Latest / AI Fire Daily / #579 Neil: AI Design Starts With ChatGPT’s Most Predictable Look

Transcript

#579 Neil: AI Design Starts With ChatGPT’s Most Predictable Look · AI Fire Daily

• 0:00
  You look at an AI -generated layout. It is clean.
• 0:04
  It is polished. And yet, beat, you get this eerie
• 0:08
  feeling. You have seen it 47 times before. Oh,
• 0:11
  man. Yeah. We have all been there. It is basically
• 0:14
  the absolute curse of the algorithm. Welcome
• 0:16
  to the Deep Dive. I am really glad you are joining
• 0:18
  us today. So glad to be here. We are exploring
• 0:22
  AI design today, and more specifically, how to
• 0:26
  stop it from being so aggressively predictable.
• 0:29
  Right. Because today, we are exploring this really
• 0:32
  brilliant five -step workflow for chat GPT. Instead
• 0:36
  of begging the AI to be more creative and just
• 0:39
  getting the exact same design in a slightly different
• 0:41
  color, we are going to learn how to trick it.
• 0:43
  Trick it into true originality. Exactly. And
• 0:46
  the way we do this is actually completely counterintuitive.
• 0:49
  How so? Well, we do it by forcing it to be boring
• 0:51
  first. I love this concept. But before we get
• 0:54
  into the solution, we should probably unpack
• 0:55
  the problem. Definitely. We need to talk about
• 0:57
  the make -it -creative trap, because I think
• 1:00
  that is where basically everyone starts. It is
• 1:04
  exactly where they start. You ask ChatGPT for
• 1:06
  a design concept. It spits out something deeply
• 1:09
  average. You look at it, you frown, and you reply,
• 1:12
  make it more creative. I have to admit something
• 1:14
  here. Oh boy. I still wrestle with prompt drift
• 1:18
  myself. Really? Yeah, where I just keep typing
• 1:21
  make it pop or be edgier until it is just a neon
• 1:26
  mess. I know better, but it is such a natural
• 1:30
  human habit. It totally is. We think vague enthusiasm
• 1:34
  equals a better result. We all do it. We treat
• 1:36
  the machine like a human artist. We think it
• 1:39
  understands abstract enthusiasm and what actually
• 1:42
  happens under the hood. It gives you the exact
• 1:45
  same outfit. It just puts it in a different color.
• 1:47
  The font changes a tiny bit. The gradient gets
• 1:50
  mathematically stronger. Maybe the UI cards get
• 1:53
  more rounded. Oh, absolutely. And Chat Cheap
• 1:56
  PT will confidently tell you this is a fresh
• 1:58
  and highly unique design. Which it is not. No,
• 2:01
  it is not. It is just the same underlying structural
• 2:03
  skeleton wearing a slightly louder shirt. Now,
• 2:06
  I know everyone loves Claude for writing right
• 2:08
  now. Yeah. My own instinct would normally be
• 2:10
  to jump into Claude for complex reasoning, but
• 2:13
  our sources actually specify chat GPT here. They
• 2:16
  do, yeah. So why does chat GPT win this specific
• 2:20
  workflow? Well, Claude is fantastic for nuanced
• 2:23
  text, but for generating direct iterative visual
• 2:26
  systems in this specific structured loop, ChatGPT
• 2:30
  leaves Claude eating dust. Really? Just entirely?
• 2:33
  Yeah. The way ChatGPT integrates with its image
• 2:37
  generation and holds on to rigid mathematical
• 2:40
  system instructions across a long multimodal
• 2:43
  context window, that is precisely what makes
• 2:46
  this five -step process work. Which brings us
• 2:48
  to the mechanics of the prompt. Right. Vague
• 2:50
  prompts are essentially like giving a painter
• 2:52
  a canvas but no edges. The AI just gets entirely
• 2:55
  lost in its own data. Exactly. If the context
• 2:58
  is too vague, ChatGPT just fills the gaps probabilistically.
• 3:01
  It goes to the safest bets. Always. It defaults
• 3:04
  to the safest, most common choices in its training
• 3:06
  data. That is why everything looks like a Canva
• 3:08
  template. So step one is defining the AI design
• 3:11
  target. explicitly. Very explicitly. And I mean
• 3:13
  explicitly. Before you even mention visuals or
• 3:16
  aesthetics, you must set the rigid boundaries
• 3:18
  of the business. You have to tell it exactly
• 3:20
  who this is for. You define the industry. You
• 3:23
  define the target demographic. You establish
• 3:25
  the core impression you want to leave. We were
• 3:27
  talking about the product context, the brand
• 3:30
  positioning, and crucially, any visual directions
• 3:33
  you already know you absolutely do not want.
