Latest / AI Fire Daily / #598 Neil: Claude Design Builds Reusable Rules From One Reference

Transcript

#598 Neil: Claude Design Builds Reusable Rules From One Reference · AI Fire Daily

0:00
We've all been there, right? You spend, I don't
0:01
know, an hour crafting the perfect AI image prompt.
0:05
And the result you get, it looks truly amazing.
0:07
It's exactly what you wanted. But then you ask
0:10
for a really simple second version, just one
0:13
small tweak. And it all falls apart. Completely.
0:16
Suddenly the fonts drift into something unrecognizable.
0:19
The colors completely shift away from your brand.
0:21
The whole layout just... It just breaks. It's
0:24
this deeply universal frustration with AI design.
0:28
It really is the worst. I mean, you finally get
0:29
that perfect aesthetic look, right? And then
0:31
the AI just seemingly forgets everything you
0:33
just taught it. It's like you're back to square
0:35
one. But today, we actually fixed that problem.
0:39
We're finally solving it. We are. We're going
0:41
to reverse engineer a PERSIC design. We'll turn
0:44
it into an automated, completely reusable AI
0:47
system. And we're doing this entirely using Claude
0:50
Design. Welcome to the deep dive. We have a very
0:54
clear roadmap for you today. We're exploring
0:56
a specific six step workflow. We'll move from
1:00
just a single reference image and we'll end up
1:02
with a completely foolproof visual system. I
1:05
really want to take this slow today. I want to
1:08
understand the underlying logic here. You know,
1:10
how this actually works under the hood. Yeah,
1:12
I think we need some context first before we
1:14
get into the steps. So there are several cheaper.
1:18
you know, Chinese AI models available today.
1:20
Right, highly capable ones. Very capable, incredibly
1:23
fast. But Claude Design has one massive advantage
1:26
over them. It can study a really good visual
1:29
design and actually find the mathematical rules
1:31
hitter behind the pixels. And then it builds
1:35
a reusable framework from that math. Which means
1:37
you never have to start from zero again. Okay,
1:40
so let's talk about that starting point. Step
1:42
one. We obviously need to feed the AI some sort
1:46
of inspiration. Right. I assume we don't just
1:48
give it a blank canvas and say, you know, do
1:50
your best. No. We must upload a very clear visual
1:54
reference. So it could be a vintage poster or
1:56
maybe a modern website. But what kind of instructions
1:59
are we actually giving it when we upload that
2:01
image? So we give Claude a highly detailed brief.
2:04
We're giving it very strict factual instructions
2:07
to follow. We tell it to deeply analyze the overall
2:10
layout. we ask it to study the visual hierarchy.
2:14
We really want it looking at the typography very
2:16
closely. Like the specific fonts. Exactly. Yeah.
2:19
It needs to check specific font sizes, the weights
2:21
of the text. It also checks the spacing and the
2:24
structural alignment of the whole page. So it's
2:27
breaking down the actual anatomy of the design.
2:30
Yeah. It analyzes the specific use of empty space.
2:34
It looks at signature visual details across the
2:36
canvas. Subtle things, right? Like borders or
2:39
drop shadows. All of it. But how do we control
2:41
what it actually takes? Because if it just copies
2:44
it, that's a problem. Right. So we give Claude
2:46
a very strict boundary here. We explicitly tell
2:49
it not to copy the original brand. It must never
2:52
copy the original text exactly. Instead, we have
2:55
it separate the visual rules into two distinct
2:57
buckets. Two buckets. Yeah. Bucket one holds
2:59
the elements that must stay consistent, the fixed
3:02
stuff. Bucket two holds the elements that flex
3:05
with new content. And this separation generates
3:08
our very first draft. I have to say, I've tried
3:10
something vaguely similar before, and the AI
3:13
usually just ends up ripping off the original
3:15
logo anyway. Yeah, that happens a lot. So how
3:18
do we ensure the AI is actually learning the
3:21
style? We really need to prevent actual plagiarism
3:24
from happening here. It really comes down to
3:26
the level of abstraction we're asking for. OK.
