Pablo Stanley - Protecting your creativity in an AI world

2026-07-21 - 52 min - source - Read full transcript

Red (host)Pablo Stanley

Key insights

Building a tool you don't personally use is a trap that's easy to fall into with AI-era tool-building.

Pablo spent months on Ecto, a 3D shader-based design tool, before realizing he wasn't dogfooding it - his actual creative work had already moved to hand-drawn sketches in TLDraw and terminal-based agent coding. He frames the ideal as needing to 'use a factory to build a factory.'

ai-native-tool-design

AI tools work best when built on the substrate the model already understands, not the one that's visually elegant.

Pablo spent a solid month and a half forcing an LLM to manipulate a 3D canvas with shaders and flexbox-style layout, which kept breaking because the underlying math was too complex for the model to reason about. Rebuilding the same tool with plain HTML and Tailwind classes - the LLM's native language - made everything work instantly.

ai-native-tool-design

Knowing when to abandon a failing approach matters more than 'pushing through.'

Pablo pushes back on the common startup advice to always keep pushing when things get hard. After a month and a half stuck on the 3D canvas approach, he had to consciously decide 'this is dumb' and restart from scratch, which he says is a harder and more valuable skill than persistence.

ai-native-tool-design

Solo agent-driven work is technically collaborative but socially isolating.

At Vercel, designers and engineers each work alone with agents in their own terminal 'black void,' then only surface finished PRs to Slack for review. The feedback loop happens after the fact with humans, not with the agent inside the team conversation - which motivated Pablo to build his own tool bringing agents into shared chat channels.

ai-collaboration-workflow

Directing/orchestrating agents feels less personally 'creative' than hands-on making, and that gap is an open identity crisis for experienced makers.

Pablo describes genuine uncertainty about whether AI-orchestrated output is 'even mine' when he isn't physically drawing the rectangle or writing the code himself. He contrasts an older hand-crafted webpage he's proud of against a newer AI-directed one that 'feels robotic' by comparison, despite similar effort.

orchestration-vs-craft

Master artists have delegated physical execution for centuries; the artist's judgment and signature, not the manual labor, was always the creative act.

Pablo cites Michelangelo, Rembrandt, and Diego Rivera, who directed workshops of painters and muralists rather than executing every stroke themselves, and Andy Warhol, who openly named his studio 'the Factory' and signed work he didn't physically make. He argues AI orchestration is a continuation of this pattern, not a break from it.

historical-precedent-delegated-art

Sol LeWitt's written wall-drawing instructions make him 'the original prompter,' decades before AI.

LeWitt gave museums literal text instructions (e.g., draw random points on a wall and connect them with lines, or pay draftsmen a set hourly rate to fill a grid) and called the resulting, museum-executed work his art - even after his death, museums still install pieces from his written instructions.

historical-precedent-delegated-art

The junior-designer mentorship pipeline is quietly being redirected from humans to models.

Historically, senior artists like Rembrandt and Diego Rivera transferred craft judgment to apprentices who executed under guidance, and those apprentices carried that knowledge forward into their own careers. Pablo argues that knowledge transfer is now going into correcting the LLM instead of mentoring junior designers, though he holds out hope AI itself could become that mentor for newcomers who bring curiosity rather than pure delegation.

orchestration-vs-craft

AI should amplify the spark of an idea, not replace the thinking that produces it.

Pablo's stated core belief, repeated at both the start and end of the episode: the danger is delegating too much thinking to AI because it 'feels like an addiction, because it's so good.' The right use is letting it expand and magnify an existing idea or curiosity rather than originate it.

ai-creative-identity

Pablo deliberately practices a skill without AI assistance to protect against tool-dependence atrophy.

He forced himself to hand-make roughly 10,000 pixel-art characters without AI, noting that at every step his brain wanted to redirect the project into 'build a pixel-art tool with AI' instead of doing the art itself - a concrete personal countermeasure to the pull toward full delegation.

ai-creative-identity

Prototyping has moved from static design files into large, often-unshipped draft PRs against the real codebase.