• 3:36
  Exactly. If we look at a classic example from
• 3:38
  the sources, a brand for a cafe. Okay, let's
• 3:41
  use that. If you just type, make a modern cafe
• 3:43
  brand, that is way too broad. We all know what
• 3:46
  happens then. ChatGPT immediately runs to the
• 3:49
  safest zone imaginable. Yep. Beige colors. A
• 3:52
  minimalist serif font. Yeah. A logo that is just,
• 3:55
  you know, a stylized coffee bean or a steaming
• 3:58
  cup. It looks very artisanal, very clean. Yeah.
• 4:01
  But you have seen it on every street corner for
• 4:04
  the last 10 years. So we lock it down. We tell
• 4:06
  it. This is a small, high -volume urban cafe.
• 4:10
  License specific. It targets professionals in
• 4:13
  their 20s and early 30s. The core impression
• 4:15
  should be casual, bold, and slightly unexpected.
• 4:18
  And then you add the negative constraints right
• 4:20
  up front. You tell it. Avoid luxury hotel vibes.
• 4:25
  Right. Avoid that overly rustic reclaimed wood
• 4:29
  coffee branding. But wait, if we know exactly
• 4:31
  what we want the brand to feel like, why shouldn't
• 4:34
  we demand a highly unique, out -of -the -box
• 4:36
  design right here? in step one. Because without
• 4:40
  constraints, forcing creativity early just makes
• 4:42
  the AI panic and revert to the safest templates
• 4:45
  it knows. It doesn't actually know what unique
• 4:48
  means without a baseline. Right. So forcing early
• 4:51
  creativity just triggers safe templated panic.
• 4:54
  Exactly. You have to build the playground first,
• 4:57
  which takes us to step two. Step two. Where we
• 5:00
  do the exact opposite of what feels natural.
• 5:02
  This is the part that fascinated me. Yeah. Your
• 5:04
  brief is locked in. The audience is clear. And
• 5:07
  now you explicitly ask the smartest machine on
• 5:10
  earth for a cliche. You literally prompt it to
• 5:13
  generate the design it would normally produce
• 5:15
  nine times out of ten. You demand the safest,
• 5:18
  most painfully obvious direction. You really
• 5:20
  do. I was thinking about this from a philosophical
• 5:22
  angle. We are essentially asking the machine
• 5:25
  to show us its reflexes. Okay. Beat. It is fascinating
• 5:29
  to systematically map its default neural pathways.
• 5:32
  That is a brilliant way to put it. You are forcing
• 5:35
  it to externalize the mathematical average of
• 5:38
  its training data. Right. But the trick here
• 5:40
  is you don't just ask for a quick summary. You
• 5:44
  demand the full comprehensive visual system.
• 5:47
  We want the color palette. Yes. The typography
• 5:50
  hierarchy. The logo direction. Layout rules.
• 5:54
  Photography style, packaging materials, common
• 5:57
  motifs. You want the entire mood board mapped
• 5:59
  out in text. And crucially, you ask the AI to
• 6:02
  explain its own choices. You have to. Like, why
• 6:04
  are these specific colors commonly used for urban
• 6:07
  cafes? Why this font? Right. You want it to articulate
• 6:09
  why it feels predictable. You have it summarize
• 6:12
  the repeated visual patterns that make this specific
• 6:14
  output a stereotype. But let me play devil's
• 6:17
  advocate for a second. Go for it. We are spending
• 6:19
  time and tokens generating something we know
• 6:21
  we hate. Does staring at this totally predictable,
• 6:24
  boring stereotype actually help the human driving
• 6:27
  the prompt? Absolutely, because it externalizes
• 6:30
  all the cliches, making it incredibly obvious
• 6:33
  what visual crutches you need to actively dodge
• 6:36
  in the next step. You cannot fight an invisible
• 6:38
  enemy. Yeah, seeing the cliche mapped out makes
• 6:41
  it easier to spot what is missing. And that realization
• 6:44
  is the entire foundation for step three. The
• 6:47
  avoid lists. Exactly. We take that boring AI
• 6:50
  design and we weaponize it. We turn it into a
• 6:54
  strict avoid list. This is where the magic really
• 6:56
  starts happening. It really does. We recognize
• 6:59
  the predictable output. Let's say it gave us
• 7:01
  that cream background, the serif font, the clean
• 7:03
  minimalist photography. We do not throw that
• 7:05
  output away. No. We use it as structural fuel.
• 7:08
  We build an impenetrable mathematical fence around
• 7:11
  the cliche. You explicitly write a prompt that
• 7:13
  says, review the stereotypical design above.
