All projects
Open Source2026

World-Class Design

An open-source skill for getting distinctive design out of AI coding agents

8
Techniques
~30
Design Variants Tested
2
Real Products Tested

Left alone, an AI coding agent returns the centre of its training distribution: purple gradient, text on the left, graphic on the right, eyebrow above every heading. You can ask it to be more creative, but that usually produces a more elaborate version of the same thing.

Then Anshu Chimala and Lenny Rachitsky published a process for pushing AI away from that default and pulling it back towards something shippable. I turned their three stages and eight techniques into an executable skill for Claude Code and Codex, with preflight checks, working prompts, honest fallbacks, and checkpoints where the person using it has to exercise their own taste.

World-Class Design screenshot

Then I Tried to Break It

I ran the skill end to end on two real products — a live SaaS landing page and the personal site you're reading now — across roughly 30 design variants. Every technique was used at least once. That matters because the useful lessons weren't hiding in the clean first draft. They showed up when three different random seeds produced the same kind of design, when a critic gave the same score to two visibly different pages, and when a generated video worked perfectly but was placed so quietly that nobody could see it.

Each failure became a rule in the skill. Seeds change aesthetics more reliably than structure, so the way the model interprets them has to vary too. Critic scores are useful for triage, not progress, so you measure whether old gaps disappear. Generated media cannot stand in for product evidence. Copy is allowed to move while you're exploring, then locks once the direction is chosen. And three moments belong to the user: choosing the ambition, deciding what to remove, and rewriting the final words.

What It Became

The result is an open-source plugin that works in Claude Code, Codex, or any agent that reads SKILL.md files. It is deliberately not a magic 'make this beautiful' prompt. It's a repeatable way to explore several directions, judge the rendered work in a fresh context, use real imagery and motion where they earn their place, and then subtract until the result feels like somebody made choices. The evidence is still small-sample and the fast-moving API references will age. The repo says that plainly too.

Process
Discover · Define · Deliver
The 8 Techniques
Seed strings · Ambitious briefs · Critic loop · Image generation · Video generation · Subtract · Kill AI tells · Rewrite copy
Works With
Claude Code · Codex · Any agent that reads SKILL.md

Watch me build it, mistakes included

Full Stack Founder is a weekly note on shipping products with AI as a team of one. Free, and you can leave whenever.

Already convinced?Subscribe on Substack
Every week
  • 01What I shipped, and what it actually took.
  • 02The tools and workflows earning their keep right now.
  • 03Where I got stuck, and the things that failed.