Latest / AI Fire Daily / #39 Robin: Coding with Taste - How Claude Code and DESIGN.md are Killing Generic AI T
Transcript
- 0:00You stare at the screen. The Figma canvas is
- 0:02completely blank. Or maybe you just generated
- 0:05another landing page using an AI tool. Right.
- 0:08And it has that same generic purple gradient.
- 0:10You know the exact look I am talking about. It
- 0:12looks perfectly competent, but it feels, well,
- 0:15entirely soulless. It lacks any real perspective.
- 0:18Yeah, it's totally flat. Beat. Welcome to the
- 0:21Deep Dive. I am really glad you're joining us
- 0:23today. We are unpacking something that completely
- 0:25changes how we think about creativity and systems.
- 0:29We really are looking at a fundamental shift
- 0:31today. Our mission is decoding a methodology
- 0:35called Claude Code Design. We're exploring how
- 0:39you take abstract design rules, you know, the
- 0:42things we normally just call taste, and turn
- 0:44them into a repeatable plain text system. Which
- 0:47sounds impossible at first. It does. But the
- 0:49promise here is massive. Because once you build
- 0:51this underlying architecture, you never have
- 0:53to start from zero again. That concept of codifying
- 0:57taste is exactly what hooked me. So let's start
- 1:00by looking at the baseline problem we are all
- 1:01facing. Everyone recognizes current AI design.
- 1:05Oh, absolutely. It defaults to a highly specific,
- 1:08totally forgettable aesthetic. It is the digital
- 1:11equivalent of elevator music. You ask an AI to
- 1:15make a modern website, and it immediately outputs
- 1:17a massive, centered hero section. Always. Below
- 1:21that, you inevitably get three perfectly spaced
- 1:24feature cards. The buttons are always softly
- 1:27rounded. Right. And there's usually a subtle
- 1:29drop shadow on everything. It's clean, sure.
- 1:32But it has absolutely no point of view. It is
- 1:35technically proficient, but artistically vacant.
- 1:38And the root cause isn't a lack of computing
- 1:41power. The source material points directly to
- 1:43vague prompting. Yeah, the classic five -word
- 1:46prompt. We sit down and type, you know, make
- 1:48a modern fintech site. And that is simply not
- 1:50enough structural information. You really can't
- 1:52expect a machine to invent your brand's soul
- 1:54from a five -word prompt. It's just going to...
- 1:57guess. I still wrestle with prompt drift myself.
- 1:59I'll sit there late at night tweaking an AI prompt
- 2:02for hours. Oh man, we've all been there. I'll
- 2:04adjust the wording, ask it to make things sleeker
- 2:07or more dynamic. My eyes will literally start
- 2:10to hurt from staring at the variations. And after
- 2:13all that manual labor, the final result still
- 2:16looks artificially generated. That frustration
- 2:18is universal right now. The fix. requires a completely
- 2:22different approach. We have to translate our
- 2:25visual preferences into strict, measurable rules.
- 2:28Right. We need to dictate the exact typography
- 2:31scales. We need to define the spacing tokens.
- 2:33We really need to establish strict boundaries
- 2:36for layout and colors. It reminds me of working
- 2:39with a master chef. If you walk into a Michelin
- 2:42starred kitchen and just tell the chef, cook
- 2:44me something good. They're going to play it safe.
- 2:47Exactly. They will probably fall back on a standard
- 2:50safe dish. They don't know your palate. But if
- 2:53you hand them your grandmother's specific family
- 2:55recipe, complete with exact ratios for spices
- 2:58and precise resting times, you are guiding their
- 3:01technical expertise. That chef analogy perfectly
- 3:04captures the dynamic. You aren't just handing
- 3:06them the ingredients. You are defining the technique.
- 3:08The source material contrasts this beautifully.
- 3:11The weak prompt just asks for a modern website.
- 3:13But the better prompt. acts like that detailed
- 3:16recipe. It demands a, quote, clean fintech landing
- 3:20page with sharp pricing cards, dark navy accents,
- 3:23and strict sans serif typography. That specific
- 3:27language bridges the gap. You're stripping away
- 3:30the AI's ability to guess. You dictate that headlines
- 3:33must be oversized. You force the layout to maintain
- 3:36massive open sections. You restrict the color
- 3:39palette to three specific hex codes. Right. It
- 3:42forces the output out of that generic purple
- 3:45gradient territory. Which brings up an interesting
- 3:47philosophical point about the tools we use. So
- 3:50is it that the AI lacks the capability to design
- 3:53well, or do we just lack the vocabulary to instruct
- 3:55it? It is entirely about our instructions. The
- 3:59AI has the capability. It just needs us to define
- 4:01the boundaries of taste. Right. We just need
- 4:03to translate visual taste into specific text
- 4:06rules. Exactly. So if the fix is giving the AI
- 4:08a highly specific technical recipe, how do we
- 4:11actually write that document? That leads us straight
- 4:13into the mechanics of text -based design. Two
- 4:15sec silence. This is where the methodology gets
- 4:18incredibly practical. We are looking at a real
- 4:20-world repository called Awesome Designer AI.
