Meng To - Codex superpowers for designers

2026-06-30 - 36 min - source - Read full transcript

Vid (host)Meng To

Key insights

Meng writes 0% of his own code now, down from roughly 95% self-written 16-18 months earlier, using Codex for the entire build process.

He frames this as the natural endpoint of a shift that started with Cursor and Sonnet 3.5: Codex introduced a thread-based way of building everything that he says Cursor has since converged toward, making the two apps feel nearly interchangeable except for the underlying model.

ai-native-design-workflow

Staying ahead of AI-generated visual cliches requires constantly moving the target, because whatever look becomes common becomes 'AI slop' within months.

Meng cites purple gradients as an example of a style that looked cutting-edge five years ago and now reads as generic; he predicts serif fonts and certain default animations are next in line to become dated, so designers must treat taste as a moving target, like playing chess a step ahead of the baseline.

avoiding-ai-slop

The single most reliable defense against AI slop is grounding prompts in a designer's own accumulated personal material rather than starting from a blank prompt.

Meng references his own archive of fonts, past Figma designs, and code he's written over his career as a 'starting point' fed to the AI; because that material is inherently his and not generic, the AI's output stops defaulting to generic AI-slop patterns.

avoiding-ai-slop

Designers should build their own internal tools rather than rely only on existing SaaS products, because building has become nearly free and reveals new design possibilities along the way.

Meng built a local, markdown-based alternative to Notion because Notion's AI doesn't function offline; the example shows building a personal tool is now cheap enough that the exercise itself surfaces workflow improvements (e.g., auto-generating cover images from document context) that wouldn't occur to someone using an off-the-shelf tool.

local-first-tools

Knowing your own product deeply is what makes AI-assisted building valuable; without it, competitors can replicate any single AI-generated output in one prompt.

Meng argues that a single polished output means little because someone else could reproduce it in one prompt three months later; what can't be easily copied is the accumulated product knowledge from thousands of prior iterations, which is why he estimates using at least 10,000 prompts per project.

solo-builder-productivity

Screenshots, voice, and direct on-element annotation are all faster and more precise inputs than typed text for directing AI on visual work.

Meng uses a screenshot-capture shortcut (command-command in Codex) as his default way to start almost any fix, recommends voice because people speak three to five times faster than they type and add detail they'd otherwise skip, and prefers annotating a specific UI element directly over screenshotting a whole screen and verbally locating the problem.

prompting-technique

Feeding a video of a desired animation to the AI and asking it to generate the descriptive prompt removes the need to know motion-design terminology.

Meng says his team now shows Codex a video and asks 'how should I prompt this animation,' letting the model translate visual motion into technical language rather than requiring the designer to already know animation vocabulary.

prompting-technique

Frontier models only pay off when given large, context-rich prompts; a simple request wastes the model's extra capability.

Meng says a basic landing-page prompt produces nearly identical output whether run on Fable 5 or a mid-tier model like Opus 4.8/Sonnet 4.6, because the task doesn't require the larger model's power; the value of a high-end model only shows up when the prompt bundles a full design.md, reference screenshots, prior context, and detailed section-by-section requirements.

prompting-technique

A design.md file is the AI-era equivalent of a designer's personal style guide, encoding years of accumulated taste into a document a model can read and apply.

Meng explains design.md holds colors, typography, spacing, corner radii, and other design-system tokens; a capable model like Fable 5 can take that file and extend it meaningfully, while a base model cannot use it effectively, making the model tier and the design.md quality interdependent.

ai-native-design-workflow

Persistent project memory files (agents.md, sometimes called runbook.md) should be built incrementally by asking the AI to save decisions after the fact, not front-loaded with everything in advance.

Meng deliberately avoids stuffing agents.md with exhaustive detail up front because it gets re-read on every single prompt; instead he tells the AI mid-session to 'remember this workflow in memory,' letting the file accumulate only proven, reused patterns.

ai-native-design-workflow

Most of Design+Code's roughly 5,000 templates are now produced by generating a visual first with GPT Image 2 and converting that screenshot into HTML, rather than designing directly in a UI tool.

Meng describes GPT Image 2 as having become good enough that this screenshot-to-HTML pipeline is now his default path for landing pages, hero sections, illustrations, wireframes, and slides, replacing manual Figma-based design work.

local-first-tools

As AI compresses the cost of building software, the scarce resource shifts to accumulated personal knowledge and taste rather than raw execution speed.

Meng frames the entire episode's advice around this idea: with agents now doing the coding, designing, and even personal-assistant work, curiosity and continuously refining one's own workflow become the differentiator, since the tools themselves are increasingly available to everyone.

solo-builder-productivity

Companies

Design+Code - Meng's company and template library; he says it now holds roughly 5,000 templates, most generated from GPT Image 2 screenshots converted to HTML.
OpenAI - Maker of Codex, the coding-agent app Meng has switched to as his primary design and build tool, and of GPT-5.5, which he calls the best model he's used this year.
Cursor - Meng's previous primary tool, now converging in UI and capability with Codex; he says Composer 2.5 inside Cursor is his fastest option for quick single-project iteration.
Anthropic - Maker of Claude, referenced as one of the frontier model options (Opus 4.8, Sonnet 4.6) alongside GPT-5.5 and Fable 5.
Notion - Cited as the cloud tool Meng replaced by building his own local, markdown-based version because Notion's AI doesn't work offline.
Figma - Meng says he no longer opens Figma to create designs, only occasionally to reference historical work like old logos.
Descript - Mid-episode sponsor read for Descript's Surfaces feature, which lets designers prototype directly on top of an existing production interface.
Paper - Mid-episode sponsor read; highlighted for a new feature that copies layers, images, and SVGs directly from Figma into Paper's canvas.
Remotion - Named as a plugin/integration Meng uses for making videos as part of his Codex-based workflow.
Hyperframe - Named as another video-related tool integrated into his workflow.
Granola - Named as an example third-party plugin/integration users might install depending on their workflow.
Mr. Beast - Referenced anecdotally: Meng recounts Mr. Beast saying his accumulated knowledge let him launch a faceless YouTube channel to 20 million subscribers in 6 months, which Meng uses as an analogy for how knowledge (not just tools) drives results with Codex.

