跪拜 Guibai
← Back to the summary

Killing the AI Template Look: unslop Refactors a Generated Page in 7 Minutes

用 unslop 拯救 AI 生成网页的塑料感,7 分钟重构落地实测

Ask AI to make a web page and nine out of ten come out looking identical: a big blue button, a little ✨ badge, a few patches of glow floating in the background, and a row of green checkmarks under the button. The heavy template plastic flavor is recognizable at a glance.

Starting over is too much work, and hand-editing the CSS piece by piece is hard to know where to begin. This time I put to a practical test an open-source design skill built to solve exactly this problem: unslop.

See the before/after comparison first. Above is the official site for the note app "Suiji" generated by Doubao from a single sentence; below is the actual result after unslop made adjustments in WorkBuddy for 7 minutes:

Before: home page built by Doubao from one sentence: ✨ capsule badge, big blue button, background glow blobs, gray "App screenshot placeholder"

After: serif headings, off-white paper texture, black button, with a proper-looking App interface drawn on the right

Same product, same foundation of copy, and the visual texture of the adjusted result is completely free of the original AI template flavor.

What unslop is

unslop is an open-source design skill package by developer Damir Sokolovsky (GitHub ID: heydamir), purpose-built for interfaces that "look AI-generated at a glance." The repository was created on October 29 and had already gathered 18,330 stars by November 18, a very rapid climb.

Its core logic is not to have the LLM rewrite everything outright. It first performs a page audit, listing the typical AI clichés one by one; then, based on the product positioning, it defines an independent design language; finally it makes changes in rounds, in this order: spacing, color, icons, copy, fonts, layout, details, visuals, and motion.

The tool was originally designed for environments like Claude Code and Codex. I plugged it straight into WorkBuddy for this test, so it can be reproduced without fiddly terminal configuration.

Step 1: Generate a typical "AI-flavored" test page

To have a baseline for comparison, first use Doubao to generate a home page quickly, with no aesthetic hints attached:

Help me make an official site home page for my AI note app "Suiji".

Typing this one sentence into a new Doubao chat, with no design requirements added

Doubao outputs the entire page code directly. The source of the AI clichés is plain in the implementation: ✨ emoji, gradient backgrounds, text shadows, and typical Tailwind preset styles.

The code Doubao generated: ✨ in the badge, gradient background and text shadows on the hero

Have the model consolidate the code into a standalone HTML file and save it as before.html. If you already have a page you want to optimize, use that directly.

One note: the prompt only said "Suiji is an AI note app," and the model filled in features like recording transcription, on-device encryption, multi-device sync, and membership benefits on its own. When generating pages with AI, always check that business claims match reality before the code ships.

Step 2: Load the unslop skill package in the environment

Create a new task in WorkBuddy and hand the model the open-source repository address directly:

I want to install a design skill package called unslop. Can you load a skill package like this? Tell me yes or no directly. Also, please install unslop from github.com/heydamir/unslop, and tell me how to use it after it's installed.

Dropping the GitHub address into WorkBuddy to have it install unslop

Installation takes about one to two minutes. Once loaded, the model lays out unslop's execution chain: complete the audit first, settle the tone, then split into nine rounds of item-by-item adjustment.

Step 3: Run the page refactor with one command

Create a new task, specify the source file path, and ask explicitly to save a copy so the original stays safe:

Please use unslop to strip the AI flavor from (your page path)\before.html, save it as after.html in the same folder, and don't overwrite the original. Tell me how long it took when you're done.

The most distinctive thing about this flow is that it doesn't blindly rewrite code; it delivers a detailed structural audit report first.

The AI cliché self-check items the audit distilled

The audit report sorted out 15 issues in total, clearly marking the specific code nodes and the visual defects:

The latter half of unslop's audit table, and the design direction it set for "Suiji"

These traits are common across today's AI-generated pages, so they work well as a developer self-check checklist:

Once the audit finished, the tool output a clear overall refactor plan:

Because a unified direction was fixed before the refactor, the later rounds of style changes stayed consistent with one another.

Page changes, item by item

The full modification took 6 minutes 56 seconds and consumed 21.49 credits (based on WorkBuddy's default DeepSeek-V4.1-Flash).

In the feature section, the six light-gray rounded cards with big blue icons were rebuilt into a two-column list with table-of-contents typography, and the section title became "Just six features."

Before, feature section: six light-gray rounded cards, each with a blue icon

After, feature section: two-column list with monospaced numbering 01–06, titled "Just six features"

The copy got de-fluffed in the same pass. The original "Open and write; supports quick input, clipboard capture; jot down a sudden thought in one second" was rewritten as "Open and you're in an input box; a sentence from your clipboard comes in with one tap," trimming the adjective pile-up and showing the interaction logic directly.

In the footer, the combination of white FAQ cards over a cold black footer became a Q&A block divided by thin rules over a warm black footer, and the icons were trimmed to text links.

Before, footer: white FAQ cards, cold black footer with social icons

After, footer: Q&A separated by thin rules, warm black footer, social icons replaced with text links

Exception and boundary declaration

After the changes, the tool proactively listed the three details where it did not fully follow its own rules: it used local scripts for screenshots because the environment lacked the Playwright dependency; it did not load the 5 reference images bundled in the skill package; and because the original page had no membership pricing, it removed the "Membership" entry from the nav bar on its own and flagged it for a human to fill in.

The three things unslop admitted it did not do, with the rebuilt feature section on the right

For claims fabricated during generation, like "basic plan free forever" and "on-device encrypted sync," the tool did not unilaterally decide to delete them; it kept them as-is and issued verification warnings. Clearly marking business and logic boundaries is a very necessary trait for this kind of helper tool.

Notes for actually shipping this

First, the App mockup on the right after the refactor is a front-end-generated placeholder; the note content inside it is sample data. Before real production, replace it with actual native screenshots of the product.

Second, avoid selecting elements frequently in the workbench's live preview panel. The preview container may inject large numbers of debug markers such as data-page-node-id into the HTML file (259 were injected in this test). Before a real build ships, write a script or have the model strip these debug attributes.

Choosing between unslop and pre-aesthetic tools

Previously tested pre-aesthetic skill packages establish rules at page-generation time. For example, taste took about 60 minutes per page and hallmark about 42 minutes, and you still had to hover around for permission confirmations at the end.

unslop, by contrast, is a post-hoc governance approach that focuses on fixing template traits in pages that already exist, and it took under 7 minutes.

Here's one way to combine them in everyday front-end and indie-development workflows:

If you have AI-generated pages of your own, pick a copy and test it; use the 15 audit items it produces to re-check your visual code.


👋 I'm Meng Jian, former Tencent T11 and ByteDance tech lead, now doing AI programming full time.

🔥 More AI programming in practice:

Find this useful? Like + save is the biggest support 🙏

This article was compiled by me with AI assistance, and the content was field-tested and reviewed by me.

Comments

Top 1 of 2 from juejin.cn, machine-translated. The original thread is authoritative.

快乐是一切

Hmm, honestly, the before and after comparison in the article just feels like it went from a DeepSeek flavor to a GPT flavor.

孟健AI编程

GPT doesn't have that feel.