Anthropic has a one-person Growth Marketing team that built an agentic workflow for ad copy with two specialized sub-agents (p. 15), according to a PDF report by Anthropic, "How Anthropic teams use Claude Code", created on June 3, 2025.
TL;DR: The Growth Marketing team at Anthropic is one non-technical person (p. 15) who covers paid search, paid social, app stores, email and SEO. The report says that this person built a workflow that reads a CSV file of existing ads with their performance data, finds the weak ads, and writes new versions within strict length limits. One sub-agent writes headlines and one writes descriptions. The report also describes a Figma plugin for ad image variations and a server that pulls ads data into a chat app. The report states that ad copy creation went from 2 hours to 15 minutes and that creative output rose 10x (p. 16). The report gives no method for either figure.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | PDF report, 23 pages (the Growth Marketing section is on pages 15-16) |
| Source published | 2025-06-03 (PDF creation date) |
| Speakers | No named author. The report is based on interviews with Anthropic staff |
| Industry | AI research and products |
| Function | Marketing |
| Techniques | Sub-agents, batch generation from CSV, a memory log of past tests, an MCP server for ads data |
| Tools named | Claude Code, Claude.ai, Claude Desktop, Google Ads, Figma, Meta Ads API, MCP |
Independent summary of public material. Anthropic is not affiliated with Taskade.
The system Anthropic built
The report describes the Growth Marketing team as one person (p. 15). That person runs performance marketing across paid search, paid social, mobile app stores, email marketing and SEO. The report says that the team is non-technical and uses an AI coding tool to automate repeated tasks and to build agentic workflows that traditionally need significant engineering resources.
The report lists these tools that the team built (p. 15):
- Ad copy generation. A workflow reads CSV files that hold hundreds of existing ads with performance metrics (p. 15). It finds the ads that perform badly and writes new versions for them.
- A Figma plugin. The plugin finds the frames in a design file and makes up to 100 ad variations by swapping headlines and descriptions (p. 15).
- An ads data server. An MCP server connects to the Meta Ads API, so that the team can ask about campaign performance, spend and ad results inside a chat app.
- A memory log. The system logs each hypothesis and test across ad versions, so that it can pull earlier results into context when it writes new versions.
Architecture of the system
The ad copy workflow splits the work between two specialized sub-agents (p. 15). One writes headlines, and one writes descriptions. Each output must meet the length limits of the ad platform: 30 characters for a headline and 90 characters for a description (p. 15). With this split, the report says, the system can write hundreds of new ads in minutes (p. 15).
The report gives three tips from the team (p. 16):
- Look for repeated tasks in tools that have an API, such as ad platforms, design tools and analytics platforms.
- Split a complex workflow into separate sub-agents for separate tasks, such as a headline agent and a description agent. The report says that this split makes debugging easier and gives better output for complex requirements.
- Plan the whole workflow in a chat assistant first, have it write a full prompt and code structure, and then build step by step instead of in one request.
Results the source reports
- The report states that ad copy creation went from 2 hours to 15 minutes (p. 16).
- The report states a 10x increase in creative output, from automated generation and the Figma plugin (p. 16).
- The report states that the Figma plugin cut hours of copy and paste to half a second for each batch (p. 15).
- The report states that the one-person team now does tasks that once needed dedicated engineering resources (p. 16).
Critical assessment
The source is a vendor report about its own product, written from interviews with its own staff. The report does not say how the team measured the 2-hour and 15-minute times (p. 16), how many ads went into the 10x figure, or whether the new ads performed better than the old ones. It gives no ad performance results at all. The text layer of the scanned PDF repeats words, so a text extract reads "22 hours" where the printed page says "2 hours" (p. 16). The figures here come from the page images. No outside party checked the numbers. The split into a headline sub-agent and a description sub-agent, each with a hard length limit, is a clear and reusable pattern.
Build this in Taskade
- Build a team of AI agents with multi-agent collaboration, with one agent for headlines and one for descriptions. Start at AI agents.
- Start from a template in digital advertising agents.
- Run the steps around the agents, such as a weekly copy refresh, with marketing automations.
- Read ad rows from a sheet and write new versions back with the Google Sheets integration, one of 100+ integrations.
- Connect an outside MCP server through an automation with the MCP Client connector. The connector works on every plan.