Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc.

Read the talk

Design at the Speed of Adjectives

Paul Bakaus explains how Impeccable turns design vocabulary into contextual instructions for coding agents, while keeping people responsible for direction, iteration, and taste.

From a talk by Paul Bakaus

At a glance

Ideas worth remembering

  • A useful steering command connects an intention to specific operations and constraints. Bolder means hierarchy, scale, and decisive typography that respects the design system; harden addresses conditions such as device performance and responsiveness.

  • Context and iteration remain essential to Bakaus’s approach. Audience, emotional direction, references, and boundaries inform the choices that a plausible generated page may otherwise lack.

  • The appropriate level of control changes throughout the workflow. Contextual commands can help with crafting and iteration, while Bakaus still sees a need for human work in the final 5, 10, or perhaps 20 percent that moves a result from good to great.

  • Impeccable’s refusal of automatic mode follows its core design decision: help people steer agents and amplify their craft, while keeping direction and taste in human hands.

A small prompt, two different interpretations

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 138 seconds
A small prompt, two different interpretations

Paul Bakaus opens with Impeccable, a design skill that works through an existing coding harness. Its purpose is to improve the agent’s design work and, potentially, help the person directing it become a better designer. The presentation concerns the approach behind the tool: how to give people useful control over generated design.

The demonstration starts from an intentionally basic website. Bakaus describes it as acceptable but bland, with at least some color. He asks the agent to make its workflow section bolder with Impeccable installed. He considers the result bolder, but immediately identifies a weakness: section numbers that he regards as a recurring generated-design tell. The skill can guide a transformation without eliminating every unwanted convention.

To address the possibility that the model alone would produce the same improvement, he presents a comparison using the same project and the same prompt without Impeccable. He invites the audience to judge the difference. This is an illustrative comparison, rather than a quantified evaluation: his account establishes what he changed between the examples, but the spoken description does not establish their comparative visual quality independently. It introduces the central question of the talk—what an adjective, or a verb, actually tells an agent to do.

0:120:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:01 · section reference included

Finding control between pixels and delegation

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 272 seconds
Finding control between pixels and delegation

Bakaus places the tool in a changing relationship between design and engineering. He sees designers moving into code and engineers moving into design, with more people becoming design engineers. Users tell him that Impeccable supplies a shared design language. In his view, the conventional sequence of product definition, design, critique, handoff, and engineering implementation is beginning to collapse in both small and large companies.

He describes two extremes in the actual work. Direct manipulation means changing margins, padding, and other properties in tools such as Figma, where the person operates on the design or final output. At the other extreme, a person simply asks an agent to design or build something. The first offers detailed control; the second delegates broad execution. Bakaus’s interest lies between them: choosing the level of control and the moment at which a person should enter the loop.

This makes tool choice a question of matching the operation to the job. Bakaus argues that, for most work, adjusting padding, margins, and spacing is too low a level of abstraction. A person may know the desired change in emphasis without needing to specify every spacing value. The opportunity is to express that intention at a higher level while retaining enough control to direct the result.

2:502:52
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:50 · section reference included

Generated design changes its defaults

Broad delegation has its own problem: recurring visual defaults. Bakaus contrasts older generated pages associated with purple gradients against what he now calls Claude beige, characterized by beige palettes, Instrument Serif, and italics. His observation is that the familiar tells keep moving. Recognizing one set of patterns does not mean the underlying tendency toward sameness has disappeared.

He does not consider that newer style inherently bad. The criticism is that many outputs adopt the same appearance. A design can be visually competent yet fail to distinguish the particular product it represents. Here, the concern is uniformity across generated work, rather than the quality of any individual font or palette.

5:085:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:08 · section reference included

Design needs decisions before execution

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 435 seconds
Design needs decisions before execution

Bakaus’s first explicit position is that design cannot be completed in one shot. He presents this as his judgment about present capabilities, with uncertainty about whether it might ever become possible. Effective design needs context about the purpose, audience, and thing being built. It also needs iteration. Team members and users bring different opinions, and even someone working alone must develop the result through a process.