Pablo and a colleague wrote more code into GitHub than many software engineers on the v0 team while exploring dozens of interface directions, with most PRs left in draft and never shipped. He argues this surfaces real edge cases (performance, dependencies, actual data) that static Figma mockups can't.

ai-collaboration-workflow

New creative technologies have always faced 'this isn't real' criticism that later fades.

Red and Pablo draw a parallel to 1980s hip-hop and electronic music, which critics dismissed as not 'real' music because it used samples and synthesizers rather than analog instruments - a skepticism they see mirrored in current debates about whether AI-assisted design and art counts as authentic creative work.

historical-precedent-delegated-art

Media referenced

It's Always Sunny in Philadelphia - show - Red compares Pablo's old Loom-video habit of walking people through a hand-drawn conspiracy-style map to Charlie's conspiracy board.
Sol LeWitt's wall drawing instructions - other - Pablo cites LeWitt as 'the original prompter' - LeWitt wrote text instructions for museum staff to execute artworks he never physically touched, decades before AI prompting.
Andy Warhol's The Factory - other - Pablo invokes Warhol's studio, where assistants produced work Warhol signed, as an early precedent for delegated creative execution.

Companies

Vercel - Pablo's employer; he is a staff designer working on v0.
v0 (Vzero) - Vercel's AI product-building tool; Pablo shipped its design mode and annotations feature and describes prototyping directly against its production codebase.
Figma - Reference point Pablo repeatedly contrasts his tools against; he built agent-collaboration features into his own tool before Figma shipped similar ones.
Framer - Episode sponsor; Pablo demos Framer's new AI agents for building production sites with CMS management.
Paper - Episode sponsor; a design canvas tool that now supports pasting layers directly from Figma.
TLDraw - The hand-drawing/whiteboard tool Pablo actually uses day to day for sketches, comics, and rough design direction instead of a traditional canvas tool.
Whimsical - Diagramming/wireframe tool Pablo and Red nostalgically discuss as a possible comeback candidate for pairing rough wireframes with coding agents.
Balsamiq - Older low-fidelity wireframing tool Red recalls using heavily before AI coding agents existed.
Jitter - Animation tool Pablo cites as inspiration for editing shape-based animations rather than raw keyframes.
GitHub - Where Pablo's agent-driven PRs land; his new collaboration tool connects directly to his repo.
Claude Code - Terminal agent Pablo runs with the 'dangerously skip permissions' flag as his default daily driver.
Ghostty - The terminal app Pablo uses to run his coding agents.

Techniques and frameworks

Dogfooding your own tool - Pablo realized he had spent months building a design tool (Ecto) he personally never used, since his real work had moved to hand-sketching and terminal agents.
Matching the tool substrate to what the LLM understands - After 1.5 months fighting an LLM to manipulate a 3D/shader canvas, Pablo rebuilt the tool on plain HTML/Tailwind, the LLM's native language, and it immediately worked.
Drawing instructions as prompting - Sol LeWitt's written execution instructions for museum installers, cited as a decades-old precedent for prompt-based creation.
Asking the agent to ask you questions - Pablo prompts agents to interrogate his own ideas back to him during ideation, surfacing gaps he hadn't considered.
Prototyping directly against production code - Instead of Figma files, the v0 design team ships large draft PRs straight into the real Vercel codebase to surface edge cases static mockups can't.
Guideline-not-gospel disclaimer prompting - Pablo repeats a stock instruction - don't copy the exact sketch, prioritize the real design system - roughly 500 times when handing rough wireframes to an agent.

Summary

Red opens Dive Club's conversation with Pablo Stanley, a staff designer at Vercel working on v0, by having him demo Ecto, a self-built 3D/shader-based design tool with an embedded coding agent. Pablo walks through why he built it (frustration that no canvas tool let him collaborate with AI agents directly) and how badly the underlying approach fought him: he spent a month and a half trying to get an LLM to reason about flexbox-style layout inside an actual 3D canvas space, which kept breaking because the math was too complex for the model to track. Rebuilding the same tool on plain HTML and Tailwind classes - the LLM's native vocabulary - fixed everything almost instantly. He frames the lesson less as a tooling detail and more as a discipline: sometimes the right move is not to keep pushing through a hard problem, but to recognize a wrong-direction rabbit hole and restart.

