AI Engineer

One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer

1680 summary words 7 min summary Watch video

Start with the signal

7 min read

Summary

At-a-Glance

  • Verdict: Watch fully
  • Core thesis: A single designer can reliably produce hundreds of conference assets by treating design as a reusable system, using AI agents to generate and update outputs, and reserving human attention for exceptions and experience-level judgment.
  • Why it matters: This is a concrete example of an AI-enabled operating model: structured inputs, reusable primitives, agent execution, visual QA, and fast exception handling replace a slow designer-to-engineer handoff loop.
  • Best use: Use it as a lightweight case study for designing agent-assisted creative and operations workflows, especially where structured data must become accurate, branded, production-ready outputs.

Executive Summary

Vincent Wendy describes how he supported AI Engineer's conference design operation as effectively one designer serving a 7,000-attendee event with 140+ sponsors, 300+ speakers, 600+ sessions, and hundreds of physical and digital deliverables. His central point is not that AI creates good design autonomously; it is that a well-defined design foundation lets an AI agent produce, adapt, and validate outputs at operational scale.

The workflow begins with a design system: typography, colors, components, taglines, and atomic-design-style building blocks. These constraints matter because otherwise LLMs improvise details such as font sizes and create inconsistent output. Once the website and core brand system are established, the same rules can be reused for email, flyers, speaker graphics, signage, and themed assets.

Wendy uses Devin in Slack, connected where possible to Figma/MCP context and supplemented by annotated spec sheets, to turn structured event data into pixel-accurate schedules and speaker-announcement assets. This removes the traditional friction between design intent and engineering implementation: instead of passing a static Figma file through an engineer and multiple feedback rounds, he asks the agent to refine an implementation against explicit visual specifications.

The most transferable lesson is operational rather than aesthetic. Automation handles repeatable production and checks, including speaker-photo identification and sponsor-logo comparison, while the designer focuses on anticipating failure modes, walking the attendee journey, and resolving exceptions such as schedule changes. The speaker frames scale as a prompt to think smaller: enumerate the details that can fail, structure them, and make each one editable and automatable.

Key Takeaways

  • Claim: AI enables a one-person creative operation only when the work is systematized into reusable design constraints rather than generated from scratch asset by asset. | Evidence: Wendy supported a conference with 7,000 attendees, 140+ sponsors, 300+ speakers, 600+ sessions, and deliverables spanning landing pages, signage, digital screens, stickers, speaker graphics, and sponsor materials; his working "design team" was himself, Devin, GPT, and Figma. | Implication: For high-volume output systems, invest first in reusable primitives and data schemas; agent productivity compounds only after this foundation exists. | Caveat: The scale claim depends on substantial prior setup of the brand system and structured inputs; the talk does not show that raw AI generation alone can maintain quality.
  • Claim: A strict design foundation is the control layer that prevents LLMs from producing visually inconsistent outputs. | Evidence: Wendy explicitly defines desktop and mobile typography, colors, and other brand rules because Claude and other LLMs otherwise choose arbitrary font sizes. He applies atomic design principles—small components combined like Lego pieces—to create complete deliverables. | Implication: Treat design tokens, component definitions, and rendering specifications as executable constraints supplied to agents, not merely as documentation for humans.
  • Claim: The practical production pattern is to give an agent structured event data plus visual specifications, then generate production outputs directly rather than manually recreating designs in Figma. | Evidence: For room schedules, Wendy asks Devin to pull the latest data, render a requested room schedule, export it to PNG, and place it on venue screens. For speaker announcements, a tool generates personalized portrait and landscape graphics, trading cards, and theme variants from speaker details and headshots. | Implication: Conference, marketing, and internal-ops teams can convert recurring visual assets into parameterized generation tools instead of treating each new asset as a bespoke design task. | Caveat: The transcript asserts that outputs are accurate and pixel-perfect but does not provide error rates, implementation details, or a comparison with a conventional coded template system.
  • Claim: Agent access to design context can compress the designer-engineer feedback loop that normally causes visual drift. | Evidence: Wendy describes the old workflow as design handoff followed by non-pixel-perfect engineering implementation and repeated feedback. His newer loop is Slack to Figma and back to Slack, where Devin can use MCP-connected context or a Figma-generated spec sheet containing spacing, font sizes, colors, and annotations. | Implication: The effective control plane is not a chat prompt alone: agents need access to source design context plus a machine-readable or clearly annotated visual spec. | Caveat: He notes that Figma layer names are often poor or generic, so implicit canvas structure alone is not a dependable interface; explicit specifications remain important.
  • Claim: Visual AI can serve as a second QA layer for high-cost omissions and matching tasks. | Evidence: Wendy says he asked Devin to compare a sponsor-logo graphic against the sponsor list and found 100% accuracy in his testing. He also used an image-matching workflow to identify a speaker, Jason Liu, from photographer photos so the correct image could be placed into a thumbnail. | Implication: Use multimodal agents as systematic checkers for completeness, identity matching, and asset reconciliation, particularly where humans are prone to overlook small items across large grids or lists. | Caveat: The 100% result is an informal, unspecified test rather than a robust benchmark; visual-agent outputs should remain subject to human review when errors have contractual, reputational, or identity consequences.
  • Claim: After automation, the designer's highest-value work shifts to user-journey reasoning and exception handling. | Evidence: Wendy walks the venue mentally as an attendee—from registration through wayfinding, QR codes, and room discovery—to ensure the system is connected. When a schedule changed after launch and the original tool lacked editing capability, he asked Devin to add an edit button, updated the schedule, exported a new PNG, and redeployed it to the screen. | Implication: Build production tools around change management from the outset: identify likely exceptions, expose edit paths, and optimize the path from corrected data to redeployed artifact.