A missing design vocabulary makes that process frustrating, particularly when stakeholders cannot express what they want. Bakaus proposes questions that establish direction before asking for execution: what emotional territory should the design occupy, what should it never feel like, what references matter, and who is the audience? These questions supply both positive direction and boundaries. His analogy is a human design studio: even an excellent design director would need to understand the brief before going away to make the work.

Returning to a quickly generated page, he identifies the missing ingredient as deliberate decisions. In his assessment, the page may be competent, but it is empty because nobody chose what it should communicate. The limitation is substantive: producing a plausible arrangement of elements does not itself establish a purpose or a point of view.

5:586:01
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:58 · section reference included

Compressing design language into commands

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 510 seconds
Compressing design language into commands

Impeccable’s proposed control surface consists of adjectives and verbs: make something bolder or quieter, distill it, polish it, make it denser, or harden it. Bakaus defines distilling as simplification. Hardening concerns whether the design works across conditions, including poor performance on certain devices and missing responsiveness. The vocabulary therefore covers both visual direction and the practical behavior of the interface.

The motivation comes from his observation of designers and engineers attempting the same task with the same model and harness. He reports a stark difference in their outputs based on the language they use. Impeccable tries to compress that design language into a skill so that people can express intentions more effectively. The important capability is contextual interpretation: a steering word should resolve to what that change means for the project, rather than merely invoking the model’s general associations.

8:108:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:10 · section reference included

What bolder should change

Bolder illustrates why the word alone is insufficient. Without a project-specific meaning, the agent can follow its own references, inventing colors or gradients. Impeccable instead directs boldness toward hierarchy, scale, and decisive typography. Those changes are intended to raise the user’s attention while respecting the existing design system. The command describes an effect to achieve and constrains the kinds of changes used to achieve it.

Bakaus describes a file loaded when the bolder instruction is invoked. It includes a challenge to imagine showing someone the result and telling them that AI made it bolder; if that attribution seems believable, the work has failed the challenge. He says the agent is instructed to apply this test to itself and sometimes reflects that its work was poor. This supplies a self-critique heuristic aimed at avoiding obvious generated styling, though it does not establish that the agent can reliably judge how another person will perceive the result.

The broader mechanism is to imbue a familiar word with a more specific meaning. The model already associates that word with many things; the skill translates it into the domain the person cares about. Bakaus also limits the claim. This level of control cannot solve every design problem, and the appropriate level changes during the work. He leaves room for exploratory work at the beginning and human finishing work at the end. In his judgment, AI still cannot replace people for the last 5, 10, or perhaps even 20 percent needed to move a design from good to great, although it can improve the work.

9:409:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:40 · section reference included

Placing commands along a messy workflow

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 772 seconds
Placing commands along a messy workflow

After more than 20 years building tools, Bakaus says he begins by mapping the target user’s workflow. Design is nonlinear, involves different people, and passes through several kinds of work. He looks for points where a tool can help: initialization and shaping, crafting, iteration through adjectives and verbs, hardening and polishing, and finally bringing the result back into the design system while cleaning up technical and design debt. These are intervention points along a process, rather than a claim that every project follows a fixed sequence.

Overdrive shows a more playful intervention. Bakaus was unsure whether to ship this command, which tries to create something deliberately over the top and began partly as a joke. He reports that people loved it. His example is an event horizon shader that he had already built, then pushed further with overdrive. He explicitly says the example was not made in one shot. The command intensified existing work; it did not account for the whole process of creating it.

His decision process includes testing uncertain commands with the community and observing whether people find value or joy in them. But the underlying requirement remains explicit meaning. An adjective with nothing behind it is only a nicer prompt. Because the model has encountered many meanings for the same word, the instruction must explain which meaning applies to this work.

12:0512:10
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:05 · section reference included

Why steering remains the person’s job

Selected presentation frame from Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc. at 870 seconds
Why steering remains the person’s job

Bakaus closes with a firm product decision: Impeccable will have no automatic mode. He reports frequent requests for the tool to do all the work without the user choosing commands, and says he intends to close a pending pull request for that reason. Choosing the command is central to the tool’s purpose. It gives the person a way to steer toward a desired result; removing that choice would move it toward the one-shot design tool he does not intend to build.

