Turning Your AI Into a World-Class Designer

Eight techniques for getting distinctive, non-generic design out of coding agents.

CheatsheetLayout, Section, NumberedItem, ExampleSection, CodeBlock, Callout, } from "@/components/CheatsheetLayout";

Why AI design is generic by default

A language model predicts the most likely next token. Every design decision, from colour to layout, becomes the choice most likely to please everyone. The result is design by committee: a purplish gradient, text on the left, graphic on the right, the same structure every time. Great design does the opposite. It starts with a feeling and makes memorable, unexpected choices. The job is not to find a better model but to push the one you have past its most predictable choice at each step. The process is a Double Diamond rebuilt for agents: Discover many directions, Define one design identity, Deliver by polishing away the rough edges.

Discover: inject real variety with seed strings

Asking for "something totally unique, make every decision at random" does not work. The model predicts tokens that sound random but are not: same palette, same awkward metaphors, every run. Randomness has to come from outside the model. The String Seed of Thought technique (published by Sakana AI) has the agent generate a random string and treat it as inspiration:

Four runs of the plain prompt in Claude Code gave four near-identical pages. With the seed string, each run had its own palette, fonts and ideas, and no two matched.

Discover: be much more ambitious

The other lever is specificity. Name an inspiration (a video game, an interior trend, an art installation) and say how it should shape the output. Three prompts from the source, all for the same productivity app: A bold pixel art theme where each section feels like a still from a video game, yet works as a landing page. An isometric living 3D city where features are neighbourhoods or buildings. A radically asymmetric layout with dissonant colours and uncomfortable negative space, "break all the rules but still make it look good". Original ideas are the hard part, and asking AI for them directly returns the same average list everyone else gets. Use a three-step system:

List ideas, intentionally shallow

Ask for as many bold, unique design language ideas as it can produce, each with a short high-level description, going broad rather than deep. The list is there to spark your imagination, not to be used as-is.

Visualise your favourites and react

Pick one (say, Industrial Control Panel) and note your reactions: tactile clicky buttons, yes; cartoony skeuomorphism, no; grey gradients are boring, add texture and some colour. Then ask the AI to sharpen the idea based on your tastes.

Iterate, then have AI write the build prompt

Once the direction feels right, hand the prompt-writing back to the model:

Because you steered the direction, the result is something only you could have produced. Try ideas that sound terrible. Keep the prompts that fail and rerun them when new models ship.

Define: feedback loops with a critic subagent

Even seeded designs lean on stale patterns: nav bar up top, text left, CTA below, graphic right. Asking the same agent to improve its own work fails because it is not objective; it reviews its own code and rationale. Instead, have the coding agent ask a separate design critic. The critic sees only a screenshot, in a fresh context, with no code or earlier critiques. It names the aesthetic the design is going for, imagines how a top studio would execute it, outlines the biggest gaps and scores the design out of 10. The implementer keeps iterating until the critic independently gives 9 or higher, and that threshold is never told to the critic. A cheap, fast model does the grunt work while an expensive model provides taste. In the source's runs, Claude Fable 5 as critic used under 10% of output tokens; asking it to redesign directly would have cost about twice as much. Bad: "judge if our design looks beautiful, not AI-generated". Too subjective, results vary wildly. OK: "visualise how a top design studio would execute this aesthetic, then judge against that bar". Great: give it 4 professional examples and 1 screenshot of yours, "rank them by polish and taste level". Provide example images as a moodboard or baseline, never as a target to copy. Set stopping criteria carefully. Run one or two iterations first and check the loop is converging before adding more. Bigger model for the critic; the implementer can be smaller, but still able to execute a direction well.

Define: enrich with image and video generation

Coding agents love to write code, so they reach for gradients, shapes and basic patterns instead of images. Those are strong giveaways of AI design, and most agents underuse image tools until told to.

Images for personality

Tell the agent the design is too plain and to add personality with image generation, combining images with shaders or 3D effects, then verify the result frame by frame in the browser. Codex, Antigravity and Grok Build have image generation built in. Claude Code can call the Codex CLI against a ChatGPT subscription, or use an OpenAI or Gemini key. Give the agent its own key with a tight spend limit, kept in a gitignored .env.agents file and noted in CLAUDE.md as development-only.

Video for motion

Use an aggregator like fal.ai so one key covers many models. Two patterns: generate a looping clip on a solid background and matte it out, so the animation layers into the UI without looking like a video (the source turned a static crystal into a splintering, refracting glass clip). Or interpolate between keyframe stills to create transitions scrubbed by scroll. A suitcase demo (floating, landing and popping open, contents dropping in) took one prompt in Codex with Seedance 2.5.

Deliver: the edit pass is yours

AI builds striking visuals; your judgement makes them usable. Most of the polish effort goes into removing things.

Cut what does not add value

A calorie tracker asked for as "clean, minimalist" still came back with pink glows, random text highlights, redundant labels and custom controls worse than native iOS ones. The fix: an image-centric grid, no gradients or glows, Apple-native components. Deleting feels risky to a model, so the push has to come from you.

Remove AI tells

Every model overuses certain patterns. Do not ban them up front; that makes the model overthink and invent stranger ones. Review the finished design against a list of tells and ask for alternatives one at a time.

Rewrite copy by hand

Treat generated copy like Lorem ipsum: placeholder text that shows structure. A human reads every line and rephrases it in one consistent voice. The human version is almost always shorter and simpler. Pair this with AI Prototyping for PMs for the tooling basics; this sheet covers what to do once the prototype works but looks like every other AI page. The model cannot be random and will not take risks on its own. Variety, ambition and restraint all have to come from outside it, which means from you. Focus on the feeling you want users to have, and the design will stand out whether AI built it or not. Source: Anshu Chimala, "How to turn your AI into a world-class designer", Lenny's Newsletter (Sep 2026). );