Detailed Brief

Design around the actual problem, not current model limits

  • Claims: Wendy argues that having a concrete operational problem is now an advantage because available AI tools can be assembled around it.; He encourages designers to work around model limitations rather than waiting for a model to perfectly execute a requested artifact format.
  • Evidence: Referencing a Simon Willison exercise, he notes that basic models still struggled to generate a usable vector file of a pelican riding a bicycle.; His workaround was to request a static PNG instead and vectorize it in Figma, making the result usable despite the model's failure at direct vector generation.
  • Caveats: The workaround still requires human design judgment and tool fluency; it is not evidence that generated graphics meet a professional standard without editing.
  • Implications: Specify the business outcome separately from the requested intermediate format; if an agent fails at one modality, route the work through an alternate format and downstream tool.

The speaker's five-part operating model

  • Claims: Wendy summarizes the system as: foundation first, reusable design, automated workflows, validated output, and friction removal.; His final heuristic is to "think small": enumerate the individual details and foreseeable failure modes inside an apparently large scale problem.
  • Evidence: He characterizes "1,000 details" as "1,000 ways to fail," citing examples such as a missing sponsor logo or incorrect speaker schedule.; The transcript's closing section is repeated almost verbatim, reinforcing that the intended conclusion is prevention and exception readiness rather than generalized AI enthusiasm.
  • Caveats: The talk is a practitioner conference anecdote, not a controlled evaluation of Devin, GPT, Figma, or MCP.; The final portion of the supplied transcript is duplicated.
  • Implications: When selecting AI automation candidates, prioritize processes with repetitive structured outputs, objectively checkable requirements, and expensive manual coordination.

Notable Concepts & Terms

  • Atomic design: A component-based design approach in which small reusable elements are composed into larger interfaces and deliverables; Wendy uses it to make branding reproducible across many asset types.
  • Design system / design tokens: The explicit typography, colors, components, and visual rules that constrain agent output and keep generated assets on-brand.
  • Devin: The AI coding agent Wendy uses as the production and iteration layer for schedules, asset generators, visual checks, and late-breaking tooling changes.
  • MCP: The mechanism Wendy references for connecting the agent to relevant design/tool context, reducing ambiguity between Figma intent and implementation.
  • Figma spec sheet: An annotated specification artifact listing spacing, typography, colors, and related properties, used to make pixel-accurate implementation more reliable.
  • Visual QA: Using multimodal AI as a second checker for omissions and mismatches, such as missing sponsor logos or identifying the right speaker image.
  • Exception handling: The remaining high-value human-and-agent work after repeatable production is automated—especially handling schedule changes, missing edit paths, and other deviations from the planned flow.