His rationale rests on a distinction between amplifying taste and manufacturing it. He calls the former amplified craft: tools can sharpen a person’s taste and help express more of it. In his view, taste arises from context, culture, and accumulated experience. When everyone replicates the same thing, its distinctiveness becomes muddy. He acknowledges that this is a contested opinion and says his tools are not an attempt to solve taste itself.

He ends by identifying the aim as finding the right level of control for design, then thanks the audience and invites questions. The closing position preserves a role for both parties: the tool helps express and execute direction, while the person continues to choose where the design should go.

14:0914:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:09 · section reference included

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:12

    Cool. Uh, let's get this going. Welcome

  3. 0:15

    everyone. Thanks for coming. Um, my name

  4. 0:18

    is Paul and, uh, I'm going to try to

  5. 0:21

    save you from the midday slump. So,

  6. 0:24

    let's, uh, let's see if we can do that.

  7. 0:26

    Um, okay. So this talk, this cold open

  8. 0:29

    of this talk will make a lot more sense

  9. 0:31

    if you know uh what impeccable is. Uh

  10. 0:34

    but if you don't know what it is, it's a

  11. 0:35

    design, it's a design skill that turns

  12. 0:37

    your coding harness into a better

  13. 0:39

    designer and hopefully you into a better

  14. 0:40

    designer. Um and you can try it out. It

  15. 0:43

    works across all harnesses, cloth code,

  16. 0:45

    github, copilot, cursor, codex, etc. Um,

  17. 0:50

    and this talk is really about uh my

  18. 0:53

    approach to how to building this tool as

  19. 0:56

    opposed to the tool itself. Um, but

  20. 0:58

    let's get started with a pretty typical

  21. 1:00

    website. So, this is a intentionally

  22. 1:04

    very basic website and I want to give

  23. 1:06

    you that one as context to what I'm

  24. 1:08

    going to show you next. So, you know,

  25. 1:10

    it's it's a it's it's an okay website.

  26. 1:12

    It's not great. It's pretty bland. Um

  27. 1:15

    but it at least it has some color some

  28. 1:17

    something. Um so now here is um the

  29. 1:22

    command make it bolder make the workflow

  30. 1:25

    section bolder with impeccable.

  31. 1:28

    Um so this is with impeccable installed.

  32. 1:31

    So now there are it's not perfect. There

  33. 1:35

    are some problems with this. for

  34. 1:36

    example, the um the section numbers that

  35. 1:39

    you have there, that's a very typical AI

  36. 1:41

    slop tail uh that uh that GPT loves to

  37. 1:44

    do and I mean Claude actually loves to

  38. 1:46

    do as well and

  39. 1:48

    it sometimes creeps in. So I wouldn't

  40. 1:50

    say this is perfect by any means but I

  41. 1:52

    think it's I would say bolder than

  42. 1:54

    before. Um

  43. 1:56

    now a lot of people install impeccable

  44. 1:59

    try it out uh and then they ask me well

  45. 2:02

    I don't know like maybe it could be

  46. 2:03

    placebo maybe the model is good enough

  47. 2:04

    in itself right uh not sure if it

  48. 2:06

    actually works well let's compare it to

  49. 2:10

    this is exactly the same project GPD55

  50. 2:12

    on extra high um same exact prompt

  51. 2:16

    without impeccable

  52. 2:20

    so there is a slight difference I would

  53. 2:22

    say but you be the judge of

  54. 2:25

    Um, so the before and after is pretty

  55. 2:28

    different and I'm going to talk about

  56. 2:31

    why that is and um and and what the

  57. 2:34

    approach is.

  58. 2:36

    So designing at the speed of adjectives

  59. 2:39

    and um to get the elephant out of the

  60. 2:41

    room, they are not just adjectives,

  61. 2:43

    they're also verbs. Uh but some of them

  62. 2:46

    are adjectives.

