Skip to content
Blog

How to give AI coding agents better design taste

By Attuna · · Updated

You give an AI coding agent better design taste by giving it your creative direction in a form it can use: references that show what you mean, written principles about what to prefer and avoid, and constraints it must respect. Agents don’t lack ability; they lack your judgement. Without it they produce the most common answer, and the most common answer looks like everyone else’s.

What creative direction actually means

Creative direction is the set of decisions that make work recognisably yours before any specific screen exists. A creative director doesn’t draw every page. They make sure every page follows the same point of view. For an agent, that point of view breaks down into four parts:

  • References: concrete examples of what good looks like, with notes on why.
  • Likes and dislikes: the qualities you’re drawn to and the ones you reject.
  • Principles: short, actionable rules, stated as something to prefer or avoid.
  • Context: which of these apply where. A dashboard and an editorial page shouldn’t share every rule.

Why generic prompts produce generic outcomes

Language models predict likely output. Anthropic’s applied AI team calls the result distributional convergence: without direction, the model gravitates to the patterns it has seen most. Words like “clean” or “premium” don’t change that, because nearly every interface claims them.

What changes the output is information the model can’t infer. That you dislike decorative gradients. That you want 32px table rows. That a particular product’s settings page is the bar. We break down the common defaults in why AI-generated interfaces look generic.

Turn taste into something an agent can use

Collect references with reasons

A reference without a reason is a mood. A reference with a reason is direction. When you save something, write the specific quality you want: “the way secondary metadata drops to a lighter weight instead of a smaller size”. Our guide to sharing visual references with Claude Code covers how to annotate them.

Write principles as prefer and avoid

Principles work best when they’re short, specific and testable. Pair what you want with what you don’t:

Principles
Prefer: one accent colour, reserved for the primary action and status.
Avoid: gradients as decoration.

Prefer: hierarchy through weight and colour before size.
Avoid: more than five text styles on one screen.

Prefer: hairline dividers and alignment to group content.
Avoid: cards nested inside cards.

Scope it to the work

Not every principle applies everywhere. Dense dashboards and a marketing page need different rules. Group your direction by context so an agent receives what fits the task, not everything you’ve ever written down.

Keep it separate from the design system

Your taste and a product’s design system overlap but aren’t the same. The system says which components and values are allowed; taste guides the choices within them. We explain the split in design references vs design systems.

Make it reusable

Direction typed into one chat disappears with that chat. To make it part of every session, put it somewhere the agent loads or can reach:

  • Project instructions. Claude Code reads CLAUDE.md at the start of every session. Good for rules specific to one codebase.
  • Skills. Reusable instructions Claude loads when a task calls for them, which keeps design guidance out of unrelated work.
  • MCP. A server your agents query for direction and references, so the same taste works across projects and tools. See how to use MCP with Claude Code for design.

Whatever you choose, keep a loop. When you correct the same thing twice, write it down as a principle. Your direction gets sharper with every project instead of starting from scratch.

How Attuna does this

Attuna is a creative direction platform for AI coding agents, built for this workflow.

  • Projects. Each thing you build gets a project with a short brief. Attuna suggests a starting creative direction from it, and you add project rules, likes, dislikes and references. Agents read the project over MCP, with accessibility and function ranked above style.
  • Reference library. Save what inspires you, organise it into collections, and note what you like about each piece.
  • Taste Profiles. Write your direction, principles, likes and dislikes, and the contexts each profile is for.
  • MCP connection. Agents that support MCP over HTTP, including Claude Code, Codex and Cursor, connect with a personal access token. Attuna shows which client last used each token, and you can revoke one at any time.
  • Direction for a task. An agent can ask for the creative direction that applies to what it’s building, and gets the relevant approved principles, things to avoid and references, with the actual images.
  • You stay in charge. An agent can propose a principle based on your feedback, but it stays a suggestion until you accept it. Attuna doesn’t claim to learn your taste on its own.

Your library is visible only to your account. Images are kept in private storage and shown through links that expire. Attuna is in development and we haven’t announced a release date or pricing yet.

Sources