- 4:23It's a GitHub tool that basically acts as a library
- 4:26of plain text design systems. That's the part
- 4:29that feels so counterintuitive to me. High -end,
- 4:32nuanced visual taste things that usually require
- 4:34a senior art director to understand can actually
- 4:37live inside a simple markdown text file. Yeah,
- 4:40we're talking about a specific file usually named
- 4:42design .md. This mundane little text document
- 4:46becomes Claude's ultimate design c***. Wow. It
- 4:49means your brand's visual identity is no longer
- 4:52locked inside proprietary software. It isn't
- 4:54trapped in a messy Canva template or, you know,
- 4:57a complex Figma file. It lives as raw code. I
- 5:00was reading through the awesome designer AI examples
- 5:02in the source, and it is fascinating how they
- 5:04break down major brands. Stripe is a great example
- 5:07to unpack here. Oh, Stripe is perfect. Stripe
- 5:10is widely considered the gold standard for clean
- 5:12sauce design. Stripe's design language is entirely
- 5:14focused on projecting trust and technical competence.
- 5:18Their plain text system doesn't just say, look
- 5:20trustworthy. It defines specific, measurable
- 5:23rules. It dictates crisp pricing layouts. Right.
- 5:27It enforces rigid rules around using massive
- 5:30amounts of white space to let the content breathe.
- 5:33Then you look at Spotify's rules, which are fundamentally
- 5:35opposed to Stripes. Spotify's text file dictates
- 5:39deep, rich, dark themes. Completely different
- 5:42vibe. It forces the AI to use bold, vibrant neon
- 5:45accents that pop against that dark background.
- 5:48The rules actively encourage heavy visual contrast.
- 5:52You also have SpaceX in that repository. Their
- 5:55brand is about scale. and drama so their text
- 5:58rules demand cinematic layouts the markdown dictates
- 6:02sweeping product sections that stretch across
- 6:04the entire viewport an apple system is predictably
- 6:07minimal the text file aggressively enforces a
- 6:10product first hierarchy The typography rules
- 6:13are incredibly rigid, stripping away anything
- 6:15that distracts from the core device being showcased.
- 6:17And Claude Ries' design .md file internalizes
- 6:20these strict constraints and then actually builds
- 6:22functional assets. It's wild. It writes the code
- 6:24for fully responsive pricing pages. It generates
- 6:27perfectly formatted HTML slide decks. It builds
- 6:30interactive website sections that perfectly match
- 6:32the Codify brand. I do have to push back a little
- 6:35bit on the mechanics here, though. I am slightly
- 6:37skeptical that a plain text document can truly
- 6:40capture an abstract vibe. Take SpaceX's cinematic
- 6:43feel. Okay, yeah. How does Claude know what cinematic
- 6:47actually means just from reading text without
- 6:50any visual aids? Claude already has broad training
- 6:53on those concepts. The text just sets the strict
- 6:56boundaries and constraints for it to apply that
- 6:58training. Got it. Text defines boundaries and
- 7:01Claude fills in the creative execution. That
- 7:04is the perfect way to summarize it. Knowing these
- 7:07pre -made templates exist in a GitHub repository
- 7:09is fantastic. But relying on someone else's templates
- 7:13isn't enough if you want to build a unique brand.
- 7:16You have to build a custom system for your specific
- 7:18workflow. Building your own system is where the
- 7:21real leverage is. The source outlines a very
- 7:23precise five -step framework to codify your own
- 7:26work. Step one is choosing the format. Right.
- 7:30You have to be strategic here. You must pick
- 7:31the task with the highest return. You shouldn't
- 7:34pick the most exciting task or the one you enjoy
- 7:36doing the most. You pick the bottleneck. Exactly.
- 7:38If you're running a YouTube channel, your biggest
- 7:40bottleneck might be creating HTML slide decks
- 7:43for your educational videos. So you codify that.
- 7:47Yeah. If you run a creative agency doing client
- 7:49work, your highest return task is probably generating
- 7:52pristine client proposals. You build the system
- 7:55around the task you repeat constantly. Once the
- 7:58format is chosen, step two is adding your tools
- 8:01and integrations. Crucial step. Claude shouldn't
- 8:03be forced to guess what your brand looks like.