  63. 2:50

    First off, the role of the engineer and

  64. 2:52

    the design are blurring. Um, I've seen

  65. 2:54

    more people become design engineers than

  66. 2:57

    ever. Designers are moving into code,

  67. 2:59

    engineers are moving into design and

  68. 3:01

    vice versa. Some of the best comments I

  69. 3:03

    get from people who walk up to me and

  70. 3:05

    say like, well, impeccable gives me this

  71. 3:07

    shared language of design that everybody

  72. 3:09

    can communicate now in. But really,

  73. 3:12

    these worlds are all colliding in my

  74. 3:14

    opinion and the typical handoff is kind

  75. 3:17

    of broken. like the the waterfall

  76. 3:19

    process of I start with design, maybe I

  77. 3:22

    start with like the what on the PM side,

  78. 3:24

    then it goes into design and then

  79. 3:25

    there's like this handoff and this

  80. 3:27

    critique process and then ultimately the

  81. 3:29

    engineer builds it. That thing is just

  82. 3:31

    not real anymore. I don't know about

  83. 3:32

    you, but um I think that's starting to

  84. 3:35

    collapse very quickly in small and large

  85. 3:38

    companies.

  86. 3:41

    And when it comes to the actual work of

  87. 3:43

    design,

  88. 3:45

    I feel like there's right now kind of

  89. 3:46

    like two worlds that you can play in.

  90. 3:48

    Either it's direct manipulation sort of

  91. 3:51

    like in the pixel space, right? So

  92. 3:53

    you're in Figma and you're manipulating

  93. 3:55

    margins, paddings, whatever it is. Um or

  94. 3:58

    you're in web flow or you're in some

  95. 3:59

    other tool where you directly

  96. 4:01

    manipulating the final output uh or the

  97. 4:04

    design. Now that's fine, but then you

  98. 4:07

    feel like you're getting behind because

  99. 4:08

    everybody else is like agentic and going

  100. 4:11

    crazy with loops and loop maxing, right?

  101. 4:13

    And now you're going on the other

  102. 4:14

    extreme, which is simply telling your

  103. 4:16

    agent, please design this for me or

  104. 4:18

    please build this. But it feels like to

  105. 4:20

    me at least there's this there's this

  106. 4:22

    middle ground that we haven't explored

  107. 4:24

    enough like what is the exact level of

  108. 4:26

    control? How can we insert the human in

  109. 4:28

    the loop at the exact right time?

  110. 4:32

    Uh and so this this came to mind then

  111. 4:35

    it's pretty popular tweet a while ago.

  112. 4:37

    Um so centering a diff with opus. Yes.

  113. 4:42

    Uh not a great use of opus. Uh but you

  114. 4:46

    know it kind of speaks to that problem.

  115. 4:48

    It's like okay well when do we use which

  116. 4:51

    tool? When is which tool adequate to do

  117. 4:53

    what job?

  118. 4:56

    In my opinion right now for most work

  119. 5:00

    the direct manipulation of padding,

  120. 5:02

    margins, spacing and so on is actually

  121. 5:04

    at a elevation that's too low.

  122. 5:08

    And on the other hand um if you go fully

  123. 5:12

    agentic you get web pages like this now

  124. 5:15

    that you would you would get this in

  125. 5:16

    like 2022 I would say. So everybody here

  126. 5:19

    can point out the typical AI slot

  127. 5:20

    patterns. Um but the purple gradients

  128. 5:23

    were actually way beyond that. there are

  129. 5:25

    no purple gradients anymore uh in most

  130. 5:27

    frontier models. Um now we just get this

  131. 5:31

    and it's kind of wears a different suit

  132. 5:32

    as Claude would say. Um but uh but it

  133. 5:36

    really is slop as well. It's just a

  134. 5:38

    moving target. So now we have I call

  135. 5:40

    this claude beige. We have these

  136. 5:42

    instrument serif fonts. We have italics.

  137. 5:44

    And it's not necessarily a bad design.

  138. 5:46

    It's just everything that comes out of

  139. 5:48

    claw design looks that way. And when

  140. 5:51

    everything looks the same, you kind of

  141. 5:52

    get algorithmic unilo as I call it. Uh

  142. 5:56

    so also not great.

  143. 5:58

    Now here's my first hot take of today's

  144. 6:01

    session. You cannot oneshot design.

