From Generic to Guided Design: How to Escape AI Slop in Your Web Pages

PUBLISHED: 2026-08-06
AUTHOR: MANUEL PRIETO
Artificial-intelligence

If you have ever tried to generate an interface with an AI assistant without giving it concrete rules, you surely recognize the default result:

  • Recursive boxing (boxes inside boxes): Outer containers wrapping cards that in turn have duplicated borders, backgrounds, and paddings.
  • Generic fonts like Inter or Roboto.
  • Purple or gradient buttons with standard rounded corners.
  • Disproportionate spacing (meaningless paddings) and endless cloned components.

This phenomenon is known in the industry as AI SlopGlosarioAI SlopTérmino que define la morralla, bazofia o contenido basura generado en masa por modelos de inteligencia artificial sin supervisión ni valor editorial. Proviene del inglés slop (literalmente la comida de desperdicios o sobras que se echa a los cerdos). En la cultura digital representa el sucesor directo del spam: textos inflados, imágenes clónicas, código genérico y relleno algorítmico sin criterio ni alma que saturan internet y provocan el colapso de los propios modelos al retroalimentarse de su propia basura.Ver término completo → ("visual junk").

Where does the term come from? Slop was coined by the tech community and popularized in 2024 by developers like Simon Willison, drawing a parallel with the term spam from the 90s. In web development, it refers to those generic and soulless interfaces that AI automatically generates by default.

The "Recursive Boxing" Vice: Why the AI Fills the Screen with Boxes

One of the visual symptoms that most reveal an AI-generated interface is the compulsive accumulation of nested containers. The model tends to wrap each section inside a large outer card with a border and background, and then boxes each item again within its own border and padding.

This recursive boxing (boxes on top of boxes) creates serious design problems:

  1. Reduces usable space: Accumulating external paddings in wrapping containers compresses the actual content on screen.
  2. Generates visual noise: Duplicating borders and background layers breaks the natural hierarchy of the web flow.
  3. Creates "div soup": Overloads the DOM tree and makes CSS maintenance difficult.

To achieve a clean and professional interface, we must prohibit redundant wrapping paddings in our design contract: content elements (such as accordion dropdowns or information cards) must rest directly on the page flow without unnecessary outer boxes.

To escape this default template, you do not need to change frameworks or learn complex programming tricks. The solution lies in applying a method of guided design: creating a centralized file called DESIGN.md that serves as the aesthetic source of truth for the AI, and making every instruction in your project revolve strictly around it.

Why the AI Delivers Generic Responses: The "One Prompt, Three Jobs" Trap

To understand the solution, we first need to understand what happens inside the model when we ask it for a screen with a vague phrase like "make me a modern and beautiful interface".

Think of the AI as a designer who has studied millions of web pages from all over the internet. When you make an abstract request without restrictions, you are asking it to make three complex decisions at the same time:

  1. The art direction: Choosing colors, fonts, contrasts, and visual personality.
  2. The user experience (UX): Deciding the layout of elements and the screen hierarchy.
  3. The code architecture: Writing the HTML and CSS rules that build the interface.

By forcing the AI to take on all three roles at once in a single message, the assistant applies the default standard response (the average "wildcard" from its training). It chooses the most common fonts on the web and the most common color combinations to avoid risking giving you something that does not fit or fails to compile.

Cargando diagrama...

Practical Example: What NOT to Do (Vague Prompt) vs. What TO Do (Guided Prompt)

To visualize with absolute clarity the difference between delegating blindly and guiding the AI, let's analyze the model's behavior facing two requests.