- 8:06You have to feed it context from your actual
- 8:08business operations. You plug it into your existing
- 8:11ecosystem. You connect it to your GitHub repositories.
- 8:14You link it to your Gmail to understand your
- 8:16communication style. You feed it massive amounts
- 8:18of meeting transcripts. Whoa, imagine generating
- 8:21hundreds of perfectly branded slide decks just
- 8:23from meeting transcripts. You finish a one -hour
- 8:26Zoom call, and the system automatically outputs
- 8:29the client presentation. That is the reality
- 8:32of this. workflow. But to get that level of accuracy,
- 8:35the source highlights a critical tool for this
- 8:37step called FireCrawl. Let's define that quickly
- 8:39for the audience. FireCrawl is a tool that copies
- 8:42website data into text for AI. It acts as a bridge.
- 8:46It crawls your existing live website. It pulls
- 8:49down your exact logo files. It maps out your
- 8:52entire DOM structure. Wow. Yeah. It extracts
- 8:55your precise brand colors and topography scales.
- 8:58Yeah. It basically turns your live website into
- 9:00data Claude can read. Which leads us directly
- 9:03into step three, providing references. Tools
- 9:06like Firecrawl give the AI factual information.
- 9:09But references give the AI taste. Right, because
- 9:12information isn't enough to make something beautiful.
- 9:14You have to show Claude what your version of
- 9:17excellent looks like. How do you practically
- 9:19do that within a text -based workflow? The source
- 9:22mentions using mid -journey. Yeah, you can generate
- 9:24conceptual images in mid -journey that capture
- 9:26your desired aesthetic. You feed those images
- 9:28into Claude. Claude's vision model analyzes the
- 9:32flat image. It maps out the spatial relationships,
- 9:34extracts the implicit color hierarchy, and translates
- 9:37those visual cues into... to structural code
- 9:39rules. You can also use your own historical assets.
- 9:43You feed it your best performing Canva templates
- 9:45or past agency designs. You are giving it a visual
- 9:49target to aim for before it writes a single line
- 9:52of code. Which brings us to step four, generating
- 9:55the first version. The crucial mindset shift
- 9:58here is that you are only aiming for an 80 %
- 10:01draft. Because expecting perfection on the first
- 10:04output will just lead to frustration. Exactly.
- 10:07The goal of step four is purely structural. You
- 10:10want the layout to be logically sound. You want
- 10:12the typography to match your brand rules. You
- 10:15want the primary colors placed correctly. You're
- 10:17just trying to escape the terror of the blank
- 10:19page. A flawed 80 % draft gives you something
- 10:22tangible to react to. Reacting to a draft is
- 10:25always faster than creating from scratch. That's
- 10:27why step five is so important. Step five is refine
- 10:30and save as a skill. This step is where the actual
- 10:33magic of codification happens. The feedback loop.
- 10:36But you can't use traditional client feedback
- 10:38here. You can't just tell Claude, make it pop
- 10:41or make it feel more dynamic. No, vague feedback
- 10:44breaks the system completely. Your feedback must
- 10:47be hyper -specific. You have to speak in structural
- 10:50rules. Instead of saying, it looks too busy,
- 10:52you tell Claude, limit the layout to three hex
- 10:55codes. Instead of saying make it smoother, you
- 10:58command it to use slower motion on the CSS transitions.
- 11:01You're literally tweaking the machine's parameters.
- 11:04Reduce padding by 20 pixels. Increase the contrast
- 11:07ratio on the subheaders. Like stacking Lego blocks
- 11:10of data. Yes. And once you dial in those parameters
- 11:13and the output looks perfect, you save that entire
- 11:16prompt chain as a reusable workflow. It becomes
- 11:19a permanent skill in your library. Beat. I am
- 11:22curious about step two again, though. Why is
- 11:24Firecrawl necessary if you can just attach a
- 11:26screenshot of a website? Well, screenshots are
- 11:28totally flat. Firecrawl pulls the exact text
- 11:31codes, typography, and HTML stature. Ah, Firecrawl
- 11:34pulls exact code and colors, not just a flat
- 11:37image. Exactly. That underlying data makes the
- 11:40whole system robust. Sponsor. All right, let's
- 11:43zoom out now. We've mapped out this exact step
- 11:45-by -step framework for codifying taste. What
- 11:48does this mean for the industry at large? We
- 11:50are witnessing a fundamental paradigm shift.
- 11:53The source material refers to this as the code
- 11:56first design future. Code first design. That
- 12:00phrase carries a lot of weight. It changes the
- 12:03entire objective of creative work. Under this
- 12:06model, you aren't actually making a design. You
- 12:08are building the machine that makes the design.