• 7:16
  turn it into a clear avoid list. And you ask
• 7:19
  it to categorize that list. You group it by colors,
• 7:22
  typography, layout, imagery, and overall branding
• 7:25
  style. Yes. Everything organized. This is such
• 7:28
  a powerful concept, it immediately reminds me
• 7:31
  of the philosophy of via negativa. Oh, improvement
• 7:34
  by subtraction. Exactly. We are defining the
• 7:37
  masterpiece by chiseling away everything that
• 7:40
  is boring. That is exactly what is happening
• 7:42
  under the hood. You are actively countering the
• 7:45
  baseline model behavior. Let me just clarify
• 7:47
  that term for everyone. Sure. Baseline model
• 7:50
  behavior. That is the default habits the AI uses
• 7:54
  when lazy. Spot on. Left to its own devices,
• 7:57
  a large language model wants to be lazy. It really
• 8:00
  does. It is literally designed to give you the
• 8:02
  most probable mathematically average answer.
• 8:05
  So by giving it an avoid list, we are placing
• 8:08
  negative weights on those probable paths. Yes.
• 8:11
  We are building a fence against the average.
• 8:13
  But there is a rule here, right? A critical rule,
• 8:16
  yeah. While you are walling off the visual cliches,
• 8:19
  you must command the AI to keep the original
• 8:21
  target audience entirely unchanged. The business
• 8:24
  goal stays the same. Exactly. The core impression
• 8:27
  stays the same. We are only banning the lazy
• 8:30
  visual execution of those goals. But you might
• 8:32
  be wondering, why go through all this trouble?
• 8:36
  So what happens if you skip making this strict
• 8:38
  list and just tell it to think outside the box?
• 8:40
  The AI just shifts from one predictable style
• 8:43
  to a slightly different predictable style because
• 8:46
  the word different is too vague for it to process
• 8:48
  structurally. Ah, different is vague. Yeah. An
• 8:51
  avoid list builds a hard mathematical wall. Exactly.
• 8:55
  Outside the box just means move to the second
• 8:58
  most common Pinterest board in my database. Right.
• 9:01
  The avoid list destroys the box entirely. Which
• 9:04
  brings us to step four. generating different
• 9:07
  directions. The fun part. The walls of cliche
• 9:10
  are finally blocked off. The AI is now forced
• 9:13
  into uncharted territory. It has to navigate
• 9:16
  around the massive avoidless you just dropped
• 9:18
  in the middle of its latent space. You prompt
• 9:21
  ChatGPT to create five clearly different visual
• 9:24
  design directions. Five is a great structural
• 9:26
  number. It is. It forces genuine variety without
• 9:29
  overwhelming the context window, but the rules
• 9:32
  still strictly apply. Same audience, same core
• 9:35
  impression. Yep. The only thing changing is the
• 9:37
  visual language. You have to explicitly instruct
• 9:40
  the model. Every direction must move away from
• 9:42
  the stereotypical patterns in the avoid list.
• 9:45
  Do not replace one predictable style with another
• 9:48
  predictable style. For each of those five directions,
• 9:51
  you demand extreme specificity. Extreme. A clear
• 9:54
  concept name. The main creative idea driving
• 9:57
  it. Color direction. Typography rules. The logo
• 10:01
  approach. Layout structures. Main graphic elements
• 10:04
  and materials. You want to see real structural
• 10:06
  differences here, not just a font swap. And here
• 10:08
  is the genius part of the prompt architecture.
• 10:11
  Tell me. Before it shows you any of these directions,
• 10:13
  it must internally check its own work. Oh, wow.
• 10:17
  It has to cross -reference its new ideas against
• 10:19
  the avoid list. So if it catches a predictable
• 10:21
  pattern sneaking in, it must revise it internally
• 10:25
  before outputting the final text to you. Whoa.
• 10:28
  B, just imagine scaling that. Generating five
• 10:32
  completely distinct, fully realized brand systems
• 10:35
  in seconds, all rigorously pre -checked for originality.
• 10:39
  It is wild. It is staggering. It completely changes
• 10:42
  the creative pipeline from generating ideas to
• 10:45
  curating systems. Once the text is dialed in,
• 10:47
  you have ChatGPT generate a visual presentation
• 10:50
  for each concept so you can compare them fairly.
• 10:52
  You are looking at the actual design cohesion
• 10:54
  now, but I do have a concern here. What is that?
• 10:58
  How do we stop the AI from subtly altering our
• 11:03
  target demographic just to make its new weird
• 11:06
  design fit better? You have to explicitly command
• 11:09
  it in the prompt to anchor the audience completely,
• 11:12
  forcing only the visual aesthetics to shift.
• 11:14
  You don't let it change the customer to fit the
• 11:17
  art. Got it. Anchor the audience completely.
• 11:19
  So only the aesthetics are forced to shift. Once
• 11:22
  you have those five distinct audience anchored
• 11:25
  directions, we move into the final stage. Step
• 11:28
  five. Step five is refinement. The wild exploration
• 11:31
  phase is over. Yes. We have found our anomaly
• 11:34
  and now we have to turn it into a reliable system.