The Wrong Case: The Vague Prompt that Generates AI Slop

When a developer writes a naïve instruction without aesthetic or contractual restrictions, the assistant applies the standard average market template (the classic AI Slop). You can click Launch Prompt to simulate the generation. If you click Re-launch Prompt again, you will see how the lack of a contract causes the AI to invent a different generic design on each attempt (switching between 3 AI Slop variants):

Caso Erróneo: AI Slop (Sin Contrato)
Prompt Enviado (Vago sin Contrato Visual):
"Créame una tarjeta de producto moderna para una tienda online de tecnología con una imagen, un botón de compra bonito y una lista de especificaciones."
>_
Simulador de Respuesta del Asistente
Haz clic en el botón "Lanzar Prompt" para enviar esta instrucción y ver el resultado visual que genera la IA en tiempo real.

Analysis of the Errors in the Generated Response:

  • Generic typefaces: The AI resorts to Inter or Roboto (font-sans), missing the opportunity to give the project visual identity.
  • Cloned gradient button: The button uses the typical violet/indigo gradient (from-indigo-600 to-purple-600) with bloated rounded corners (rounded-full), a worn-out pattern found in thousands of free templates.
  • Inflated and disproportionate spacing: The container uses a 40px padding (p-10), generating artificial voids that force the user to scroll unnecessarily.
  • Excessive shadows: The large-radius floating shadow (shadow-2xl) gives it a bulky appearance misaligned with modern technical precision.

The Correct Case: The Prompt Guided by Tokens and Contract

Instead, let's observe what happens when we write an instruction tied to the specifications of our DESIGN.md and protected by the three-level architecture of AGENTS.md (for a detailed analysis of the three-layer technical specification, see our guide on Guided UI Design for AI Agents). Press Launch Prompt to verify how, with a DESIGN.md contract in place, the AI generates 100% deterministically the one exact design required:

Caso Correcto: DESIGN.md (Determinista)
Prompt Enviado (Guiado por DESIGN.md):
"Construye la tarjeta de producto para los 'Auriculares Pro X' consumiendo estrictamente el sistema de DESIGN.md: - Maqueta la tarjeta con el badge 'PRODUCTO DESTACADO', título principal, descripción y rejilla de 2 especificaciones técnicas. - Consume los tokens de las secciones 2 (Sistema Tri-Font), 4 (Tarjetas Compactas) y 5 (Botones de Acento) declarados en DESIGN.md sin hardcodear estilos."
>_
Simulador de Respuesta del Asistente
Haz clic en el botón "Lanzar Prompt" para enviar esta instrucción y ver el resultado visual que genera la IA en tiempo real.

Comparing both approaches reveals the fundamental principle: the AI does not design poorly due to technical inability, but due to a lack of design boundaries made explicit in the prompt.

The Guided Design Method: DESIGN.md as the Center of Gravity

To break this inertia and get the AI to create unique interfaces, the key is to separate the creative stage from the implementation: centralizing all visual identity in a single document (DESIGN.md) and making all prompts in your workflow depend on it.

Cargando diagrama...

The 5 Steps to Build Your Project Around DESIGN.md

1. Develop Your Own DESIGN.md (The Source of Truth)

Instead of improvising colors and typefaces in each chat message, we create a DESIGN.md file at the root of our project.

For a DESIGN.md to be truly effective, the first indispensable step is to include a System Guardrails block that restricts default options through contrast rules:

Let's avoid relying on unbranded generic fonts (such as Inter or Roboto by default)
The reason why many AI-generated web pages look identical is that models always converge on this same group of standard fonts. By defining specific typography rules in our DESIGN.md contract, we ensure that the AI applies fonts with their own identity aligned with the project.