  145. 6:05

    Um now many people want to and maybe

  146. 6:07

    we'll get there eventually. I just don't

  147. 6:09

    think it's possible at the current

  148. 6:11

    moment. Uh maybe not ever. Um, I think

  149. 6:14

    in order to build something that really

  150. 6:17

    looks and feels great and is effective

  151. 6:19

    design, you need to have something

  152. 6:22

    that's context context rich which has

  153. 6:24

    been informed by what you're trying to

  154. 6:26

    do, what your audience is, what you're

  155. 6:28

    trying to build. And also it will have

  156. 6:30

    to be multi-shot. You will have to

  157. 6:32

    iterate on your design. Uh, and you I

  158. 6:35

    mean even if you're a soloreneur, but if

  159. 6:37

    you're not a soloreneur, you might have

  160. 6:39

    different people who have different

  161. 6:40

    opinions. Users have opinions. I mean

  162. 6:42

    design is messy. it's process and

  163. 6:45

    wishing it away and just saying like

  164. 6:47

    solve this problem for me agent will not

  165. 6:49

    unfortunately solve your problem. Um

  166. 6:54

    now the other problem is if you don't

  167. 6:56

    know the language of design or if your

  168. 6:58

    stakeholders don't know the language of

  169. 7:00

    design gets really frustrating really

  170. 7:01

    quickly too. So if you've ever designed

  171. 7:03

    something um you kind of know like this

  172. 7:05

    kind of experience um and then you know

  173. 7:08

    the inner feeling builds up of like

  174. 7:11

    and at some point you're like

  175. 7:15

    uh

  176. 7:16

    so really what you have to ask first is

  177. 7:18

    what's the emotional territory? What

  178. 7:20

    should this never feel like? What's

  179. 7:22

    what's the reference? What's the

  180. 7:24

    audience? These are all really important

  181. 7:26

    questions to figure out. Like you can't

  182. 7:27

    go to a human design studio even if it's

  183. 7:31

    the best design studio in the world. You

  184. 7:33

    go to design director and you say like

  185. 7:35

    you know what I want to design for my

  186. 7:37

    brand and a design dire not nods. It's

  187. 7:40

    like got it and then walks away. That

  188. 7:43

    doesn't make any sense. Like they

  189. 7:45

    obviously have to figure out what it is

  190. 7:46

    that you're trying to build and what

  191. 7:47

    you're trying to get out of it.

  192. 7:50

    Importantly on this example even sort of

  193. 7:52

    like the 2026 version of AI slop you

  194. 7:55

    know quickly vibe coded page nobody

  195. 7:58

    decided anything here

  196. 8:01

    like you just oneshotted this nobody

  197. 8:03

    decided anything on this page this was

  198. 8:05

    just I mean it's maybe competent but

  199. 8:08

    completely empty

  200. 8:10

    so my uh thesis that I make here and

  201. 8:15

    what I've been trying to build over the

  202. 8:16

    last couple months is a tool that gives

  203. 8:19

    you just the amount of control needed to

  204. 8:22

    steer your agent in the right direction

  205. 8:24

    with adjectives and verbs.

  206. 8:26

    And these are the types of words I'm

  207. 8:28

    talking about. Talking about words like

  208. 8:30

    make this bolder, make this quieter, uh

  209. 8:33

    distill this. So basically

  210. 8:34

    simplification, um polish this, um make

  211. 8:37

    it denser or harden this. Um harden for

  212. 8:41

    instance would mean uh make sure that

  213. 8:43

    this design works across the board.

  214. 8:44

    Maybe the performance is bad on certain

  215. 8:46

    devices. Maybe it's not responsive. Um,

  216. 8:49

    and so these are these are words

  217. 8:51

    that we use in the world of design. And

  218. 8:55

    I noticed when seeing two people attempt

  219. 8:57

    the exact same task and one is an

  220. 8:59

    engineer who's never really touched

  221. 9:00

    design at all and the other one is a

  222. 9:02

    designer um there's a stark difference

  223. 9:06

    even when they use the same model and

  224. 9:07

    same harness everything in the output

  225. 9:09

    that they're getting based on the

  226. 9:10

    language that they use. So, I've been

  227. 9:14

    trying to put that language basically

  228. 9:17

    compress it into a skill uh and into a

  229. 9:19

    system to be able to express yourselves

  230. 9:21

    better. And so, for example, you could

  231. 9:24

    have like a baseline um over here, but

  232. 9:27

    then you make it bolder and the agent

  233. 9:30

    actually knows what bolder in your

  234. 9:31

    context means. You make it quieter, it

  235. 9:33

    knows what it means. You distill and it

  236. 9:36

    knows what it means.

