Claude Design's 4 Tools That Change How You Build
Claude Design adds visual editing, annotation, and drawing tools to AI-generated assets. Here's what those four tools actually do—and where Claude Chat still wins.
Written by AI. Yuki Okonkwo
Photo: AI. Dante Nwosu
Editor's note: This article covers Anthropic's Claude Design tool—an AI and tech story. The category mismatch flagged in revision #1 is noted; this piece sits squarely in AI/Tech, not Business. Reassignment is not warranted.
There's a version of this story that writes itself: AI company ships design tool, creator makes carousel, everyone claps. But what Omar Farook actually documents in his latest video is more specific and more interesting than that. It's a live stress test of whether Claude Design's new layer of visual controls—four tools called Tweaks, Comment, Edit, and Draw—actually close the gap between "AI-generated" and "ready to post."
The short answer: mostly yes, with some friction. The longer answer is worth unpacking.
What Claude Design is (and isn't)
Before we get into the tools, some context. Claude Design is Anthropic's research-phase design environment. Unlike Claude Chat—where you prompt, get an artifact, and work through text—Design starts with a brand-specific design system. You feed it your brand information, it infers the relevant components (typography, color, spacing, iconography), and from there you generate assets that actually look like your brand rather than a generic AI output.
Farook built his design system around Granular, his cohort-based program. Claude inferred the brand from the materials he provided and generated a full system: UI marketing kit, slide templates, components, logo lockups. When he compared it to the design system it generated for his other product, Blitz, the difference was stark—different sections, different hierarchy, different visual language. The system isn't applying a template; it's interpreting brand context and making decisions.
That's genuinely interesting behavior. It's also slightly unnerving if you care about design consistency, because "it interpreted the brand" is not the same as "it followed the brand guidelines." But that's a separate conversation.
The more pressing question in this video is: once you have a generated asset, how much control do you actually have?
The four tools, explained honestly
Farook builds a carousel post from scratch—theme: "the wild west of Claude code skills," a.k.a. the chaos of every AI influencer telling you to install 20 different tools before you've even written a line of code. The first slide generates at 1080x1350, looks polished, captures the prompt's energy. Then he starts editing.
Comment is the most intuitive of the four. You select a specific element on the canvas, drop a comment, and Claude treats it as a targeted instruction. Farook used it to remove an "eyebrow" label he didn't want—"Why field notes? Is this necessary? If not, let's remove it"—and it worked cleanly. The key thing here is specificity. Instead of re-prompting the whole asset and hoping Claude figures out what you mean, you're pointing directly at the element. That's a meaningful UX improvement over chat-based iteration.
Edit is where it gets more like a real design tool. You can select elements and adjust font size, colors, tracking, margins—standard properties, dragged or typed. Farook spent a chunk of time here fighting with a top margin that was too large, nudging font sizes from 104px up to 200px (too big) down to 140px (better). His honest read: "The dragging value preview is a bit whack, but it works." That's the kind of candid assessment that actually helps people decide whether to use a tool.
The interesting tension he surfaces is that Edit and Comment work best together. He'd use Edit to manually adjust values, then switch to Comment when he hit something he couldn't fix directly—like a layout overlap where the stats block was covering subtext. Some problems are better solved by telling Claude what's wrong; others you just want to fix yourself. Having both options without leaving the environment is the actual value proposition.
Tweaks surprised him. He'd been thinking of it as a prompting shortcut, but after asking Claude to fix a slide, he discovered it had generated a custom controller panel—sliders for header size, top padding, stack gap—specific to that slide's structure. "Custom controls for your specific design asset," he said. "That is amazing." It didn't work perfectly (the stack gap slider apparently did nothing), but the concept—Claude generating its own UI controls based on what it built—is genuinely novel. You're not editing code or re-prompting; you're using a purpose-built panel that Claude created for this asset.
Draw is the wildcard. It's an annotation layer: you sketch or write directly on the canvas, and Claude uses that as context for its next edit. Farook circled a left sidebar element he wanted removed, typed "remove these left bars," sent it, and Claude deleted them. Then he pushed it further—sketching a rough gradient shape and describing it as a radial gradient background treatment. Claude built it. Not approximately. It nailed it.
"Not only can you use annotation to point at things, give it more context and make adjustments, but you can also use it to annotate and sketch your ideas," Farook observed. That's a meaningful extension of the input modality. You're not describing spatial intent in text; you're showing it.
