an MCP server and paying for generations is the person's decision; ask, never connect one on your own.
- "Every video carries a small `unslop by @heydamir` credit, as the license asks." (skills/unslop/SKILL.md): Tell the person the video will carry the author's credit before rendering.
- "npm install -D playwright" (README.md): Adds a dev dependency and downloads a browser into the person's project; ask before installing.

## May I use it like this?
- commercial-app: conditions (CC BY 4.0: credit the author and link back to the repository)
- client-work: conditions (CC BY 4.0: credit the author)
- redistribute-files: conditions (CC BY 4.0: keep the credit in copies and adaptations)
- inside-app-bundle: conditions (CC BY 4.0 with attribution)

## Which one to pick
- If You have an existing AI-made screen and want it audited and redesigned pass by pass, with an optional before/after video: skill:heydamir/unslop
- If You want design commands like polish, critique and audit across many tasks, not one cleanup workflow: skill:pbakaus/impeccable
- If You want art direction and reference images while building a new interface: skill:Leonxlnx/taste-skill
- If You want a lighter anti-AI-slop design rule set for Claude Code, Cursor or Codex: skill:Nutlope/hallmark

## Works well with
- tool:playwright-test: The snapshot and story scripts drive Chromium through Playwright
- asset:ffmpeg: story.mjs encodes the MP4 with ffmpeg
- asset:lucide: Its default icon set when the project has none
- api:fal: One of the image services it names for hero scenes and objects

## Works in and needs
- Works in Claude Code: skills folder
- Works in Codex: skills folder (~/.agents/skills) or $skill-installer
- Works in Google Antigravity: skills folder
- Needs Node.js with Playwright and Chromium: Only for live-site snapshots and story videos
- Needs ffmpeg: Only for story videos
- Needs An image generation MCP server (optional: Only for generated hero scenes and objects

## How to tell it works
Copy the skill, then point your assistant at the repository's examples/dashboard.html and say "/unslop this dashboard". It should first list the signs it found and a five-line direction, and stop to ask you before editing. That pause is how you know the skill is on.

## Safety
The two scripts launch a browser and, for snapshots, download a live page's styles and images; run them only on sites you mean to copy. Image generation can cost money; the skill asks before each run.
