Across our design and engineering work, Claude has become less a finished tool and more a layer between a brief and whatever actually produces the output.
Adam Gudmundsson, our Head of Client Services, didn't open Midjourney, Figma or Shopify Sidekick when he wanted to test a design idea. He opened Claude first and asked it to write the prompt for him.
"I asked Claude to prompt Sidekick for me. Results may surely vary, but it's an interesting way to work when you're exploring what these tools can do together," says Adam, our Client Lead.
That's a small habit, but it points at something bigger about where Claude actually sits in a design workflow. It doesn't design. It reasons, briefs and, increasingly, builds, sitting as the layer between a strategic decision and whatever tool produces the visual or the code that follows it.
Claude is a large language model built by Anthropic. It processes and generates text and can reason across writing, analysis, coding and structured thinking. It does not generate images natively. It works in language: describing, briefing, structuring and critiquing, rather than producing a visual output directly.
That puts it in a different place to image generation tools like Midjourney or Adobe Firefly. Claude sits earlier, at the thinking and briefing stage, not the output stage. The line is blurring though. Claude can write prompts for image generation tools, structure a design system, generate front-end code from a brief and reason about layout, hierarchy and user experience in ways that are genuinely useful to a design team. With Claude Code, Anthropic's agentic coding tool, the gap between design intent and live implementation is narrowing further still.
Claude is also not a single fixed tool. Anthropic ships new versions with different capability profiles, and the gap between what Claude could do eighteen months ago and what it can do now is significant. A team that tried it early and found it limited is likely working from an outdated picture of what it can actually do today.
Brief writing is one of the most immediate use cases. Claude can take a set of inputs, brand guidelines, a campaign objective, a target audience, a handful of reference points, and turn them into a structured creative brief a designer or design tool can work from. For a team managing several clients at once, that compresses the time between a strategic decision and creative output.
The quality of the brief is directly proportional to the quality of the inputs. A vague set of inputs produces a vague brief. Specific brand guidelines, clear objectives and defined constraints produce something a design team can actually use. The discipline required is the same discipline that makes briefing a human designer effective. Claude just processes it faster and hands back a structured output immediately.
Adam's experimentation goes a step further than brief writing: rather than prompting an image generation or design tool manually, Claude acts as an intermediary, taking a higher-level brief and translating it into the structured prompt language that tools like Shopify Sidekick, Midjourney or Firefly respond to most effectively. Output quality from these tools is heavily dependent on prompt quality, and most teams underestimate how much a prompt's framing and specificity changes what a generative tool produces. Claude's ability to reason about what a tool needs, and to iterate on a prompt based on feedback, makes it a useful layer between a human brief and a generative tool. It removes friction from the exploration stage and lets a team move faster between concept and visual output, with a person directing at the strategic level rather than managing prompt mechanics.
The same language strength shows up in UX copy. Button labels, error messages, empty states, onboarding flows, product descriptions: Claude handles a solid first pass on all of it. UX copy is often deprioritised or left until late in a build, so having a tool that can produce a working draft quickly changes the order things happen in. Copy gets into the design earlier, and the visual and verbal layers of an experience develop together rather than one after the other.
There is a more direct integration worth separating out from the conversational tool: Claude Code. Where Claude reasons and advises, Claude Code acts. It reads a codebase, writes and edits files, and executes changes directly. That opens up a practical workflow for a commerce build: design a component or layout with Sidekick or another AI design tool, then use Claude Code to build and implement it in the actual storefront.
On Shopify, this has a reasonably high chance of working well. The theme architecture is well structured and widely documented, so Claude Code has strong reference material to draw on. The constraints are known, the patterns are established, and the implementation path is fairly predictable. On a headless or composable build, where the architecture is custom and there is no standardised framework to anchor to, the results are less reliable and need more developer oversight to validate.
That doesn't make Claude Code less useful on headless builds. It changes how it gets used. On Shopify, it can move with less supervision. On a custom composable stack, it works better as a starting point than a finished implementation, generating a scaffold a developer refines rather than an output that goes straight to production. Claude can also write HTML, CSS and React directly from a description or wireframe brief. A designer describes a product card, a navigation pattern or a checkout step, and gets a working scaffold a developer can refine rather than build from scratch, which shifts engineering time from writing to reviewing on most projects.
Claude does not generate images, illustrations or visual assets. It cannot replace a visual designer, a brand strategist or a UX researcher. It has no access to a brand's specific assets, guidelines or historical work unless they are given directly in the conversation. Every session starts from what it is handed.
Its output is a starting point, not a finished product. Design work produced with Claude's help still needs human judgment, creative direction and production craft to be commercially effective. Claude reflects back what it is given: a thin brief produces thin output, in exactly the way an over-relied-on template would.
It also has no visual perception in the way a designer reviewing a Figma mockup does. Claude can reason about a design from a written description, and multimodal capabilities that let it process images do exist and are improving, but the depth of that visual analysis still falls short of a human design review. Describing a design to Claude and asking for feedback is useful. Expecting it to catch visual inconsistencies or judge aesthetic quality at the level a senior designer would is not realistic yet.
The teams doing this well, ours included, share a few habits. They treat input quality as the main variable: the more context Claude gets, brand voice, target audience, platform constraints, design system conventions, the more useful the output. A team that briefs it vaguely gets a vague result and concludes the tool isn't useful. A team that briefs it the way it would brief a senior designer gets something worth working from.
They use it at the thinking stage, not the production stage. Claude is most valuable early, structuring a brief, exploring directions, generating options to react to. Pushing it to produce finished output directly tends to need heavy refinement afterwards. The earlier it enters a workflow, the more it shapes the direction rather than just filling in details.
They treat it as a layer between tools rather than a standalone solution. Adam's approach, using Claude to prompt Shopify Sidekick, is a good example of the pattern: Claude reasons well, a specialist tool executes well, and the combination beats either alone. Direction gets set by a person, Claude translates it into a prompt, the specialist tool generates the output, and a person evaluates it, faster and more consistently than any single-tool approach.
They also match the tool to the platform. On Shopify, Claude Code can move fast with less supervision because the architecture and documentation are already standardised. On a headless build, more developer review is needed to validate what it produces. And they iterate rather than expect a finished result from one prompt. Treating Claude as a conversation, refining, redirecting, pushing back on output that isn't right, produces noticeably better results than a single shot at it. The first output is a starting point for a dialogue, not a deliverable.
AI's integration into design workflows is early and moving fast. Claude, Midjourney, Firefly, Figma AI and Shopify Sidekick are not the finished version of what AI-assisted design will look like in two or three years. The workflows are still being worked out, by the tools themselves and by the teams using them.
What is already clear is that the design process won't look the same going forward. The time between brief and first concept is compressing. The cost of exploring directions is falling. The skill that matters is shifting towards creative direction, taste and judgment, knowing what good looks like and being able to steer AI tools towards it, rather than execution alone. A designer who can brief AI tools precisely and evaluate their output critically will be more valuable than one who can't, regardless of technical production skill.
For commerce brands, that matters because design is a commercial input. Better briefs produce better creative. Faster iteration produces more options to test. A component that moves from brief to implementation faster means less time between idea and live test. We're working this out in practice, alongside our clients, and if you want to talk through how AI tools fit into your own design and development workflow, get in touch.