May 6, 2026 · 51m · startup-ideas
My AI Design Workflow That Doesn't Ship Slop
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Host Greg Isenberg sits down with master designer Meng To to break down high-end AI design workflows centered around Google's design.md framework. They explore how combining modular Markdown blueprints, multi-task generative tools, and obsessive human iteration enables founders to create stunning, production-ready interfaces without shipping generic AI slop.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 16.7% of the talking time here. How this is scored →
Greg as informed peer 3.5 Guest teaching 7.2 Guest disagreement 0.5 Greg pushing back 0.1
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 37:00 AI does not make creators lazy
Meng rejects the conventional assumption that generative AI enables passive one-click shortcuts, insisting that true craftsmanship demands thousands of obsessive prompt iterations.
Hardest push from Greg ▶ 41:06 Probing solopreneur team structureGreg gently challenges Meng's solo builder framing, directly questioning whether he is operating entirely alone given his existing company footprint.
Biggest teaching moment ▶ 35:20 Correcting the 20-prompt estimateMeng directly schools Greg after Greg estimates a build takes a couple dozen prompts, revealing that professional-grade AI craftsmanship requires over a thousand micro-adjustments.
Greg holds their own ▶ 13:31 Homogeneous cityscape aesthetic analogyGreg demonstrates strong product and cultural insight by comparing generic web templates to sterile, identical modern cityscapes that customers instinctively reject.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Welcoming Meng To and AI Presentation Setup | 3 | 4 | 0 | 0 | Greg sets the stage by welcoming Meng To back to the show and introducing the topic of design.md. Meng graciously credits Greg's podcast for multiplying his revenue before sharing his screen for the presentation. | |
| Unpacking design.md as the Soul of Design | 2 | 7 | 0 | 0 | Meng delivers a technical breakdown of design.md, explaining how markdown files act as the soul and design system for AI agents across multiple media formats. | |
| Greg Isenberg's AI Startup Workshop Announcement | 3 | 6 | 0 | 0 | Following an ad break, Meng explains the structural difference between finished HTML and markdown blueprints. Greg prompts Meng on how non-designers can discover visual identities. | |
| Eliminating Design Drift Beyond Static Templates | 3 | 8 | 1 | 0 | Meng breaks down the core problem of 'design drift' in AI tools, arguing that locked community templates in Figma or Framer fail when translated into autonomous agent workflows. | |
| Combating Generic Aesthetics with Local Design Memory | 5 | 7 | 1 | 0 | Greg draws an architectural analogy about homogeneous cookie-cutter cityscapes, which Meng validates before detailing how local design memory in Codex and OpenClaw protects proprietary workflows. | |
| Live Demo: Remixing Skills and Landing Pages | 3 | 8 | 0 | 0 | Meng performs a live demonstration showcasing how modular skills, WebGL effects like lasers, and markdown files are fed directly into prompts to generate unique interfaces on Aura. | |
| Multi-Task Generative Queuing in Newform and Stitch | 5 | 7 | 1 | 0 | Meng evaluates Google Stitch against his startup Aura and explains the efficiency of parallel generative queues in Newform. Greg connects this queuing dynamic to his past experience with Midjourney. | |
| The Skills Maxi Philosophy and Modular Prompts | 4 | 8 | 0 | 0 | Meng explains the 'skills maxi' philosophy, advocating for modular workflow skills over bloated global agent files. He demonstrates prompts for skeuomorphic shading, 3D elements, and precise typography. | |
| Iterative Craftsmanship and Infusing Soul into AI | 3 | 9 | 1 | 0 | When Greg guesses that Meng uses 20 to 25 prompts per product, Meng corrects him by revealing he executes over 1,000 iterations to infuse genuine craftsmanship and soul into AI-assisted builds. | |
| Solo AI Velocity Versus Long-Term Team Craft | 4 | 7 | 0 | 1 | Greg probes Meng on whether he is truly operating as a solopreneur. Meng explains his hybrid model of rapid solo prototyping backed by a trusted human team to provide long-term craft and authenticity. | |
| Cultivating Taste and Niche Moats Against Commoditization | 4 | 8 | 1 | 0 | Meng summarizes the ultimate defensible moat against AI commoditization: developing refined, niche taste and refusing to settle for generic aesthetic tropes. |
Statements from this episode (18)
To: Google released design.md as a cross-medium design blueprint
“Google just came out with design.md, and this is one way to create a blueprint to take one beautiful design, Whether you made it or not, and then spread throughout the, all the formats and the mediums for your startup idea, right?”