Operator Notes / Why Ken Should Care

  • Pilot a parameterized asset-generation workflow for one recurring output class—such as event schedules, speaker cards, sales one-pagers, or partner-logo boards—with structured source data and direct export formats.
  • Require agents producing visual interfaces to consume explicit tokens/specifications for typography, spacing, colors, and breakpoints; do not rely on screenshots or generic prompts alone.
  • Add an independent visual reconciliation step for high-stakes asset lists, but retain human sign-off for logos, identities, and externally visible commitments.
  • Audit recurring creative workflows for exception paths before launch: identify likely late edits, make the underlying data editable, and test redeployment from a changed record to the final channel.
  • Evaluate whether MCP-connected design context plus generated annotated specs can replace portions of current design-to-engineering handoff and revision cycles.

Source/Metadata

  • Title: One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer
  • Transcript words: 2514
  • Duration seconds: 1008
  • Timestamp note: No timestamps or chapters were present in the supplied transcript; the final discussion segment is duplicated.
Search this video Resume transcript at 0%
Full transcript 2226 words · 11 min read
0:14

Hello everyone. Hope you guys are having a good time at the conference. So before we start, how many of you are actually designers? A product designer? Hey, one hand and other. Okay. And how many, I assume that the rest of you are engineers? Is that correct? Yeah, pretty much. Okay. So today's talk is a non-technical talk, but more of a real-world experience: how I created the design for AI Engineer this conference and our other past conferences as well, and how AI has helped me. And so the talk today is "One Designer Plus AI," which is me as the designer, and hundreds of deliverables. All right, let's start. So my name is Vincent Wendy.

0:50

I am a senior creative designer at AI Engineer. And at AI Engineer, it's a very small team. So we only have around 12 to 15 people at the moment. And everyone has been doing their own thing. And I think AI has been a really helpful way to help everybody doing everything. And even at the scale we have, we have a problem, obviously, right? And the problem is the scale problem, or I would call it the challenges. And how to overcome it? So it's basically automation. And we get to that in the later part of this talk. So when I prepared this talk, we only expected 6,000 attendees, and now it's 7,000. Well, good for us.

1:42

And then we have 140 sponsors, more, 140 plus sponsors, and then 300 plus speakers, 600 plus sessions, and one designer. And everybody needs design, right? Every single thing needs design. Sponsors need assets, speakers need graphics, you need signage so you don't get lost. And this is basically what we do, what I do. So from stickers, do you like your swag, your stickers? Well, I hope you do, because I created that design too. So I'm going to go to a landing page, speaker announcement, track mascot, all the stuff that you see, most of the stuff that you see here, from signage to digital signage, landing page, everything is a deliverable.

2:17

And 1,000 details means 1,000 ways to fail, right? Because a missing sponsor logo is going to be a huge issue. And speakers that have a wrong schedule also issues, right? And it seems impossible to handle that many deliverables, but yeah, meet my design team. And I'm going to be a little bit more specific. So it's me, and Devin, GPT, and Figma. And right now we are at the stage where tools isn't the problem anymore, but having a real problem is our advantage. So for example, when someone asked me, what inspired you when designing the AI Engineer? I don't know the answer back then, but after I think about it, it's a problem that inspired me to design the AI Engineer.

3:04

And we'll get to that in the later part of this talk. So have you guys seen the talk by Simon Wilson in 2025? Yeah. Yeah, and it's pretty interesting, right?

3:26

So I think he asked every LLM to create a vector file, which is basically a pelican riding a bicycle. And it is basically to test, and I tested it again, and it's still doing this for the basic model. And it's not usable for me as a designer. But as a designer, we have to think outside the box. And we could simply ask it to create a static image, a PNG of a pelican riding a bicycle, and then I can factorize it on Figma. And we can ship that now. So we have to think outside the box here, regardless of the capabilities of the LLM. And so how to solve this scale problem, right? Basically, five things.