3:29
Claude is specifically analyzing the spatial
3:32
relationships on the page. It's looking for the
3:35
underlying geometric logic. It is not looking
3:38
at the actual colored pixels. It ignores the
3:42
original creative content entirely. It just simply
3:45
maps how the different elements interact with
3:48
each other. So it steals the underlying blueprint,
3:50
not the actual house. Exactly. It just wants
3:53
the architectural math. It completely leaves
3:55
the original furniture behind. OK. That makes
3:58
perfect sense. So we have this first generated
4:00
draft. Yeah. But in my experience, the first
4:02
draft is rarely perfect. It usually gets us,
4:05
I don't know, maybe 80 % of the way there. If
4:07
you're lucky, yeah. Right. And if we extract
4:09
the rules right now, we're basically extracting
4:12
flawed rules. Honestly, I still wrestle with
4:15
prompt drift myself. Oh, we all do. It's incredibly
4:18
hard to manage over multiple iterations. You
4:21
fix one thing, another thing breaks. And that
4:23
is exactly why this next phase exists. Step two.
4:26
We do not extract the rules from that rough first
4:30
draft. We create a separate, completely reusable,
4:34
clawed skill instead. We call this specific utility
4:37
the design loop. The design loop. Yeah. It performs
4:41
four specific, very rigorous review passes on
4:44
the draft. Tell me more about those specific
4:47
review passes. What exactly is it looking for
4:49
in the very first pass? So pass one specifically
4:52
checks the overall composition. OK. It's looking
4:54
for structural balance and a clear focal point.
4:56
It basically wants to know where the eye travels
4:58
first when you look at it. Right, the visual
5:00
anchor. Exactly. Then Pass 2 focuses heavily
5:03
on typography and structural hierarchy. It asks
5:06
if the main headline has enough visual impact.
5:09
It checks for strict font consistency across
5:11
the entire design. So it's actively critiquing
5:14
its own previous work. Very critically, yeah.
5:17
That's fascinating. And then Pass 3? Pass 3 is
5:19
all about spacing. and exact alignment. It checks
5:23
the margins, the invisible grids. It makes sure
5:26
no specific area feels too visually crowded.
5:29
It's acting like an art director. Basically,
5:31
yeah. So what happens in the final pass of this
5:34
loop? Pass four is the final visual consistency
5:38
check. It verifies the exact color palette is
5:41
applied correctly. It checks that repeated details
5:44
match perfectly throughout the entire canvas.
5:46
Wow. Yeah. And we run this loop on the first
5:49
draft inside our original chat. Oh, I see. We
5:51
keep the core visual concept completely intact.
5:55
We just systematically polish the weak areas.
5:57
We just do this until we have a flawless source
5:59
design. I understand the function of refining
6:02
the draft. It makes total sense. Yeah. But I'm
6:04
curious about the mechanical setup here. Why
6:06
build this design loop as a completely separate
6:08
skill? Why not just prompt the main chat to review
6:11
its own work right then and there? Because you
6:13
want to completely isolate this specific function.
6:17
By isolating it, you create a permanent utility
6:20
tool. You never have to rebuild this critique
6:22
system again. You never have to re -explain the
6:25
four passes for future design projects. They're
6:27
just ready. It just sits there, perfectly ready
6:29
to work. You just call on it whenever a new draft
6:31
needs polishing. Build the tool once, reuse it
6:34
forever without retyping instructions. Exactly.
6:36
You save yourself just hours of repetitive frustration.
6:40
It automates the absolute most tedious part of
6:43
the creative process. So after this rigorous
6:45
loop, we finally have a perfect source design.
6:47
Yes. But right now, it is just a beautiful flat
6:51
image, essentially. We need to extract its soul
6:53
before we can actually automate it. Right. We
6:55
have to translate all that visual beauty into
6:57
usable data. How do we actually pull the rules
7:00
out? What is step three? So we call this specific
7:04
extraction the design DNA. Design DNA. Yeah.
7:07
Claude analyzes the flawless design across 10
7:10
specific vectors. 10. 10 of them. Yeah. And this
7:13
analysis is incredibly thorough. It is deeply
7:17
mathematical. It looks closely at the typography,
7:20
the overall layout. It maps out the grid and
7:23
the specific spacing used. It measures proportions
7:26
alongside the entire color system. It must also
7:29
map the negative space. Right. That's huge in
7:32
design. Oh, definitely. Yes, it charts negative
7:34
space and all the little visual details. It outlines
7:36
the hierarchy and the distinct signature choices.