The Claude Chat question
The video's most useful section is the comparison Farook draws at the end, because he resists the easy take. He doesn't declare Claude Design the winner. He doesn't dismiss it as beta noise. He actually thinks through when each tool is the right one.
His case for Claude Chat: it holds full conversation context. You can brain-dump, research, iterate, and then say "okay now build the asset"—and Claude builds it with everything you've discussed baked in. Design systems, project memory, free-form reasoning. It's a complete workflow for people who think out loud.
His case for Claude Design: it has the visual control layer Chat simply doesn't. "These are powerful tools to get more control in the assets that you create—which Claude Chat does not have." For assets where you know what you want and need to get it pixel-close, Design wins.
The honest tension he lands on: these two tools should probably talk to each other. "I do also hope that Claude figure out a way to then use sort of Claude Projects as maybe your brain that could then connect up into a Claude Design project to execute an asset." Research in Chat, execute in Design. Right now that handoff is manual. Whether Anthropic closes that gap is the real product question hanging over this whole demo.
What's still rough
A few things Farook flagged that are worth noting for anyone considering jumping in:
Claude Design runs on a separate credit limit that depletes faster than Chat. He hit it after his first session and had to wait seven days for a reset. For production use, that's a real constraint.
Export options defaulted to PPTX or PDF rather than PNG, which he would've preferred for social posting. Small thing, but tells you something about where the tool's assumptions are.
Version loading is manual—he had to click "load new version" after each edit rather than seeing changes update automatically. Minor friction that adds up across a lot of iterations.
And the category organization in the generated design system wasn't quite right for his use case. "Backgrounds being in spacing just feels a bit off," he noted. Adjustable through prompting, apparently, but you'd want to know that upfront.
None of these are dealbreakers. All of them are what you'd expect from a tool in research beta. The question is whether you're willing to work with rough edges in exchange for genuine capability—and that's a question only you can answer based on what you're trying to build.
The more interesting question, to me, is what it means when an AI tool starts generating its own UI controls based on what it produced. Tweaks doing that—even imperfectly—points toward something. Design tools have always been about the interface between human intent and visual output. Claude Design is trying to make that interface adaptive rather than fixed. Whether "adaptive" ends up meaning "more powerful" or "less predictable" probably depends on how much you trust the model's interpretation of your intent.
We don't have enough data to answer that yet. But it's the right question to be asking.
By Yuki Okonkwo, AI & Machine Learning Correspondent
AI Moves Fast. We Keep You Current.
Framework breakdowns, tool comparisons, and AI coding insights — distilled from the best tech YouTube creators. Free, weekly.
More Like This
Pencil.dev Promised Design-to-Code Magic. Here's Reality
AI LABS tested pencil.dev's design-to-code workflow and found it wasn't automatic. Here's what they built to fix it and what it means for AI design tools.
NVFP4 vs INT4: The Quantization Format That's 27% Faster
Nvidia's NVFP4 quantization outperforms traditional INT4 by 27% while maintaining quality—but the real story is what this reveals about benchmarking.
Anthropic's Claude Design Tool: What Actually Changed
Anthropic released Claude Design for UI prototyping. We tested it to see if it escapes the 'vibe-coded' look that plagues AI-generated interfaces.
ChatGPT vs Claude: The Visual Explainer Battle Nobody Saw Coming
OpenAI and Anthropic released competing visual tools within 48 hours. We tested both—one's faster, one's smarter, and the differences matter.
Claude Design Isn't Coming for Figma—It's After Something Else
Anthropic's new design tool targets a different workflow than established players. Early users reveal what it's actually good at—and the hard limits.
Apple's New CEO Inherits a Paradox: Did Doing Nothing Win AI?
John Ternus takes over Apple amid questions about whether the company's AI inaction was genius or fumble. Plus: Google forms a coding strike team.
Claude Opus 4.7 Promises Coding Dominance—With Caveats
Anthropic's Claude Opus 4.7 crushes coding benchmarks and builds impressive demos, but token consumption and quirks suggest the 'best' model depends on context.
Harness Engineering: The New Frontier in AI Development
AI companies are shifting focus from better models to better infrastructure. Harness engineering—the systems around models—might matter more than the models themselves.
RAG·vector embedding
2026-05-17This article is indexed as a 1536-dimensional vector for semantic retrieval. Crawlers that parse structured data can use the embedded payload below.