4:03

So foundation first, reasonable designs, automated workflows, validated output, and also remove friction. So the foundation is definitely the core part that we need to set up right. The design system, typography, colors, components, and other stuff. And once this is set up, for example, when we create the website, it's all set up within this thing. And yeah, this is just an example. So we have the colors, primary, and then also the accent colors, the typography, and also the tagline, all the other stuff. And also, have you guys heard of atomic design? So yeah, my previous background is product design.

4:38

So I'm pretty familiar with creating user-centric design and atomic design, right, where we create the smallest part possible and then combine it into basically Lego pieces and then into deliverables. And this is pretty useful in my job right now.

4:52

So once we set up all of those foundations, we basically need to create. For example, we use Devin a lot. At the office, everybody uses Devin. Everybody is using Devin, for example. Yeah. And so in this case, I just need to say, hey, we use this desktop typography and this mobile typography. Because we know Claude or any other LLMs love to throw some random font size, right? And if we don't define it, it just delivers a mess. And yeah, typography, color and stuff. And then it comes to reusable design. So once we set it up right, the website has the branding to it, all the other themes on the AI Engineer, for example the marketing themes, can create everything basically.

5:33

They can create an email design based on that. They can create a flyer, a document, just based on the website because it's already defined early. And yeah, once you get the design, you can rinse and repeat. For example, the mascot, it all has pretty much the same design and it's rinse and repeat. And if you already define those things, you can basically create one design that works for all. And this is the part that I'm most interested to talk about: automated workflows. Before, for example, if you take a look outside the room, there's a schedule, right? The schedule for each and everyone. So we used to do it manually on Figma, but now we use Devin for it.

6:13

And let me show you. Hey. So right now we just pull the latest data. I just ask Devin, hey, I want this room at this base. And then we can export it, download it to PNG, and the data is accurate. And then we can ship it to the flash drive and put it on the screen. And it was impossible before because the friction is just too much between the designers and the developers. We cannot make things pixel perfect because once we tell the designer, hey, this is the design. And then the engineers that created the design, for example, hey, I need this to be delivered. And then they don't create it pixel perfect. It's a lot of feedback, right?

7:07

But with Devin, we just say, hey, can you make this more accurate? We can connect it to MCP. And then if it doesn't work, we can give a spec sheet or something that can be defined like what's the spacing, what's the font size, et cetera. And this is what we do for the speaker announcement. So we have 300 plus speakers. And it's impossible for me to handle one by one, right? So we create this thing, which you can also access: speaker announcement. And you can also try it yourself, this one, for example. You can select it right here. And then you can also change your name. Well, yeah. For example, you can change the name to whatever you want.

8:08

And we also have the landscape mode, which can also be loaded. If the speaker also has the headshot and all the details, it will automatically export. And we also have the trading cards, which is surprisingly pretty popular. And we have different themes. And this is all pixel perfect. For example, this one. This is inspired by TBPN. So yeah. And how do I deliver this pixel perfect? Let's jump into it. So the process here is before when I started my career as a product designer, it used to be research. We need to build product, design thinking in general, right?

9:16

And then feedback loop and stuff. But right now, it's outdated for me. In my case, we just go to Slack, Figma, and then send it back to Slack, because our Devin lives in Slack. And then ship all the things that he needs. For example, if we can connect the MCP or also the spec document, which is for example, the spec sheet, which is a plugin in Figma, if you're interested. It's free. And it basically gives annotation to the PDF. And yeah. All designers don't name their layers. So yeah. This is just random frame three, frame four. But the LLM will get it. And it's basically defining all the spacing, all the font size, and all the colors and stuff.

10:07

It's definitely going to help you develop a pixel perfect product. And we also have just recently, today, have photos, which we have to create the thumbnail for speakers, right? And then we asked Devin, hey, who is this person? And yeah, it kind of did. I made a Tinder detection if this is the same person or not. And I think it's pretty accurate.