7:39
Okay. And finally, it clearly defines the specific
7:42
design refusals. Doing all this, it's like stacking
7:47
Lego blocks of data. That's a great way to put
7:49
it. It just builds a complete structural profile
7:51
of the aesthetic. I think that analogy is perfectly
7:54
accurate. It really is like building blocks.
7:56
Yeah. But these resulting rules, they must be
7:58
completely measurable. Correct. They can't just
8:01
be vague, artistic suggestions to the AI. Like
8:05
we cannot just say something like, use bold typography
8:09
or make it pop. No, no. Make it pop will ruin
8:12
everything. Right. The rule must be rooted in
8:14
strict mathematics. Absolutely. The source material
8:17
gives a really great factual example of this.
8:19
Instead of vague instructions, the AI actually
8:22
writes a strict formula. It says, the main headline
8:26
should take 35%. to 45 % of the content width.
8:30
Oh, wow. Right. It measures everything in precise
8:33
percentages and strict ratios. And this extraction
8:37
process typically yields about 10 to 15 core
8:40
rules like that. That is so precise. Yeah. I
8:42
want to circle back to one term you mentioned,
8:44
though. Sure. You brought up something called
8:46
design refusals. Ah, yes. What exactly are they
8:49
and why are they a necessary part of the DNA?
8:51
I really want to understand their specific function
8:53
in this whole system. Refusals are essentially
8:56
negative constraints. for the AI. They clearly
9:00
define what the design should actively avoid
9:02
doing. OK. So they might permanently ban specific
9:06
colors from being used, like never use neon green.
9:08
Right. They might forbid the use of overly busy
9:10
layouts or central line text. They actively prevent
9:14
the core visual identity from breaking down.
9:16
They keep the AI from making messy creative leaps.
9:19
Defining the boundaries is just as crucial as
9:21
painting the picture. Sometimes knowing what
9:24
not to do is actually more important. It just
9:25
keeps the automated system safely on the rails.
9:28
OK, so we've successfully extracted this mathematical
9:31
DNA now. We have our 10 to 15 rules. But isolated
9:35
DNA still needs a proper host to survive. We
9:38
must organize these scattered rules into a master
9:41
file. We have to build the actual structural
9:44
system. This is step four and five. Right. So
9:46
how do we organize these core rules? We organize
9:49
the DNA into a 12 -part AI design system. It
9:53
basically becomes a massive, highly structured,
9:55
technical document. It includes all the foundational
9:58
elements, all the specific typography rules.
10:00
It clearly maps the grid, the spacing, and the
10:02
scale. It defines reusable components alongside
10:06
fixed versus flexible rules. It also outlines
10:09
usage rules and a strict consistency checklist.
10:12
Exactly. And then we have to save this entire
10:15
system document. We save it as a KLAIU .md file
10:18
in the project root. Let me quickly clarify that
10:20
specific term for our listeners who might not
10:22
be developers. A KLAIU .md file is basically
10:25
a hidden rule book the AI reads before it starts
10:27
any work. That is exactly what it is in practice.
10:30
It completely preconditions the AI before you
10:32
even type a prompt. It just knows. It just knows.
10:35
And when you realize the true potential here...
10:38
Whoa, imagine scaling to a billion queries. Seriously?
10:42
You could apply this master rule book across
10:44
an entire enterprise's design output. It guarantees
10:48
total brand consistency globally, completely
10:50
without human oversight. Every single generated
10:53
image instantly follows the corporate math. The
10:56
scale of that is genuinely hard to wrap your
10:58
head around. It's staggering. It's massive. But
11:00
I do see a potential practical issue here. OK,
11:03
what's that? When we start generating a completely
11:06
new design, things naturally change. What happens
11:09
if a rigid rule contradicts the new text content
11:12
we want to use? What if the new headline is just
11:15
way too long for the allocated grid? That exact
11:17
tension is why the system is structurally split.
