Claude Code + Klaviyo's MCP: 9 use cases we built for 15 clients, without engineers
Generated by Klaviyo AI
At Restore.ai and Marketech, a non-technical marketing team runs Klaviyo for 15 clients using Claude Code and Klaviyo's MCP server, without any engineers or dev shop involved. The post walks through 9 use cases split into a 4-step content calendar pipeline and 5 standalone prompts, showing how each replaced work that used to require manual data pulls, spreadsheets, or QA time.
- A repeatable content pipeline: The first four use cases turn campaign and subject-line performance data into content pillars and angle categories, then run product sell-through analysis and a "divergent thinking" step to keep ideas from all looking the same — taking the team from a blank calendar to around 50 concepts they narrow down to roughly 15 sends.
- Standalone prompts replace manual reporting: Five additional prompts handle A/B test analysis, pop-up funnel reporting, detecting double sends across overlapping flows, staging drag-and-drop campaigns from Figma, and building brand identity briefs — work that previously meant exporting Klaviyo data, building spreadsheets, and doing manual QA.
- The team's role shifts from doing to managing: With assembly work automated, the team spends less time on manual reports and more time deciding inputs and reviewing output quality — a shift the post credits with letting a 15-client, no-engineer team take on more clients and more reporting than before.
At Restore.ai, a first-party revenue engine for Klaviyo, and Marketech, a data-driven and engineering lifecycle marketing agency for mid-market brands, we have no dev shop, no technical co-founder, and no sprints.
Every use case I’m about to describe was built by a non-technical marketer. All of it lives in Markdown files. And all of it runs through Claude Code via Klaviyo's MCP server.
If you’ve been hearing people talk about vibe marketing and AI-powered marketing programs and feeling like it’s a thing that happens to other teams, this blog is for you. The on ramp is lower than it looks.
Here’s the full stack of what we’re running, use case by use case.
Use cases 1–4: The content calendar pipeline
The first 4 use cases are a pipeline. Together, they take us from a blank content calendar to 50 rough ideas we can cut down to 15 messages we actually want to send.
The first 3 address what a marketing agency should be talking about for a given brand in a given month, while the last one takes that output and deliberately blows it open.
Each one is a prompt, so you don’t need to wire anything together. You run one, take the output, and feed it into the next.
AI marketing tip: Check out hundreds of tried and tested AI prompts in Klaviyo’s AI prompt library.
1. Content pillar analysis
We start by uploading 3 years of email campaign data into Claude with a prompt that asks it to identify 5–6 content pillars for the brand. From there, it categorizes every past campaign send by pillar, then runs the analysis: which pillars have the highest open rates, the highest click rates, the highest revenue per recipient.
The output tells us what to prioritize in the calendar. If educational content is performing twice as well as product highlights, we send twice as many educational emails. We use performance to guide our content mix: the highest performer will have the most campaigns, and the lowest the least, but every pillar is still represented every month.
And because every brand is different, the pillars are always different, too.
2. Subject line angle analysis
This is the same process as the one I describe above, but scoped to subject lines and preview text. Instead of going in with a fixed list of angles (e.g., urgency, promotion, social proof, etc.), we have Claude generate the angle categories dynamically for each brand.
Claude looks at the full subject line history, surfaces 5–6 categories that actually fit how that brand communicates, and then runs the analysis to show which angles drive the best results.
The key here is the dynamic categorization. A static list of angles often doesn’t fit. Letting Claude define the categories from the data means the framework matches the brand, not the other way around.
The key here is the dynamic categorization. A static list of angles often doesn’t fit. Letting Claude define the categories from the data means the framework matches the brand, not the other way around.
3. Product and category sell-through rates
The prompt we built for this approach pulls all product purchase events and calculates revenue per product per month. The goal is to find which products and categories are at their peak right now, relative to how they perform all year.
A real example: for a client that sells bike parts, Claude flagged that June is the highest month for floor pumps as a category. Even though floor pumps aren’t the brand's biggest revenue driver overall, or even the biggest revenue driver in June, it’s by far that category’s best month. So we prioritized floor pump-focused emails, and saw those significantly outperform differently-themed emails.
This type of analysis would have taken a lot of time prior to Claude, and would’ve been buried in the 15–20 different data frameworks we can now analyze on demand. With Claude, we can scan all our data and surface only the most actionable insights.