10:26

It's Jason Liu. Yes. And then we can use this for context. Before when we create the thumbnail, we have to search all the codes that the photographer has and search one by one and maybe by time if possible. But now we can just, oh, this is Jason Liu. We have to download that photo and then we can paste it into the thumbnail, right? And it's pretty amazing. I mean, the world that we live in right now is actually at the peak for me as a designer because what else can you ask for, right? I mean, we already have things to automate. We already have things to create design fast.

11:08

Basically, all you need is a problem because once you have a problem worth solving, you can basically solve anything. And back to my talk. I got sidetracked. Yeah. And then, yeah. And this is also the amazing thing that we test. So as you know, we have hundreds of sponsors, right? 140 plus. And as you can see in the lobby, we have the banner with all the sponsors. And I basically tell Devin, hi, could you compare, could you check if there are any missing logos in this graphic? And the accuracy is 100% based on the test that I do.

11:46

So that's pretty good. And we use the same thing for the T-shirt that you got for your swag. And yeah, surprisingly, Devin knows how to visualize things, right? How to detect things visually. And that is very surprising because as humans, we can make errors. Oh, it turns out there's one small logo that is missing. But with this, we can double check. So human plus AI combined. Well, you've got your own QA team. And then remove friction. So this is just a way of thinking. As a designer, we have to think as a user, not as a designer, right? Because every user has their needs. You can walk through the map plan here.

12:36

So basically, I'm imagining myself as an attendee going to registration, going to see the wayfinding and the QR code, and all the stuff. Basically, everything needs to be connected so you guys don't get lost and know how to find your rooms and other stuff. And the real job is handling exceptions. So for example, there is a schedule update, all right? And when we create this thing, it doesn't have an edit button. And then one morning, hey, this schedule needs to be updated. And we don't have those edit buttons. I could just ask Devin, hey, can you add me an edit button? And then it did.

13:07

So we can change everything now and ship it to PNG and re-plug it to the screen, which is pretty convenient, right? And those exceptions, it's not possible before when we have to do it manually. But now it just gets easier. And so the takeaway here is that to solve the scale problem, you have to actually think small. Think all the smallest things possible. Think everything that can go wrong and will go wrong and then try to solve it before. And also, right now basically you can automate everything. And at this moment, having a problem is actually going to benefit you because that's going to help you ship a better product, going to ship things that are good.

13:40

And yeah, I think that's all that I can share. Hope my talk has some benefits to you. And yeah, that's all. Thanks, guys. Oh, turns out there's one small sound that is missing. But with this kind of thing, we can like double check. So human plus AI, combine it. Well, you've got your own QA team. And then remove fiction. So this is just the way of thinking. So as a designer, we have to think as a user, not as a designer, right? Because every user has its needs. You can walk through the, for example, the map plan here. So basically, I'm imagining myself as an attendee to go to the registration, go to the, see the wayfinding and the QR code, and then all the stuff.

14:46

Basically, everything needs to be connected so you guys don't get lost and knows how to find your rooms and other stuff. And the real job is handling exceptions. So for example, oh, I have, yeah. For example, there is a schedule update, all right? And when we create this thing, it doesn't have an edit button. And then one morning, it just, hey, this schedule needs to be updated. And we don't have those edit buttons. I could just ask Devin, hey, can you add me an edit button? And then it did. So we can change everything now and then ship it to PNG and re-plug it to the screen, which is pretty convenient, right? And those exceptions, right, it's not possible before

15:38

when we have to do it manually and stuff. But now it just gets easier. And so the takeaway here is that to solve the scale problem, you have to actually think small. Think all the smallest things possible. Think everything that can go wrong and will go wrong and then try to solve it before. And also, like, yeah, right now basically you can automate everything.

16:07

And at this moment, having a problem is actually going to benefit you because that's going to help you ship a better product, going to ship things that are good. and yeah, I think that's all that I can share. Hope my talk has some benefits to you. And yeah, that's all. Thanks, guys.

Reading tools

Type to find a passage

Appearance
Ask this transcript

Notebook

Export Markdown

Add a note