Below I’ll explain what they built, why it matters, and how you can copy the patterns without a big team (with only Claude Code). Who is they you are asking? Well, Anthrophic! Read their full document here.
What they built and why it matters
They focused on four practical automations: automatic Google Ads creative generation, a Figma plugin for mass creative production, a Meta Ads MCP (Model Context Protocol) server for campaign pull, and a simple memory system - I personally recommend basic-memory for such use cases. Each automation handles a specific pain point.
For Google Ads, the heavy lift is generating many compliant variations, then testing them. Ads have strict character limits, and manual creation is slow and error prone. The Claude Code workflow reads a CSV of past ads and metrics, flags underperformers, and spins up hundreds of new headline and description variations. Instead of a human spending hours or days, the system churns out candidates in minutes.
For creative output on social, they automated what designers do manually. A Figma plugin finds frames and swaps text layers to produce batches of variations. That reduced a job that used to take hours of copy-pasting down to less than a second per batch.
For measurement, they built a small MCP server that talks to the Meta Ads API and serves campaign performance directly into Claude Desktop. No more tab switching between platforms to pull numbers and form hypotheses.
Finally, they added a lightweight memory system. It logs hypotheses, experiment parameters, and results so each new round of creative generation can reference past learnings. That makes the testing pipeline self-improving rather than starting from scratch every time.
These are small, repeatable automations that buy hours per week and let a tiny team operate like a larger one.
How the Google Ads automation works (practical details)
They start with a CSV export that includes ad text and performance columns. Think columns like ad_id, headline, description, impressions, clicks, ctr, conversions, cost_per_conversion. The workflow first filters for ads that meet a "needs iteration" rule, for example CTR below a threshold with reasonable impressions. That keeps the system focused on ads that matter.
Claude Code runs two focused sub-agents. One handles headlines and enforces the 30 character limit. The other handles descriptions and enforces the 90 character limit. Splitting the task makes each agent simpler to prompt and easier to debug. Each sub-agent is asked to produce N variants and to return both the copy and metadata like estimated tone and targeted audience hints. The workflow then writes the new variations back to a CSV, tags them for a staged rollout, and hands them off to the ad platform or human QA.
Why split agents? Prompting one model to manage all constraints tends to create edge-case failures. Two small, single-responsibility agents give predictable outputs and make it simple to add more constraints later, like legal checks or localization.
The Figma plugin for mass creative production
Designers hate repetitive resizing and swapping. The plugin they wrote scans a Figma page for frames that match an ad template. It finds the text layers that contain headline and description, then programmatically replaces those with the generated copy. It can output up to 100 variations in one go. The plugin also exports image assets ready for upload to ad platforms.
This approach does two useful things. First, it preserves layout and image composition, so designers don't lose control of the visual. Second, it scales testing by letting you attach many copy variants to a small set of visual templates. You end up testing more combinations without hiring more designers.
A few practical tips here: keep text layers named consistently across templates, and bake in a fallback font size so long headlines truncate predictably. Also add a quick visual QA step. Even automated swaps can create weird line breaks or overlap.
Meta Ads MCP server and measurement flow
The MCP server is small and focused. It pulls campaign-level metrics, spend data, and ad creative performance from the Meta Ads API. The server normalizes the data and serves endpoints Claude Desktop can query. That means the marketing workflow can ask for the latest stats, form hypotheses, and generate new creatives without switching tools.
Rate limits and data freshness matter. They implemented simple caching and incremental pulls so the server doesn't hammer the API. They also added a lightweight schema that ties creative variants to experiment IDs so test results are easy to attribute back to individual variations.
Why bother? Pulling data into the same environment where you generate copy closes the loop fast. You look at a bad ad, ask Claude to rewrite it using the last two months of learnings, and get candidate copy within the same session. That sort of tight loop speeds up meaningful iterations.
The memory system: keep experiments from getting lost
Tracking hypotheses used to be a spreadsheet nightmare for them. They built a small memory layer that logs experiments, the hypothesis, the variants used, rollout dates, and results. When generating new creative, Claude can pull past experiment summaries and avoid repeating failed approaches. It can also suggest promising directions based on winning motifs from similar audiences.
This is not a full-featured experiment platform. It’s a simple registry that prevents reinventing the wheel. Even a few structured fields and a consistent naming convention for experiments can save hours of duplicated effort and make A/B test attribution sane.
Real-world impact and trade-offs
The concrete results were simple. Ad copy creation time dropped from roughly two hours to about 15 minutes for a campaign refresh. Creative output increased by about 10x, meaning many more variants get tested. The single-person team started running workflows that used to need engineering support.
But there are trade-offs. Automated copy still needs human QA for brand voice and legal compliance. Models can hallucinate metrics or fabricate claims, so you need checks that prevent false statements from going live. API rate limits and costs are practical constraints. And automated variants are only useful if your experiment and attribution setup is solid, otherwise you’ll collect noise instead of signal.
Practical how-to and a short starter checklist
If you want to try this in your team, start small and focus on one bottleneck. For most teams that will be either creative generation or measurement.
Quick starter checklist:
- Pick one ad platform and export a CSV with ad text and at least three performance columns.
- Build two simple Claude Code prompts: one for headlines, one for descriptions, each enforcing the exact character limits. Keep prompts short and strict.
- Add a human QA step before anything goes live. Automate everything else once you trust the outputs.
After that, add a Figma export/import step if you want batch creative rendering. Then bring in a tiny MCP server for the platform you care about and connect it to the same workflow.
Common pitfalls and how to avoid them
People tend to ask Claude for one-shot, complicated magic prompts. That fails. Break the job into small tasks. Validate outputs at each step. Log every experiment with IDs so you can map results back to variants. Beware overfitting to short-term metrics, and make sure you track the right goals, not just CTR.
Also watch out for rate limits and edge-case text behavior. Long headlines that barely fit may break layouts on mobile. Automated QA should check character counts, line breaks, and legal tokens before upload.
Final thoughts
The pattern here is simple: automate the repetitive, make the outputs predictable, and keep humans in the loop for judgment. Claude Code isn't a all magic, but it lowers the engineering bar so small marketing teams can move fast.
If you keep prompts focused, split responsibilities across small agents, and make sure you can tie experiments back to results, you'll get a lot of mileage from this approach.
P.S. I suggest always starting with this prompt: "Do not waste tokens on generating implementation guides and illustrations, focus on the producing quality codes, I am satisfied with only minimal explanation."
Top comments (0)