How We Use Claude Code + Webflow MCP to Ship Faster (Without Losing Taste)

Claude Code is Anthropic's terminal-based coding agent. Webflow MCP is Webflow's official Model Context Protocol server, which lets Claude read and write to a Webflow project through the Data API and Designer API. Together they let a designer hand off the repetitive parts of a build — bulk CMS updates, schema generation, class refactors — and keep their hands on the parts that actually require judgment.

"Will AI Replace Designers?" Is the Wrong Question

We get asked some version of this every other week. Usually by a founder who just watched a Twitter demo of someone generating a landing page in 30 seconds.

The honest answer: no, and the question itself is the problem. The better question is — how do good designers use AI to ship better work, faster? That's the one we've been answering for ourselves over the last 18 months, quietly, inside real client projects at Digi Hotshot.

This post is our honest attempt at showing what that looks like. Not a demo. Not a manifesto. Just how the work actually happens when you take AI seriously and taste even more seriously.

Taste Is the Moat

"Taste is the moat" isn't our phrase. It's been kicking around the design community for a while — design engineers have written about how small teams of people aligned on taste consistently out-ship large teams that aren't. Linear's design team talks about it. Julian Lehr's essays circle it.

So what does taste actually mean? It's the bundle of judgment calls that most of us can feel but can't always name.

It's knowing when a header should say one thing instead of three. It's noticing that the whitespace on the left of a hero isn't symmetrical with the right — and deciding that's the correct answer, not a mistake. It's cutting the third CTA because it dilutes the second one. It's the restraint to leave a section almost empty because the page needs a breath before the pricing table. It's picking a typeface that says "serious infrastructure" instead of "friendly startup" because the buyer is a VP of Security and not a founder.

Now contrast that with what most AI-generated landing pages look like today. Same purple-to-blue gradient hero. Same floating chat bubble in the bottom right. Same "Introducing [Product] — the [adjective] way to [verb] your [noun]" headline. Same three-column feature grid with the same three icons from the same icon library.

This isn't an accident, and it isn't malice. It's math. Every general-purpose AI tool is trained on roughly the same distribution of the public internet. When you ask a model to generate "a SaaS landing page," it gives you the statistical center of every SaaS landing page it's ever seen. Regression to the mean is a structural feature of how these systems work, not a bug you can prompt your way out of.

The net effect: if you build a site using only AI, your site looks like every other site built using only AI. That's not a moat. That's the opposite of a moat.

Taste is the part the model can't learn from a generic training set. It's the part a real designer brings to the table.

Where AI Actually Speeds Us Up

We use AI every day at DH. Not as a gimmick, not as a toy, and definitely not to replace anyone on the team. It's a speed layer. Here's the pattern — not the recipe — for where it earns its keep:

• Bulk CMS operations. When a client has a 50-item collection that needs a new field populated across every entry, or 12 case studies that need a new schema pattern, Claude Code + Webflow MCP handles it in minutes. A designer would hate doing that by hand, and rightly so.
• Long-form draft generation. Blog posts, case studies, FAQ sections. We draft the structure, AI gets a first pass down, a human editor rewrites it in the client's voice, a second human checks it against our voice rules. The AI draft is scaffolding, not a finished product.
• Schema markup generation. Every blog post we publish ships with JSON-LD for FAQPage and BlogPosting. Generating those by hand is tedious. Generating them with AI against a fixed template is a ten-second job.
• First-pass class naming on new builds. We have an internal class naming convention. When a designer finishes a new section, we can get a clean first draft of class names in seconds and review from there.
• Spec-to-structure conversion. A Figma spec becomes a structured content outline. That structured outline becomes a CMS schema. The middle step — which used to eat an hour — takes five minutes.
• Repetitive refactors. Renaming a class across 400 elements. Updating a link pattern across 80 pages. Migrating an old component to a new version. These are the jobs nobody on the team wants. AI is excellent at them.

The through-line: every one of these is a job a human would do badly because they'd get bored. Boredom is the enemy of quality on repetitive tasks. AI doesn't get bored.

Where AI Can't Help (And Shouldn't)

Here's the other half of the honest answer. There's a long list of decisions that AI still shouldn't touch. Not because the tool isn't capable of producing output for them — it is — but because the output will be wrong in ways that matter.

• Hierarchy calls. What's the first thing the eye should land on? Why? An AI can guess. A designer knows.
• Typography decisions. The difference between a page that feels fast and expensive and a page that feels like a template is often one font choice and one line-height decision.
• The decision about what to leave out. This is the hardest call in design. AI wants to include everything it's been asked about. Good designers cut.
• Brand voice. A model that's read the whole internet cannot tell you what a specific founder sounds like in a real Slack message.
• Section transitions. The half-second rhythm between one section and the next. The thing you feel when you scroll a well-built site and can't tell why it feels right.
• When to break the grid. Rules exist so you know which ones to break. That's a taste call.
• The difference between 80% and 100%. Anyone can get to 80. The last 20 is where design lives.

If you squint, all of these are the same thing: they're judgment calls about restraint, rhythm, and honesty. They're the things that separate a site that looks expensive from a site that looks generated.

Webflow MCP: What It Is, What It Does

Quick primer for readers who haven't touched it yet.

Webflow MCP is Webflow's official Model Context Protocol server. Model Context Protocol, if you're new to it, is an open standard that lets an AI client (like Claude, or any MCP-compatible tool) connect to external systems in a structured way. Instead of the AI guessing at API calls, the MCP server tells it exactly what operations are available and how to use them.

For Webflow specifically, the MCP server exposes two sets of tools. One wraps the Data API — CMS collections, items, assets, SEO metadata, publishing. The other wraps the Designer API — creating sections and grids, applying styles, managing design variables, building components.