• 11:37
  You review the five options. You look for the
• 11:39
  one that pushes the furthest away from the baseline
• 11:41
  while still actually solving the business problem.
• 11:43
  There is a really important warning in the sources
• 11:45
  here. Oh, definitely. Don't just pick something
• 11:47
  unusual if it is too hard to use in real life.
• 11:49
  That is a crucial point for any designer listening.
• 11:52
  Weird for the sake of weird is not good design.
• 11:54
  It has to be highly functional. It is like stacking
• 11:57
  Lego blocks of data into a stable, usable structure.
• 12:00
  Exactly! A crazy asymmetrical logo might look
• 12:03
  cool in a prompt, but if it doesn't fit on a
• 12:05
  coffee cup or an app icon, it fails. I love that
• 12:09
  Lego analogy. You are building out the functional
• 12:11
  reality of the brand. Once you choose the strongest
• 12:14
  direction, you feed that decision back to ChatGPT.
• 12:17
  You say, we are moving forward with direction
• 12:19
  three. Right. Keep the main creative idea, but
• 12:22
  now refine it. You ask it to flesh out the entire
• 12:25
  functional system. Logo variations for different
• 12:28
  formats. Balance color system. Typography hierarchy
• 12:30
  for headers and body copy. Packaging design.
• 12:33
  Signage. Interface details. You ask ChatGPT to
• 12:37
  create three refined variations of this single
• 12:40
  chosen direction. Each of these three variations
• 12:43
  should feel like they belong to the exact same
• 12:45
  brand. You are not asking for new concepts anymore.
• 12:47
  No, you are stress testing the chosen concept.
• 12:50
  And again, you must force it to check every single
• 12:53
  variation against the avoid list from step three.
• 12:55
  Because the predictable patterns always want
• 12:57
  to sneak back in. It is gravitational pull. You
• 13:00
  have to actively guard against them during refinement.
• 13:02
  So what do you do if one of the three refined
• 13:06
  variations starts subtly slipping back into those
• 13:09
  familiar cliches despite the guardrails? You
• 13:12
  have to immediately intervene and pull chat GPT
• 13:15
  back to the original concept from step one, reinforcing
• 13:18
  the avoid list heavily in your next prompt. Makes
• 13:21
  sense. Just pull it back to the core concept
• 13:22
  immediately. Don't let it drift. When you execute
• 13:25
  the step correctly, you end up with a highly
• 13:27
  original, structurally sound brand system. And
• 13:32
  you can actually use it across real world applications
• 13:35
  without losing the core identity. Exactly. Let's
• 13:38
  zoom out for a second and look at the bigger
• 13:39
  picture of what we just did. The core philosophy
• 13:42
  of this entire workflow is endlessly fascinating
• 13:45
  to me. Yeah. You cannot just command a machine
• 13:48
  to be creative. It simply does not work. If you
• 13:51
  want true originality from an AI, you first have
• 13:53
  to let it show you its most boring, predictable
• 13:56
  answer. You have to patiently map the cliche.
• 13:58
  You have to understand the mathematical average.
• 14:00
  And then you have to wall it off completely.
• 14:02
  By building that fence, you force the system
• 14:05
  down the road less traveled. Sometimes the fastest,
• 14:09
  most effective way to innovation. is explicitly
• 14:11
  defining the boring. You aren't asking for vague
• 14:14
  creativity anymore. You are architecting a very
• 14:17
  specific negative space. You've shown the AI
• 14:20
  what must stay anchored and exactly what is forbidden.
• 14:23
  It completely changes how you think about prompting.
• 14:26
  It moves you from being a requester to being
• 14:28
  a director. It really does. Which brings me to
• 14:30
  a final thought I want to leave you with today.
• 14:32
  Oh, this is a good one. Think about this for
• 14:34
  a second. Beat. If an AI naturally defaults,
• 14:38
  to the average of all human design. Doesn't that
• 14:42
  mean our own human brainstorming sessions are
• 14:45
  probably just as predictable? It is a genuinely
• 14:47
  sobering thought. Human brains want to be lazy,
• 14:50
  too. Right. We are just as prone to leaning on
• 14:52
  cliches when we are staring at a blank whiteboard.
• 14:55
  Exactly. So next time you have a human team meeting,
• 14:58
  try something new. Yeah. Try making an avoid
• 15:00
  list of your industry's most boring cliches.
• 15:02
  Do it together as a team before anyone is allowed
• 15:05
  to pitch a single idea. I guarantee your human
• 15:09
  meetings will get much more creative, much faster.
• 15:11
  Because at the end of the day, you look at a
• 15:14
  layout. It is clean. It is polished. But you
• 15:18
  never again want that eerie feeling you have
• 15:20
  seen it 47 times before. Never again. Thanks
• 15:23
  for joining us on this deep dive. We will catch
• 15:25
  you next time.