4. Divergent thinking for marketing angles
Here’s the problem the first 3 steps create: if you hand Claude a tight brief, telling it which pillars work, which angles work, and which products to focus on, it will produce the same content every single time. You’ve built a box, and the AI will stay inside it.
The solution is to force divergent thinking after the convergence step.
We do this in a few ways:
- We take 8 broad marketing angles and ask Claude to generate 3 email concepts under each one.
- We take a set of viral Facebook hook formats and repurpose them for email: 6 hooks, 3 concepts each.
- We generate roughly 50 campaign ideas and expect to like 14–15 of them.
The first time we tried this approach, we averaged 14 concepts that our team was excited about and that we didn’t need to rewrite from scratch.
The solution is to force divergent thinking after the convergence step.
Use cases 5–9: Standalone Klaviyo MCP prompts
These next 5 aren’t connected to the pipeline above. Each one is a standalone prompt you can pick up and run on its own.
5. A/B test analysis
Of course, we don’t A/B test a single send because the data is too thin. Between machine opens, sample sizes, and other noise, one send won’t tell you enough. So we test the same variant across multiple sends over a month, using a consistent naming schema to tag the variants.
Normally, pulling that analysis means building a report in Klaviyo, exporting it, opening Google Sheets, parsing the variants, and building pivot tables. We replaced all of that with a single Claude prompt. It automatically categorizes the variants, runs the analysis, and outputs a structured result. The whole process is one step instead of 5. Our team used to spend 540 hours a week between downloading and uploading Klaviyo exports to Google sheets and creating data visualizations. With all of that automated, we’ve been able to take on double the clients and 2–3x the amount of data visualizations we can deliver.
6. Pop-up funnel reporting
Our next use case is the same idea, applied to the pop-up unit. One Claude prompt pulls the pop-up funnel report and surfaces what’s working and what isn’t, without a manual export.Once we got this in place, we could finally get a time series of pop-up data in our reporting.
7. Double send detection
This one came out of a specific problem: we run an identity resolution product alongside Klaviyo, which means some clients end up with multiple abandoned cart flows running in parallel. The obvious risk is double sending.
So we built a Claude prompt to detect it automatically. It flags users who receive two sends in a short window across any combination of flows.
What we found surprised us. Our own flows were clean. But when we ran it on clients who came from other agencies or in-house set-ups, a lot of them had issues. One had 10% of subscribers receiving both an abandoned check-out and an abandoned cart email within an hour of each other.
Klaviyo gives you everything you need to prevent this. The problem is that with 10 or more exclusion rules across multiple flows, the mental map of every possible user path gets complicated fast. Our prompt does the audit automatically, with carve-outs for flows where double sending is intentional (e.g., transactional emails, post-purchase, etc.) on a per-client basis.
8. Staging email campaigns from Figma
This one just got a lot easier. In June, Klaviyo released the ability to stage drag-and-drop campaigns via the MCP, and it changed our process completely.
We now have a prompt that exports from Figma, builds out a set of drag-and-drop templates, uses those templates to create a campaign, inputs the right segments, and schedules it. Before this, the same process required ~30 minutes of hands-on work per campaign. Now, QAing the email campaigns takes us about 5 minutes.
For an agency running 15 clients, having Claude Code windows running in parallel for different accounts, each pulling from Figma and staging into Klaviyo, drastically changes the arithmetic of what a small team can handle.
And again: none of the people running those processes are engineers.
9. Brand identity research
A lot of brands doing less than $30 million in revenue know their brand conceptually but have never written that identity down anywhere. Their voice and positioning are baked into their website, their social, and their press coverage. But there’s no brand documentation.
We built a prompt that scrapes publicly available sources and assembles a structured brand intelligence brief: voice, messaging, product positioning, community sentiment. From that, Claude generates a brand guide the team can actually use.
Of course, it’s not a replacement for a proper brand strategy engagement. It’s a starting point that takes an hour instead of a week, and it’s the thing that makes every downstream prompt smarter. We never start implementing what's been auto-generated.
What changes when the assembly disappears
In all of these use cases, the tools matter less than the shift in how the team thinks about our work.
When you have a system doing the assembly, the person who used to write subject lines or pull reports all day is now managing a process. What are the right inputs? Are the instructions tight enough? Is the output quality where it needs to be? That’s a different job, and it’s a higher value one.
This is available to any marketing team right now. Not after you hire an engineer. Not after some future product release. You just need a Claude account, Klaviyo's MCP server, a boatload of curiosity, and the willingness to write a prompt and see what comes back.
Related content