  237. 9:40

    Um, but then bolder how what does bolder

  238. 9:44

    mean in the context of your project?

  239. 9:45

    That's a really important question

  240. 9:47

    because as you could tell at the

  241. 9:48

    beginning of the demo uh in the cold

  242. 9:51

    open uh when you just ask GPT or claude

  243. 9:55

    whatever make this bolder it doesn't

  244. 9:56

    know anything about what you mean. Uh

  245. 9:58

    now it has its own sense of reference of

  246. 10:01

    like what boulder could mean but it will

  247. 10:03

    do whatever it wants right it will

  248. 10:05

    invent new colors it will invent new

  249. 10:07

    gradients. it will I mean it doesn't

  250. 10:09

    matter um what it actually means in

  251. 10:12

    impeccable

  252. 10:14

    not gradients not glass not neon but

  253. 10:17

    actually hierarchy scale decisive type

  254. 10:20

    so uh things that uh don't break your

  255. 10:23

    design system and adequately raise the

  256. 10:27

    attention that the user would put on it

  257. 10:31

    uh and this is an actual sentence from

  258. 10:33

    the the file that is loaded when you say

  259. 10:37

    bolder impeccable um show someone your

  260. 10:39

    work and say AI make this bolder made

  261. 10:42

    this bolder if they believe you you

  262. 10:44

    failed this is something that we

  263. 10:47

    actually instruct the agent to do for

  264. 10:49

    themselves and then the agent oftent

  265. 10:50

    times reflects and like I don't know

  266. 10:54

    think I did some bad work here

  267. 10:57

    an adjective is a light wart and a light

  268. 11:01

    wart such a German term and Matt Puk

  269. 11:05

    actually talked about this recently uh

  270. 11:06

    who also makes some really excellent

  271. 11:08

    skills. And so this is uh stolen by one

  272. 11:10

    of his tweets, but I really think uh um

  273. 11:14

    it makes a lot of sense. Um these are

  274. 11:16

    words that h that you basically imbue

  275. 11:19

    with meaning that you infuse with

  276. 11:21

    meaning and they mean something. They

  277. 11:23

    already mean something to the model, but

  278. 11:24

    then you translate this into the area

  279. 11:28

    that you're interested in

  280. 11:30

    and the right altitude really shifts.

  281. 11:34

    Um, now I don't think you can solve any

  282. 11:38

    problems at that altitude. That's super

  283. 11:40

    important too. Uh, there is still room

  284. 11:42

    for the very beginning. For instance, if

  285. 11:44

    you just want exploratory work and you

  286. 11:46

    just want to get something on the screen

  287. 11:47

    or you just want the last amount of

  288. 11:50

    polish. I don't think right now uh AI is

  289. 11:53

    good enough to replace humans for the

  290. 11:55

    last 5, 10, maybe even 20% of the work

  291. 11:57

    to get it from good to great. Uh but it

  292. 12:02

    does actually make it better. Now I've

  293. 12:05

    built tools for more than 20 years. I

  294. 12:07

    guess I'm dating myself here. Uh but

  295. 12:10

    whenever I build a tool, I actually map

  296. 12:11

    the workflow first. I think about what

  297. 12:14

    is the actual workflow of the target

  298. 12:15

    user that I'm trying to help. And the

  299. 12:18

    workflow of design is not linear. It's

  300. 12:20

    messy. Has many different stages. Uh

  301. 12:23

    there are different people involved. And

  302. 12:25

    so I kind of mapped every part of the

  303. 12:27

    process and thought about well what are

  304. 12:29

    the injection points along that route

  305. 12:32

    and how can we help along that route.

  306. 12:35

    And so you start with initialization

  307. 12:36

    with the shaping process of design uh

  308. 12:39

    you go over to uh crafting to iterating

  309. 12:42

    uh with all these verbs and adjectives

  310. 12:44

    and then to the actual harden and polish

  311. 12:46

    phase and finally you kind of bring it

  312. 12:48

    back to your design system. You do some

  313. 12:49

    tech debt cleanup design deck cleanup.