On top of that, Webflow and Anthropic shipped an official Webflow connector for Claude in February 2026. That connector gives Claude direct, authenticated access to a Webflow project so a marketer can, for example, run a site audit for broken links or missing alt text without leaving their Claude window.

Translation into plain English: Claude can now talk to your Webflow site. It can read it. It can write to it. It can publish changes. With the right guardrails, it's a legitimate force multiplier.

Claude Code + Webflow MCP in Practice

Here's a generic example so you can picture it. Names changed, details blurred, the shape of the work is real.

A client needs a new service page. The designer opens Webflow and builds the page the way they'd build any other page. Typography decisions, hierarchy calls, the image treatment, the rhythm between sections, the headline rewrite, the single CTA at the bottom instead of three. That part takes what it always takes — a couple of days of real design work.

Once the page is approved, the repetitive work kicks in. The FAQ needs JSON-LD schema. There are four related case studies that need to be linked through a CMS reference field we haven't used on this page type before. The design system has a new spacing token that needs to be applied consistently across a dozen similar pages. The client's marketing lead wants the same page translated into three variants for different industries.

That's the part where Claude Code + Webflow MCP earns its keep. The schema gets generated and embedded. The CMS references get wired up. The spacing token gets applied across 12 pages in a single pass. The variants get drafted, then handed back to a writer to edit in the client's voice.

Work that used to eat a full day now takes about an hour. But — and this is the whole point — the taste calls on the original page still took the two days they always took. We didn't speed those up. We didn't try to. The speedup happened in the repetitive layer, where speed actually helps.

Taste stays slow on purpose. Everything around it gets faster.

How to Tell a Design Was Made With Taste vs Generated

When we audit other people's sites — ours and prospects' — there are tells.

• Whitespace that isn't symmetrical, and is clearly on purpose.
• A hierarchy that surprises you a little. The thing you expected to be the biggest isn't.
• A header that says one thing instead of three. No "flexible, powerful, intuitive" triples.
• Restraint with illustrations. Not every section needs a custom spot illustration.
• One or two intentional rule-breaks. A section that bleeds to the edge. A bold color where nothing else is colored.
• Writing that sounds like a person. You can tell it was typed by someone with an opinion, not generated by someone who had to fill a box.
• No purple-to-blue gradient. Or if there is one, it's being used ironically.
• The absence of the floating chat bubble. Or if there is one, it's earned its place.

If you see a site that hits most of those, a person with taste was in the room. If you see a site that hits none of them, a model was in the room alone.

A Word About Webflow's New AEO Product

On April 13, 2026, Webflow announced Webflow AEO — a closed-loop agentic answer engine optimization product for Enterprise customers. The product does three things: it measures how often a brand shows up in AI answer engines, it recommends technical and content fixes that are specific to that brand, and it ships those fixes at scale with a review-before-publish step so a human is always in the loop.

The signal for us isn't the feature list. It's the design of the thing. Webflow is building AI into the platform at the operating layer, and they're keeping a human in the approval loop by default. That's the same pattern we've been using internally. AI does the repetitive execution. A person with taste decides what actually ships.

It's nice to see the platform we bet on five years ago arriving at the same conclusion.

A Position, Not a Preach

AI is a speed layer for people with taste. It isn't a replacement for people with taste. We use Claude Code, Webflow MCP, and our own internal tooling on most builds now — and we still have designers making the calls that no model can make.

That's the whole post.

FAQ

Does AI replace designers?

No. AI replaces some of the repetitive tasks designers used to do manually — class renaming, bulk CMS updates, schema generation, first-draft content. It doesn't replace the judgment calls that make a design feel considered: hierarchy, typography, restraint, brand voice, and the decision about what to leave out. Good designers use AI as a speed layer for the boring parts and spend the saved time on the parts that actually need taste.

What is Webflow MCP?

Webflow MCP is Webflow's official Model Context Protocol server. It lets AI clients like Claude connect directly to a Webflow project through the Data API (for CMS, assets, SEO, and publishing) and the Designer API (for sections, styles, variables, and components). An AI agent can read from and write to a Webflow site in a structured, authenticated way instead of guessing at API calls.

Can Claude Code update a Webflow site?

Yes. With the Webflow MCP server connected, Claude Code can create and update CMS items, adjust styles and variables, audit content for issues like broken links or missing alt text, and publish changes. Webflow and Anthropic also shipped an official Webflow connector for Claude in February 2026 for marketers who want the same access without running Claude Code in a terminal.

Why do AI-generated websites look the same?

Because every general-purpose AI model is trained on roughly the same distribution of the public internet. When you ask a model to generate a SaaS landing page, it produces the statistical center of every SaaS landing page it has ever seen — same gradient hero, same feature grid, same headline structure, same icon library. Regression to the mean is a structural feature of how these systems work, which is why sites built end-to-end by AI start to feel interchangeable.

What is taste in design?

Taste is the bundle of judgment calls that separate a considered design from a generic one. It shows up in hierarchy decisions, typography, whitespace, restraint with illustrations, knowing when to break the grid, knowing what to leave out, and writing copy that sounds like a person instead of a template. It's the part of design that doesn't come from rules, and it's the part a general-purpose AI model can't replicate.

How does Webflow AEO work?

Webflow AEO is a closed-loop agentic Answer Engine Optimization product that Webflow launched on April 13, 2026, for Enterprise customers (currently in private beta). It measures how often a brand shows up in AI answer engines, surfaces brand-specific recommendations for technical and content fixes, and ships those fixes at scale with a human review step before publish. It's Webflow's bet that AI belongs inside the platform as an execution layer, with a designer or marketer still approving every change.