Growth marketing might seem like an unexpected home for Claude Code. It's not a tool typically associated with marketers, and growth teams at most companies work primarily in campaign management platforms, analytics dashboards, and spreadsheets — not code editors.
But Anthropic's Growth Marketing function has built some of the most sophisticated Claude Code workflows in the company, using it to automate repetitive tasks at scale, build custom integrations that didn't exist before, and operate with the output capacity of a much larger team. For a non-technical team running a lean operation, Claude Code has become essential infrastructure.
Automated Google Ads Creative Generation
Performance marketing at scale involves managing hundreds or thousands of ad variations across campaigns. Keeping that creative fresh — replacing underperforming ads with better-performing alternatives — is a critical but enormously time-intensive task.
The Growth Marketing team built an agentic workflow to automate this entirely. The system ingests CSVs containing hundreds of existing ads along with their performance metrics. It identifies which ads are underperforming against defined thresholds. Then it generates new variations to replace them, working within the strict character constraints Google Ads enforces: 30 characters for headlines, 90 characters for descriptions.
The architecture uses two specialized sub-agents — one focused on headlines, one on descriptions — because the creative requirements and constraints are different enough that specialized agents outperform a single generalist agent. Headlines need to be punchy and keyword-relevant within 30 characters. Descriptions need to elaborate on value propositions within 90.
The result: what used to take a marketing team member 2 hours to do manually — analyzing performance data and generating replacement ad copy — now takes 15 minutes. Hundreds of new ads generated in a single workflow run.
Figma Plugin for Mass Creative Production
Once you have strong ad copy, you still need to produce the visual ad assets — and for display advertising, that means generating multiple variations with different copy combinations to test.
The team built a custom Figma plugin using Claude Code that programmatically generates up to 100 ad variations by swapping headlines and descriptions into a template. What previously required manually opening a design file, editing text layers one by one, and exporting each variation — a process that could take hours for a large batch — now takes half a second per batch.
This 10x increase in creative output isn't just a time saving. It changes what's possible to test. When generating 10 variations requires hours of manual work, you're conservative about what you test. When generating 100 variations takes seconds, you can run genuinely comprehensive creative tests and let the performance data tell you what resonates.
Meta Ads MCP Server for Campaign Analytics
One of the chronic friction points in performance marketing is the tool-switching overhead: you're analyzing campaign data in Meta Ads Manager, you want to understand something, you need to export data and analyze it somewhere else, then switch back. Each switch breaks the analytical flow.
The team built a custom MCP server integrated with the Meta Ads API that lets them query campaign performance, spending, and ad effectiveness directly in Claude Desktop — without switching platforms. They can ask conversational questions about their campaigns and get immediate answers pulled from live data.
This kind of integration would previously have required engineering resources to build and maintain. Claude Code made it buildable by a non-technical marketing team member with no prior MCP development experience.
Advanced Prompt Engineering With Memory Systems
Beyond individual automations, the team has built something more sophisticated: a self-improving testing framework for ad creative.
The system logs hypotheses and experiment results across ad iterations. When generating new ad variations, it pulls previous test results into context — so the new generation is informed by what has and hasn't worked before. Over time, the system builds an increasingly refined model of what drives performance in their specific campaigns and audiences.
This is prompt engineering taken seriously: instead of static prompts that generate generic ad copy, the team has built a memory system that makes the generation process progressively smarter with each iteration.
Team Impact
- Ad copy creation time reduced from 2 hours to 15 minutes — an 87% time reduction for a core recurring task
- 10x increase in creative output through automated Figma plugin ad variation generation
- Operating with the output capacity of a larger team — previously would have required dedicated engineering resources for these automations
- More time for strategy and campaign planning, less time on manual execution
- Campaign analytics directly accessible in Claude Desktop via custom MCP server, eliminating platform-switching overhead
Top Tips From the Growth Marketing Team
Identify API-enabled repetitive tasks first. The highest-value Claude Code applications in marketing are workflows that currently involve repetitive actions against tools that have APIs — ad platforms, analytics tools, CRM systems, design tools. If you're doing something manually that has an API behind it, that's a strong signal that Claude Code can automate it. Make a list of your most repetitive tasks and check which tools offer API access.
Break complex workflows into specialized sub-agents. When building automation pipelines, resist the urge to make a single agent responsible for everything. The Google Ads workflow works better with a headline agent and a description agent because the constraints and creative goals are different. Specialization produces better output. Think about where in your workflow the requirements diverge and create separate agents for each branch.
Thoroughly brainstorm and prompt plan before you start coding. The team's recommended approach: spend significant time in Claude.ai first, working through the design of what you're building — the inputs, the outputs, the logic, the edge cases. Then have Claude.ai distill that planning conversation into a comprehensive prompt you can take into Claude Code to start the implementation. Starting with clear requirements saves enormous time compared to figuring things out mid-implementation.