  314. 12:53

    Um, now here's one example of a command

  315. 12:56

    that I wasn't actually sure about

  316. 12:57

    shipping and it's called overdrive. Um,

  317. 13:01

    I basically built this uh built this

  318. 13:03

    command in Impeccable that uh tries to

  319. 13:06

    create something completely over the top

  320. 13:08

    and it was kind of like a a half of a

  321. 13:10

    joke maybe. Um, but people absolutely

  322. 13:14

    are rabbit about this about this thing.

  323. 13:15

    Um, and they love it. Um, and so this is

  324. 13:18

    just one example of I mean this is not

  325. 13:20

    oneshotted. Uh this is a um sort of

  326. 13:22

    realistic event horizon shader um that

  327. 13:25

    I've built for my shader library radiant

  328. 13:27

    shaders. But I did actually use the

  329. 13:29

    overdrive command for this to uh

  330. 13:32

    overdrive it even more uh and make it

  331. 13:34

    more ridiculous. So sometimes I come up

  332. 13:36

    with these uh commands I'm actually not

  333. 13:38

    sure about and I test them with the

  334. 13:39

    community and once it sticks uh I

  335. 13:42

    realize okay well there are more people

  336. 13:44

    who are getting joy out of this.

  337. 13:46

    Um

  338. 13:49

    oh here we go. Here we go. Okay. Um, an

  339. 13:52

    adjective with nothing behind is just a

  340. 13:54

    nicer prompt. So, you really have to

  341. 13:56

    tell the agent what you mean by the word

  342. 13:59

    that you're using. Uh, it it has so many

  343. 14:01

    different meanings uh within its

  344. 14:04

    training materials that you really have

  345. 14:05

    to tell it. Um, now here's another

  346. 14:07

    thing.

  347. 14:09

    I've gotten so many requests

  348. 14:12

    for an automatic way of using

  349. 14:13

    Impeccable. So many. It's so

  350. 14:17

    frustrating. Like so at least once a

  351. 14:19

    week somebody tweets at me say like you

  352. 14:21

    know what would be cool just let

  353. 14:22

    impeccable do all the work by itself. I

  354. 14:24

    don't have to decide on any command

  355. 14:25

    anymore. That's not the point. The point

  356. 14:27

    is to give you a way to steer what you

  357. 14:30

    want to end up with. It's never going to

  358. 14:33

    be a tool to oneshot design. That's not

  359. 14:35

    the intent. And I'll never I mean

  360. 14:37

    there's a pull request standing uh

  361. 14:39

    standing out right now that I'm going to

  362. 14:40

    close because of that.

  363. 14:43

    There is no auto and there will be no

  364. 14:45

    auto

  365. 14:47

    in my opinion and uh this is a bit

  366. 14:49

    awkward maybe for some of the next

  367. 14:51

    speakers but I think taste cannot be

  368. 14:54

    amplified uh sorry I think taste can be

  369. 14:56

    amplified I think I call this amplified

  370. 14:58

    craft uh and I think you can work with

  371. 15:01

    tools to sharpen your taste and bring

  372. 15:04

    out more of your taste I really do not

  373. 15:05

    think it can be lab ground lab grown I

  374. 15:07

    think it is uh taste by definition sort

  375. 15:11

    of is uh contextual It's cultural. It's

  376. 15:15

    scars. And once everybody replicates the

  377. 15:17

    same thing, it starts to become really

  378. 15:20

    muddy and we don't think of it as taste

  379. 15:23

    anymore. That's my opinion. And I know

  380. 15:24

    this is like a hotly contested topic.

  381. 15:26

    Uh, but I'm not trying to solve taste

  382. 15:28

    with my tools.

  383. 15:30

    Um,

  384. 15:32

    that's what I got. Uh, this was about

  385. 15:34

    finding the right level of control for

  386. 15:36

    design. Thank you very much. And, uh, I

  387. 15:38

    think I have time for a few questions if

  388. 15:40

    you have any. Thank you.

  389. 15:56

    >> [music]