Techniques and frameworks

Design.md / style-guide-as-file - A markdown file encoding a designer's full style guide (colors, typography, spacing, corner radii, tokens) that the AI reads and applies, replacing years of internalized design taste for a new designer using the tool.
Agents.md / runbook.md as persistent project memory - A project-level file the AI reads on every prompt or new task; Meng deliberately keeps it lean, asking the AI to save specific workflow or style decisions into it after the fact rather than front-loading it.
Screenshot-first prompting - Meng starts nearly every fix or iteration with a screenshot (via a command-command shortcut in Codex) rather than typing a text description, because it is faster and more precise than describing what's wrong.
Point-and-annotate UI feedback - Instead of screenshotting a whole screen and explaining which part is broken, Meng annotates the specific element directly and asks the AI to fix just that.
Video-to-prompt for animation and interaction design - Meng's team feeds Codex a video of a desired animation or interaction and asks the AI to write the prompt describing it, sidestepping the need to know animation/motion terminology.
Reference-based prompting across a personal archive - About half of Meng's prompts reference prior work: a folder of an old app, a page he already built, a button in another screen, telling the AI to build a new element in that established style.
Screenshot-to-HTML template generation via GPT Image 2 - Meng's team generates a rough visual with GPT Image 2 first, then converts that image into working HTML, which he says now produces most of Design+Code's roughly 5,000 templates.
Effort-level model selection matched to task - Meng recommends starting on a medium-effort model tier to conserve usage limits and cost, reserving high/extra-high effort tiers (e.g., GPT-5.5 extra high) for tasks that actually need a more powerful model, since a simple landing page shows little quality difference across tiers.

Summary

Meng To, founder of Design+Code, returns to Dive Club roughly 16-18 months after his first appearance to describe a near-total shift in how he builds software: he now writes 0% of his own code, down from about 95% self-written when the two last spoke. The change tracks his move from Cursor and Sonnet 3.5 to OpenAI's Codex, which he credits with pioneering a thread-based way of working that he says Cursor has since converged toward stylistically, even as he considers Codex's underlying maturity greater at this point. He no longer opens Figma to create designs at all, reserving it only for occasional reference to old work like logos, and describes replacing cloud tools like Notion with his own local, markdown-based system built entirely through Codex, because cloud AI tools like Notion's don't work offline.

Much of the conversation is a tour of concrete prompting tactics. Meng starts nearly every fix with a screenshot rather than typed text, using a command-command shortcut in Codex to capture and send the active window instantly; he argues screenshots and voice are both faster and more precise than typing, since people speak three to five times faster and naturally add detail they'd otherwise skip. He also describes annotating a specific UI element directly rather than screenshotting a whole screen and describing where the problem is, and a newer trick his team uses for animation: feeding Codex a video of a desired motion effect and asking the model to write the prompt itself, sidestepping the need to already know animation terminology.

A recurring throughline is defending against "AI slop." Meng argues that whatever visual style becomes common - purple gradients five years ago, certain fonts and animations today - quickly becomes a marker of generic AI output, so designers have to treat taste as a moving target, always a step ahead of the baseline. His primary defense is grounding every prompt in his own accumulated material: past Figma files, fonts, code, and previous projects, fed to the AI as a starting point specifically because that material is inherently unique to him and therefore can't default into slop. He extends this into two file conventions: design.md, a markdown style guide encoding colors, typography, spacing, and design-system tokens (an AI-era equivalent of the style guides designers used to build over a career), and agents.md/runbook.md, a persistent per-project memory file that he deliberately keeps lean, asking the AI to save proven workflow decisions into it after the fact rather than loading it up front, since it gets re-read on every single prompt.

Meng also makes the case that frontier, expensive models like Fable 5 only pay off with large, context-rich prompts; a simple landing-page request produces nearly identical output whether run on a top-tier or mid-tier model, so the model's extra power is wasted unless the prompt bundles a full design.md, reference screenshots, and detailed requirements. On image generation, he says GPT Image 2 has become good enough that most of Design+Code's roughly 5,000 templates are now produced via a screenshot-to-HTML pipeline: generate a rough visual with GPT Image 2, then convert it directly into working HTML, a path he now uses for landing pages, hero sections, illustrations, wireframes, and slides.

The episode closes on a broader claim about what becomes scarce as AI compresses the cost of execution: not raw building speed, but accumulated product knowledge and taste. Meng estimates he has used at least 10,000 prompts per project, and argues that no competitor can replicate that history in a single prompt even if they can replicate any one polished output. He frames his agents as standing in for his editors, designer, and personal assistant, and says the only limiting factor left is a designer's own curiosity and willingness to keep exploring the growing ecosystem of apps, plugins, skills, and prompt patterns.

Notable Quotes

"If you always chase the trends, you will become AI slop very soon." - Meng To

"Your starting point is you, and you are definitely not AI slop because you are unique as a human." - Meng To

"I can guarantee you that nobody can make Dream Cut in one prompt just because I went through so many iterations and so many features." - Meng To

"The best people are the editors, right? So I also think myself as more of an editor than as someone who likes to scratch from scratch." - Meng To

"Anything you that you can think of, your agent can do it. And the only thing that's stopping you is you opening to this world of apps and plugins and libraries and prompts and design.md." - Meng To