AI world-class designer: a skill for escaping the generic look

TL;DR

• The skill runs a three-stage process — Discover, Define, Deliver — with seven techniques for getting a design out of the patterns a model defaults to.
• The diagnosis is mechanical: an LLM predicts the next likely token, so at every design decision it picks the most predictable choice, and every page converges on the same purple gradient with text left and a graphic on the right.
• The fix is what the model cannot produce alone: external randomness (seed strings) and external taste (fresh-context critic subagents, your own reactions), plus image and video generation to replace the gradient-and-shapes tell, plus a final subtraction pass.
• It is one Markdown file with no dependencies and no API keys to start; keys only unlock two of the seven techniques.

What it is and who it is for

AI World-Class Designer is an agent skill, distilled from Anshu Chimala's post "How to turn your AI into a world-class designer" on Lenny's Newsletter (September 1, 2026), with credits to the author.

It is for the person whose generated pages work and look like everybody else's generated pages. The README's framing is precise: the problem is not that the model lacks skill, it is that the model is doing exactly what it is trained to do — predicting the most likely next token, which at a design decision means the most predictable layout.

The diagnosis

LLMs are next-token predictors. At every design decision they pick the most predictable choice, so everything converges to the same purple gradient and the same text-left/graphic-right landing page. No amount of "make it beautiful" in the prompt breaks the pattern, because "beautiful" is also predictable.

The skill brings in the two inputs the model cannot produce by itself: external randomness, through seed strings, and external taste, through fresh-context critic subagents and your own reactions. Then it adds image and video generation to replace the gradient-and-shapes tell, and closes with a subtraction pass to remove the remaining AI fingerprints.

The README credits the lineage honestly: Technique 1 is based on String Seed of Thought, published by Sakana AI, and the process is loosely inspired by the Double Diamond) model.

How to use it

Any agent: it is one Markdown file, no dependencies and no API keys needed to start.

git clone https://github.com/artificialguybr/ai-world-class-designer /tmp/awd
cp -r /tmp/awd/skills/ai-world-class-designer ~/.claude/skills/ai-world-class-designer

Claude Code reads ~/.claude/skills/ globally or .claude/skills/ per project. For the other agents the directory changes but the file does not: Cursor reads .cursor/skills/, Codex .codex/skills/, Gemini CLI .gemini/skills/.

Then ask in the shape you actually have:

design a landing page for my productivity app
make this page look designed, not AI-generated
restyle my dashboard — I hate the generic AI look
polish this UI to a top-studio level

The skill triggers on design, restyling and polishing requests, and drives the full loop: seed-string exploration, then an ambitious brief written with your taste, then critic-subagent iteration until the critic scores it 9 out of 10, then image and video enrichment, then the final subtraction pass.

What makes the loop worth running

Two parts are doing the real work. The seed strings force a starting point you would not have chosen, which breaks the convergence at the source rather than asking the model to be less predictable. The critic subagents get a fresh context, so they judge the rendered result instead of the conversation that produced it — the same reason a second pair of eyes reads a diff better than its author.

The subtraction pass is the part most design prompts skip. Adding detail is the model's default mode; removing the remaining AI fingerprints (the decorative blobs, the extra rounded card, the third gradient) is the step that makes the result look like a person stopped.

When to use it, when not to

Use it when the page works but looks generated, when a dashboard needs to stop looking like every dashboard, or before you publish a landing page that will be compared against better-funded competitors. The skill is cheap to run: one file, no dependencies.

Do not expect it to make design decisions for you. The README says the model cannot produce taste alone, so the loop depends on your reactions feeding the brief — if you accept the first output, the whole approach collapses into the usual generation. And do not use it to replace image and video assets: the two techniques that need keys enrich the page, but the layout, hierarchy and subtraction still have to happen.

Sources

1. AI World-Class Designer — repository — three-stage process, seven techniques, install paths, credits and licence quoted above.
2. README.md — install commands, usage examples, attribution.
3. Anshu Chimala, "How to turn your AI into a world-class designer" — Lenny's Newsletter, September 1, 2026: the original article the skill distills.

Project page: AI World-Class Designer

For the asset side of the same problem, see generating app icons instead of gradient placeholders.

FAQ

Does it need an API key? Not to start. The README says it is one Markdown file with no dependencies; keys only unlock the image and video techniques.

What are seed strings for? They inject external randomness into the starting point, so the model does not begin from its most predictable choice. The technique is based on Sakana AI's String Seed of Thought.

What does the critic subagent do? It reviews the rendered result in a fresh context, so it judges what is on screen rather than the conversation that produced it.

Where does it work? Claude Code, Cursor, Codex and Gemini CLI through their skills directories; the file is plain Markdown, so any agent that reads files can use it.