To: Appearing on Isenberg's podcast grew MRR from $3K to $15K
“Remember the first time I came to your podcast, I started with a 3000 dollars MRR, and it shut up to 15,000 MRR just by going through your podcast.”
To builds all his software and tools solo using vibe coding
“Everything I do today, including this tool that I'm showing you, which is my own notion with slides,
Is vibe coded, and I made it by myself.”
To: Attaching design.md files transfers design systems to AI agents
“So basically this is for designers, and if you want to take the soul of the design and you want to bring that
To the agent, and you want to bring a design system, the colors, the typography that makes the design beautiful.
Well, you port it, and you put it all…”
To: design.md is the recipe, skills are ingredients, HTML is the dish
“The HTML is more like The sort of finished dish, and the MD file is more like the recipe. The skills are like the ingredients.”
To: Traditional UI templates are incompatible with AI coding agents
“You can go to any of the templates that you find on VZero, Lovable you know, all of these other place Framer, Figma community for people who are more familiar with designs, and the problem with those is that they are highly close behind the techniques and the …”
To: Foundational design systems prevent AI from being too literal
“Now, if you only provide a portion of it or something too specific, the AI tends to sort of like take that too literally. But if you provide something more, ah, like a foundational system, like a process, agents.md and what not, then you get something that giv…”
To: Outdated design tropes like purple gradients now repel users
“What used to be wow, five, 10 years ago, now is generic, which is funny, because like, if you look at the purple gradient, everyone was like, wow, purple gradient, I want to use that. But nowadays, if you see something with, you know, purple gradients, you jus…”
To: design.md creates portable design memory across AI tools
“I like to think of design.md like a design memory, and this is great because it can transfer between platform. So, if you're working in lovable, and you like what you've done, or in cloud design, and then you want to move it to Figma, or you want to move it to…”
To: Local-first architecture is the key moat against Notion and Figma
“First of all, for me, when I make apps, I look at what is the mode? What is something that I can do with my app that Notion cannot do or Figma cannot do? For me, that mode right now is the fact that everything is local.”
To: Basic typography and colors are no longer enough for landing pages
“Like, nowadays, just having the typography Just having the colors is not enough unless it's a secondary page or unless it's something that is very serious, but if you're talking about a landing page or a promo video or, you know, a slide, you need a moat.”
To spent nearly $500,000 on AI tokens for his startup
“And I just realized, you know what, like, I spent almost half a million dollars in token for my startup right now.”
To: AI builders must rapidly adopt new models to maintain moats
“As someone who works in AI, your mode is basically, you have to be new, you know, you have, like, I don't, I'm sure anyone can relate, But you always, when there's a new model that comes, you have to talk about it. Otherwise you're kind of losing a lot of that…”
To: Global agent prompt files consume more tokens and lack specificity
“If you only change the agent.md, oftentimes, you know, first of all, it costs a lot more tokens, but also it's too general, right? It doesn't apply to every single workflow out there.”
To: AI shifts design from moving pixels to high-frequency judgment
“So, what I believe is that everything is changing, and we're doing a lot less of the moving pixels, moving rectangle and resizing things. We're doing a lot more judgment per minute. In fact, you know, For me, the agent is more like someone who's doing a lot of…”
To: Products built purely with AI will show visible defects
“They want real people, they want real craftsmanship behind all of these pixels at some point, because, you know, the cracks are going to show if you only use AI. The, you know, the whole way.”
To: Derivative product design slashes perceived value by up to 100x
“Niche is the key word here because you cannot compete anymore with people who are generalized. It's too difficult. Nowadays we want authenticity, we want niche, we want taste, and if something looks like another thing, it kind of reduced the value by, like, al…”
To: Only delegate workflows to AI if you have already mastered them
“Please use your, ah, you know, your taste every step of the way, and don't let AI do everything for you, and even if you do let AI do a lot of things for you, make sure that it's part of the workflow that you already mastered, right?”