
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Design.md is a portable design memory for AI agents — Meng explains it as the “recipe” that carries typography, colors, spacing, and motion rules across tools like Claude, Codex, Cursor, Google Stitch, and Aura so your app doesn’t drift into generic page two and page three designs.
The real win is consistency across mediums, not just one pretty landing page — with a design.md plus HTML and reusable “skills,” the same visual DNA can become a website, promo video, slide deck, mobile UI, and new landing-page sections without losing its soul.
Skills are the missing layer that turns bland AI design into something distinctive — Meng calls them the “ingredients,” using examples like laser effects, skeuomorphic UI, copywriting, 3D globes, and badges to add the kind of special effects that make people stop scrolling.
AI hasn’t reduced the work for great designers — it’s changed the job into judgment and iteration — Meng says he used at least 1,000 prompts on some products, sometimes 10,000, arguing that the craft now lives in taste, workflow, and micro-decisions rather than moving rectangles in Figma.
Taste is now the moat, because the baseline is high but increasingly generic — both Greg and Meng point to the death of the “purple gradient” era, arguing that websites that look like every other startup can lose 10x to 100x in perceived value, even if they’re technically polished.
A solo builder can now operate like a small design studio, but only by staying obsessively current — Meng, who says he’s spent nearly $500,000 on tokens and is building four products at once, frames the edge as adopting new models and workflows fast, then grounding them in a personal library of references and design inspiration.
Greg opens by framing the promise in almost absurd terms: by the end, you’ll be able to design better than 99.99% of the planet. The real point is less hype than leverage — if your startup looks beautiful and coherent, more people install it, buy it, and share it.
Meng introduces design.md as an open-source way to capture the “soul” of a design and hand it to an agent. His metaphor is clean: HTML is the finished dish, design.md is the recipe, and skills are the ingredients — together they let AI reproduce a system, not just a one-off screenshot.
He says most people get one great AI-generated page, then everything after that turns generic. That’s the whole reason for design.md: instead of copying a closed Figma template or a vague prompt you don’t understand, you give the model a foundational system it can keep reusing across pages and formats.
Greg pushes on the fear everyone has: copying a cool design and ending up with another cookie-cutter startup site. Meng agrees and says the baseline has gotten weirdly high and weirdly generic at the same time — what felt fresh five years ago, like purple gradients, now just makes people “run.”
Meng walks through his creative process in tools like Variant and Aura: remix fast, click fast, keep going until something clicks. Then he adds “skills” — prompts for things like lasers, 3D, skeuomorphism, badges, and copywriting — arguing that scroll-stopping design now needs special effects and distinctive flair, not just decent fonts and spacing.
In the live demo, he downloads both the design.md and HTML from a template, attaches them to a prompt, and generates a landing page for a fake startup called Aura. He makes the point that this workflow isn’t tied to one platform: the same assets can drive Aura, Google Stitch, Codex, or other agent-based tools while preserving brand consistency.
Meng then shows New Form, where he can queue multiple outputs from the same core design: mobile layouts, slide decks, motion pieces, hero sections, and more. Greg says it reminds him of Midjourney’s flow state — the magic is not just chatting with an agent, but exploring several creative directions at once and making better decisions.
Meng says the job has shifted from pushing pixels to maximizing “judgment per minute.” He claims AI hasn’t made him lazier at all — he’s building four products at once, often as a team of one, with 1,000 to 10,000 iterations behind the scenes, while insisting the real moat now is taste, niche expertise, and caring enough not to let AI flatten everything into sameness.
Share
Keep Reading
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.

Playbook
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.

Playbook
OpenAI shipped /goal for the Codex CLI. It turns a prompt into a persisted, self-continuing contract.