How to Work with AI Assistants in Design

## AI assistant is not a search engine or autopilot

The first thing to understand is that an AI assistant is not Google, which you ask and get a link to. And not the autopilot to whom you say "fly there" and go to bed.

The right analogy is a **experienced jun with good taste ** who does everything you say but does not read thoughts. He doesn't know the context of your project if you didn't give it to him. You don’t know your style unless you describe it. You don’t know what’s “wrong” unless you say what’s wrong.

It changes everything about how tasks are formulated.

## Four modes of working with AI in design

AI assistant design is used in four fundamentally different modes – and each requires a different approach.

### Mode 1 - Generation

You give the intention, the AI creates the first draft. Stitch, Figma Make, v0 – all work in this mode.

The key rule is: don’t skimp on context. A vague prompt gives an average result. A good generator contains four things:

1. Who the user is and what he does with the product
2. What specific task does this screen/component solve?
3. Feeling and style (reference is better than an adjective)
4. Which is exactly not necessary

"Make a dashboard" is bad. “Dashboard for a solo developer, manages tasks and time, style like Linear, dark background, no illustrations and gradients” – good.

### Mode 2 - Iteration

There is a draft, you specify it. This is the most frequent mode in real work.

The main mistake here is to write “redo” when you need to write “change this specifically.” AI doesn’t know what you didn’t like, it will regenerate everything and lose what was good.

Iteration rule: ** one prompt, one change**. "Make the button bigger and change the font color and remove the shadow and add the icon" are four changes. AI will do all of them, but with a high probability something will break. Better one at a time.

And always call what’s wrong: “The button is too small, at least 44px in height” works better than “the button looks weak.”.

### Mode 3 - Evaluation

Asking AI to rate its design is an underrated mode. It works especially well when you do not see the obvious problems - "eye blurring.".

Useful requests in this mode:

- Find hierarchy issues on this screen
- What the user might not understand at first sight?
- Check Readability: Is the Contrast Enough?
- What’s more, what can be removed without losing meaning?

An important nuance: AI in evaluation mode is prone to flattery - says "good" more often than necessary. Helps direct instructions: “do not praise, only criticize” or “find at least three problems”.

### Mode 4 — Explanation

Asking AI to explain the solution is useful for both teaching and presentation. “Why such a leyaut?”, “What patterns are used in this component?”, “How to justify this decision to the stakeholders?”

In this mode, AI works as a thinking partner, not an executor.

## How to give context: three levels

The more AI knows about the project, the more accurate the outcome. Context can be given on three levels.

### Level 1 - In Prompt

The simplest thing is to explain it all over again. Works for one-off tasks, but is ineffective for long-term projects - waste tokens and time replicating.

### Level 2 - In system files

DESIGN.md and AGENTS.md at the root of the project – AI reads them automatically before each task. Once described the design system, brand style, rules – and do not repeat. More about this in the article about DESIGN.md.

### Level 3 - In Skill Files

SKILL.md is a specialized context for specific types of tasks. For example: “When you make a form, always follow these rules.” It only loads when you need it. More information about SKILL.md.

For most projects, level 2 is sufficient. Level 3 is justified when the same task is repeated frequently.

## Five Mistakes That Kill Results

1. It's too long at a time

It is not necessary to describe the whole product in one prompt. It is better to start with one screen, establish a direction, then move on. AI loses focus on very long instructions.

**2. There is no reference, there are only adjectives. **

“Minimalistic,” “modern,” “clean” are not instructions, they are wishes. The AI doesn’t know your “minimalist.” “Like Linear,” “like Notion,” “like Bloomberg Terminal.” The difference in output is significant.

**3. Accept the first result or reject the whole thing

A good approach is to take the best of the first result and iterate from it. Not "redo it" but "leave the leaut, change the color and typography.".

**4. Don't check the results

AI confidently does things that look right but don’t work: unreadable contrast, illogical flow, a component without a hover state. Always check the result as a designer, not just as a customer.

**5. Try to explain "what's wrong" instead of "what should be." **

“It looks too cumbersome” – AI will guess what you mean. Reduce padding from 32px to 16px, remove the shadow from the cards.

## How to build dialogue in practice

The working scheme for any design task:

1. Context: who is the user, what is the product, what is the task
2. Objective: What specifically needs to be done
3. Limitations: what not, what style, what reference
4. Outcome format: one or more options, with or without code

Example for component generation:

Context: SaaS for task management, audience – developers,
Design system at DESIGN. md

Task: Task card for kanban board

Restrictions: dark theme, status through a colored dot (not badge)
without avatars, priority through the icon on the left

Format: One variant, React + Tailwind

After the first result, iterate point by point, one change.

## When AI won't help

An honest list of tasks where an AI assistant in design gives a bad result:

• Strategic decisions: what to build and for whom. AI generates options but does not make product decisions.
• Cultural context - design for a specific audience (local patterns, age characteristics). AI knows the statistical center, not the niche.
• Original visual language - AI reproduces existing styles well, poorly invents fundamentally new ones.
• **The final judgment of quality is ‘good or bad’.

Everything else — drafts, iterations, verification, variant generation, code translation — AI does quickly and reasonably well.

## Main rule

The AI assistant is a powerful amplifier tool, but the human remains the director of the process. The better you know how to set a task and evaluate the result, the better AI works. This is not about industrial engineering as a separate profession. It’s about the fact that clear thinking has always been a designer’s top skill – AI just made it more obvious.