- 12:10Right. You spend your energy defining the rules
- 12:13once, and then you let Claude execute the layout
- 12:15generation over and over again at near zero marginal
- 12:18cost. Let's break down the impact on different
- 12:21professions. For independent creators and startup
- 12:24founders, this shift is incredibly empowering.
- 12:27Historically, custom design has always been a
- 12:29massive bottleneck. It's expensive, it's slow,
- 12:32and it requires specialized talent. With a codified
- 12:35system, founders can launch new products infinitely
- 12:37faster. They maintain impossibly clean, consistent
- 12:41visuals across all their platforms. And they
- 12:43save thousands of dollars on routine, one -off
- 12:45design tasks. What about developers? This seems
- 12:48like a massive structural change for engineering
- 12:50teams. Developers no longer need to spend years
- 12:53mastering the intricacies of Figma to build beautiful
- 12:56products. But they do have to acquire a new hybrid
- 12:59skill. The source describes it as coding plus
- 13:03taste. Coding plus taste. They need to understand
- 13:06the underlying principles of visual hierarchy.
- 13:09They have to learn how to articulate design critique
- 13:12technically. A developer can't just look at a
- 13:15generated layout and say, make it look better.
- 13:18Right. They need the vocabulary to say, use stronger
- 13:21typographic hierarchy here or increase the negative
- 13:24space around the primary call to action. They
- 13:26act as the art director, guiding the creative
- 13:29direction, while Claude acts as the production
- 13:31designer, writing the actual code. And what is
- 13:34the impact on larger established teams? This
- 13:37might be the most practical application of all.
- 13:39Think about traditional corporate brand guidelines.
- 13:42Usually a company spends $50 ,000 on a rebrand.
- 13:45And they get a beautifully designed PDF that
- 13:47sits completely dead on a shared hard drive.
- 13:49Exactly. Nobody ever actually follows the PDF
- 13:52rules when they are rushing to finish a sales
- 13:54deck on a Friday afternoon. That is the reality
- 13:56of corporate design. But Clawed Code takes those
- 13:59dead PDF guidelines and turns them into living,
- 14:03usable engines. You transform a static reference
- 14:06document into an active system that automatically
- 14:09generates your daily deliverables. It builds
- 14:11your internal dashboards, your client proposals,
- 14:14your quarterly reports. It enforces strict brand
- 14:16consistency automatically without requiring a
- 14:20senior designer to review every single slide.
- 14:22But I do have to push back on behalf of professional
- 14:24designers here. If the machine handles all the
- 14:28execution and the guidelines are enforced automatically,
- 14:30does human taste actually matter anymore? Yes,
- 14:33the creative work is still there. You still need
- 14:36judgment and taste, but now taste becomes a system.
- 14:39So where does Figma fit into a world where Claude
- 14:42does all the heavy lifting? Figma is simply the
- 14:44final polished step. It is definitely not the
- 14:46starting line anymore. Exactly. Use Claude for
- 14:48the fast draft, Figma for the pixel perfect finish.
- 14:51That hybrid approach is how the best teams will
- 14:54operate moving forward. Two sec silence. This
- 14:57has been a thoroughly fascinating shift in perspective.
- 15:00The core truth we are uncovering is that good
- 15:04design doesn't have to be this ethereal, abstract
- 15:07concept trapped inside a creative director's
- 15:10head. No, it doesn't. It doesn't have to be stuck
- 15:12in a massive, disorganized Figma file full of
- 15:15disconnected layers. It is so liberating when
- 15:17you realize those visual rules can be extracted.
- 15:20They can be codified. They can be written down
- 15:22in plain text. They can be turned into a highly
- 15:25repeatable system that scales your specific visual
- 15:27identity. A landing page isn't just a page anymore.
- 15:31It is a localized output of a much larger system.
- 15:34That systemic thinking is the future. It leaves
- 15:37me with one final, slightly provocative thought
- 15:39to consider. Let's hear it. If our current modern
- 15:42visual taste can be perfectly codified into Markdown
- 15:45and perfectly replicated by Claude at scale,
- 15:47what happens to the value of current design trends?
- 15:50Will the most valuable design skill of tomorrow
- 15:52simply be having a wildly unique, undefinable
- 15:55perspective that an AI hasn't been trained on
- 15:57yet? That is a profound question to end on. It
- 16:01really challenges us to find the human element
- 16:03in our work. I encourage you to look at the very
- 16:06next document or presentation you create this
- 16:08week. Ask yourself, am I just doing a one -off
- 16:10chore or can I turn this into a system? Thank
- 16:13you for joining us on this deep dive.