The deeper realization Pablo surfaces is that he had built a tool he wasn't dogfooding. His actual daily creative work had already shifted away from a traditional canvas toward hand-sketching in TLDraw and terminal-based agent coding with Claude Code (run with permissions skipped so it never stops to ask questions). That shift exposed a second problem: agent-driven work at Vercel is technically collaborative but socially isolating - designers and engineers each disappear into their own terminal "black void" with an agent, then only surface a finished PR to Slack for review, with no visibility into the process or into each other's in-progress thinking. This gap is what motivated Pablo's newest and most recent project: a Slack/Discord-style tool where agents participate directly in shared team channels rather than being invisible until a PR lands.

The conversation's emotional center is Pablo's account of an ongoing identity crisis about what counts as "his" creative work when so much of it is now directed rather than made by hand. He contrasts an older, fully hand-crafted webpage he's proud of against a newer AI-directed piece that, despite similar effort, "feels robotic" in comparison because he was delegating decisions rather than making them. He resolves this partly through historical precedent: Michelangelo, Rembrandt, and Diego Rivera directed workshops of painters rather than executing every brushstroke, and Andy Warhol openly called his studio "the Factory," signing work he didn't physically make. Pablo goes further back to Sol LeWitt, who wrote literal text instructions for museum staff to execute wall drawings he never touched - calling LeWitt "the original prompter," decades before AI. The throughline across all these examples is that the artist's judgment and signature, not the manual labor, was always the creative act - a framing Pablo uses to make peace with orchestration-heavy AI work, while still holding that human friction and pushback (which LLMs are bad at providing) remain essential to producing genuinely original output.

Pablo and Red also dig into concrete workflow details: how Pablo uses TLDraw sketches and rough wireframes as loose creative direction rather than literal specs (repeating a "prioritize the design system, don't copy the exact canvas" disclaimer to agents an estimated 500 times), how he prompts agents to interrogate his own half-formed ideas back to him during ideation, and how the v0 design team prototypes by shipping large, often-unshipped draft PRs directly against Vercel's production codebase rather than working in Figma - because testing against real data and real components surfaces edge cases static mockups can't. He notes he and a fellow designer wrote more code into GitHub during v0's design-mode development than many of the team's software engineers, even though most of that code never actually shipped.

The episode closes on Pablo's clearest statement of principle: the danger of these tools isn't using them, it's delegating so much thinking to them that using AI starts to feel like an addiction "because it's so good." As a countermeasure, he describes deliberately forcing himself to hand-make roughly 10,000 pixel-art characters without AI assistance, noting that at every step his brain kept trying to redirect the project into building an AI pixel-art tool instead of doing the art. He connects this to Marshall McLuhan's observation that we shape our tools and then our tools shape us, and to Socrates' historical objection to writing eroding memory - arguing that fear of new tools eroding capability is not new, and that the resolution isn't avoidance but discipline: using AI to expand and amplify an existing spark of curiosity rather than replace the thinking that produces it in the first place. Red draws a parallel between current AI skepticism and the "this isn't real music" criticism once leveled at sampling and synthesizers in hip-hop and electronic music.

Notable Quotes

"There is a danger in using these tools, I think, because if we delegate too much of it, we're not forcing ourselves to think." - Pablo Stanley

"Ideally I should have been using a factory to build a factory. I was not doing it. I was just straight up working with the agent." - Pablo Stanley

"Andy Warhol wouldn't even hide it. He would just call it the factory. It's a factory and I'm not making it, I'm just signing it at the end." - Pablo Stanley

"There has to be a moment where you say, no, this is dumb. The usual commentary is you have to keep pushing whenever things get hard - no, no, no." - Pablo Stanley

"I use this terminal Ghostty. I use Claude and then dangerously skip permissions. I'm like, don't ask me questions, just let you cook." - Pablo Stanley