Below is a complete specification inspired by a professional light-background design system:

Real Demo: A Simple Design Faithful to DESIGN.md

Without resorting to complex prompts of thousands of words or unnecessary fireworks, below is the real result of a website created by the AI when required to directly comply with our DESIGN.md guidelines:

Real result of a web cover guided by DESIGN.md (MANU & CO)

Real result of investment models and footer guided by DESIGN.md

Visual Coherence Analysis with DESIGN.md:

  1. Respect for the Required Color Palette:

    • Main background in soft linen (#FAF7F2) giving it a print warmth.
    • Primary and featured buttons in terracotta coral (#C84B31).
    • Lower block (footer) in deep night blue (#2D4263) with light text for high legibility.
  2. Tri-Font Typographic Hierarchy:

    • Headlines (Clash Display): Headlines ("DESIGNING VENTURES", "DEVELOPMENT PROCESS", "INVESTMENT MODELS") use bold uppercase with wide tracking and enormous presence.
    • Body text (Albert Sans): Clean and readable paragraphs for descriptions.
    • Metadata and Technical UI (Fira Code): Used strictly for utility labels in uppercase (LOCATION: Zaragoza/Madrid/Barcelona, AREA, STATUS, prices in €/min and footer menu).
  3. Card and Button Structure:

    • Cards: No bloated floating shadows or excessive effects. Clean borders, adjusted paddings, and compact internal grid for the technical data of each project.
    • Buttons: Sober buttons with border-radius: 4px in terracotta coral (APPLY NOW, START PROJECT) and soft variants for secondaries (VIEW DETAILS, CONTACT).

The goal of the method is clear: the AI builds exactly the design boundaries declared in the visual contract, achieving a clean, readable interface 100% coherent with our DESIGN.md.

2. Create a Development Contract that Shields the File (AGENTS.md)

In long conversations, the AI tends to forget the instructions from the beginning due to context displacement. To prevent it from reverting to generic fonts mid-project, we create an AGENTS.md file (or .cursorrules) that acts as a System PromptGlosarioSystem PromptInstrucción de nivel superior proporcionada a un modelo de inteligencia artificial antes de procesar las peticiones del usuario. Establece las reglas de comportamiento, restricciones de estilo, contexto arquitectónico y contratos de diseño (como archivos AGENTS.md o CLAUDE.md) que el agente debe respetar estrictamente durante toda la sesión de interacción.Ver término completo →. Its main function is to force the AI to consult DESIGN.md before generating code (for a complete breakdown of the three-layer contract architecture, see our knowledge section guide on Guided UI Design for AI Agents).

Why do we use XML-style tags in AGENTS.md?

  1. Context Scoping: Modern language models (such as Claude or GPT) are trained to process XML tags. Delimiting instructions within system tags (for example, <web_designer_rules>) tells the assistant that this content is a rigid behavioral contract and not a simple conversational suggestion.
  2. Attention Anchoring: Opening and closing tags prevent key rules (like consulting DESIGN.md or not using bloated margins) from diluting as the conversation grows.
  3. Modular Referencing: Allows you to invoke specific rules in the chat by simply saying "Build this component respecting the AGENTS.md contract and the DESIGN.md tokens".

3. Fix the Style System Consuming DESIGN.md

Once your DESIGN.md is defined, don't let the AI choose libraries at random or inject hardcoded colors into the code. Explicitly require it to build elements consuming the rules declared in your document (including cards and tables with grid divs).

4. Replace Adjectives with Exact References to DESIGN.md

Words like "elegant", "clean" or "impactful" mean nothing to a machine. To get exactly what you want, replace ambiguous adjectives with direct references to the sections of your DESIGN.md.

5. Micro-interactions and Compact Density Tied to the Base

Don't ask for "cool animations" that load heavy JavaScript libraries. Ask for native CSS micro-interactions and compact designs that respect the adjusted paddings declared in your DESIGN.md.

Practical Example: The AGENTS.md Contract that Protects DESIGN.md

Below is an example of a system rule structured in XML blocks for your AGENTS.md or CLAUDE.md file. Its goal is to act as a "guardian" that prevents the AI from straying outside your DESIGN.md specifications:

Dive Deeper into the Technical Documentation

If you want to see a complete implementation with code of this design system (HSL variables, polygonal patterns with clip-path and CSS animations), you can consult our complete guide in the knowledge section: Guided Design System for AI Agents.

Frequently Asked Questions about Guided UI Design