11:21
The keylad .md file explicitly separates the
11:24
rules into two categories. It separates the fixed
11:27
rules from the flexible rules. Fixed rules strictly
11:30
protect the core visual identity, the absolute
11:33
must -haves. But the flexible rules, they gently
11:36
adapt to content links or new formats. The AI
11:40
basically knows which rules it is allowed to
11:43
dynamically bend. Fixed rules lock the identity,
11:46
flexible ones let the content breathe. Yes. Beautifully
11:49
put it perfectly prevents the design from shattering
11:51
under pressure. It gives the AI structured freedom
11:54
to actually solve layout problems Okay, but a
11:57
theoretical system is honestly pretty useless
11:59
on its own true It is only valid until it survives
12:02
contact with a completely different format. We
12:05
have to stress test this master file in the real
12:07
world Step six. Right, step six. How do we prove
12:10
that the keelyud .md file actually works? So
12:14
this final phase is the ultimate consistency
12:16
test. We purposefully prompt Claude to create
12:19
a completely new poster. OK. But this time, we
12:22
demand a square social media poster. The specific
12:25
dimensions are 1080 by 1080 pixels. Oh, that's
12:28
tricky. Very tricky. And the theoretical event
12:30
is called midnight frequencies. It features an
12:33
electronic artist named Mira Vale. So we are
12:35
moving from a tall vertical canvas to a square
12:37
canvas. The system must adapt all its rules to
12:41
this new geometric constraint. And it must do
12:44
this without losing the original visual identity.
12:47
That seems like an incredibly difficult leap
12:50
for an AI. It really does. But the outcome from
12:53
the source material is genuinely fascinating.
12:56
Claude successfully scales the typography hierarchy
12:59
specifically for the square format. It mathematically
13:02
calculates and adjusts the negative space perfectly.
13:06
It keeps the distinct accent colors completely
13:08
intact. It manages to look like the exact same
13:11
brand campaign, just square. But what happens
13:14
if the math just doesn't fit? Say a headline
13:16
with rule is simply way too strict for a square
13:18
canvas. It has to break at some point, right?
13:20
If a rule fails, Claude actually reports it back
13:23
to you. Really? Yeah. Claude provides a detailed
13:26
visual consistency report. It points out exactly
13:28
which mathematical rule caused the layout conflict.
13:31
That is incredibly helpful. It's amazing. This
13:34
feedback loop allows the user to easily update
13:36
the CLOTUDHD file. You can slightly tweak those
13:40
flexible rules for future square generations.
13:42
I'm still really hung up on the spatial translation
13:45
part of this. How does the AI know how to handle
13:48
a completely different shape? I mean, moving
13:50
from a tall vertical poster to a perfect square
13:53
is jarring. Yeah. How does it arrange the elements
13:56
without just squishing them all together? It
13:58
does not rely on visual memorization at all.
14:00
Claude relies entirely on the mathematical ratios
14:03
we extracted earlier. It looks deeply at the
14:06
grid relationships that are stored in the DNA.
14:09
It doesn't try to just force the original layout
14:11
into a new box. It fundamentally recalculates
14:14
the math for the new specific dimensions. It
14:17
translates the core math to fit the new canvas
14:19
perfectly. Exactly. It just becomes an exercise
14:22
in pure geometry at that point. It solves the
14:24
square canvas as a mathematical equation. Sponsor.
14:27
So let us zoom out for a second and look at the
14:30
big idea here. We really need to summarize this
14:32
philosophical shift for the listener. It's a
14:33
big one. Because this completely changes how
14:36
we approach AI generation. You no longer have
14:39
to start from a terrifying blank prompt every
14:42
single time. The structural pipeline we talked
14:45
about changes everything about visual creation.
14:48
It really gives you a highly repeatable scientific
14:51
process. Exactly. You start by providing a strong
14:54
visual reference. You rigorously run that draft
14:56
through the You systematically extract the mathematical
15:00
design DNA. You organize and build the master
15:03
CLIE .md file. And finally, you run the consistency
15:08
testing on new formats. This means you can turn
15:10
one single spark of inspiration into a permanent,
15:13
scalable visual engine. It really is a profound
15:16
change to the workflow. Treating design as measurable
15:19
data completely changes the creative process.
15:22
It totally removes the frustrating guesswork
15:24
from generating daily assets. I want to leave
15:28
you with one final thought today. If an AI can
15:31
perfectly map our signature design choices into
15:33
pure math, and if it can flawlessly replicate
15:36
them across any format, does visual originality
15:39
just become a highly customizable algorithm?
15:42
I invite you to try extracting the DNA of your
15:44
favorite poster this week. Feed it into the system
15:46
and see what the underlying math actually looks
15:48
like.