The team that builds Claude runs its own marketing on Claude Code.
That detail matters, because it means the people closest to the model are using it to do the unglamorous, repetitive work that usually eats a marketing team alive: writing ad copy, spinning up creative variations, pulling campaign numbers, and tracking what worked.
Here are the four workflows Anthropic's marketing team runs, what each one changed, and why this is only the surface of what a modern GTM team can build.
1. Automated Ad Creative Generation
Anthropic's marketers built an agentic workflow in Claude Code that reads CSV files holding hundreds of existing ads alongside their performance data.
The workflow does three things on its own. It scans the spreadsheet, flags the creatives that underperform, and generates fresh variations that respect the character limits each ad platform enforces. A marketer no longer has to open every row, compare metrics by hand, and rewrite the losers one at a time.
The detail that makes it work is the performance data sitting next to the copy. The agent is not guessing which lines are weak from tone alone. It reads the numbers, ties them to specific creatives, and only rewrites what the data says is dragging the account. That grounding is what turns a generic copy generator into something a paid team can trust.
The result: ad copy creation went from 2 hours to 15 minutes.
That is the same kind of problem every outbound and paid team faces. You have a pile of copy, you know some of it is weak, and rewriting it against strict length rules is slow, fiddly work. The value is not that a model can write a headline. It is that the model can hold your character limits, your performance history, and your whole ad set in context at once, which no human does comfortably at scale.
If you want to pressure-test your own short-form copy against those constraints, you can score it and get rewrites here, for free:
Email Optimizer Tool
Business email required to use the tool
2. Mass Creative Production in Figma
The second workflow moves the same idea into design. Anthropic's team built a Figma plugin driven by Claude that identifies the frames inside their ad designs.
From there it programmatically generates up to 100x ad variations, swapping headlines and descriptions automatically across every frame. The manual version of this job, copy-pasting new lines into each layout, used to take hours. The plugin now runs a full batch in about half a second.
What makes this more than a batch-export trick is that design and copy stop being separate steps. The layout, the headline, and the description move together, so the team ends up with finished, on-brand assets rather than a text file someone still has to drop into Figma by hand. The handoff that usually sits between a copywriter and a designer disappears.
The result: a 10x increase in creative output across channels.
Volume like that changes the constraint. When you can produce a hundred variations in seconds, the bottleneck stops being production and becomes ideas. The hard question turns into which angles, hooks, and offers are worth testing in the first place. A team that can only ship five ads a week rations its ideas. A team that can ship a hundred has to go find more of them.
That is a targeting and strategy problem, not a design one. You can generate campaign angles built around your ICP and positioning in seconds, for free:
Campaign Ideation Tool
Enter your email to generate your campaign ideas.
3. Real-Time Campaign Analytics With MCP
The third workflow connects Claude to live campaign data. Anthropic's team built an MCP server plugged into the Meta Ads API.
Through it, they query campaign performance in real time, analyze spending data, and track ad effectiveness, all directly inside Claude Desktop. Instead of exporting reports and rebuilding pivot tables, a marketer asks a question in plain language and gets the answer against fresh numbers.
This closes the loop with the first two workflows. The copy agent writes the ads, the Figma plugin produces the variations, and the MCP connection tells the team which of those variations are working while they are live. Generation and measurement stop being separate jobs owned by separate people on separate schedules.
The result: efficiency gains that translate directly to better ROI.
The wider lesson sits in the plumbing. MCP is the layer that lets a model reach into the tools a team already runs, whether that is an ad platform, a CRM, or a billing system. Once the data is reachable, analysis stops being a weekly ritual and becomes something you do mid-conversation. A question like "which creatives spent the most and returned the least this week" gets answered in the same window where you brief the next batch.
4. A Self-Improving Testing Framework
The fourth workflow is the one that compounds. Anthropic's team implemented a memory system inside Claude Code that logs hypotheses and experiments across ad iterations.
When the team generates new variations, the system pulls previous test results back into context. Every new round of creative is informed by what already won and lost, which creates a self-improving feedback loop rather than a series of disconnected tests.
Without a memory layer, testing at high volume works against you. Ship a hundred ads a week and the learnings scatter across dashboards, docs, and people's heads, and by the next sprint half of them are forgotten. The framework captures the hypothesis behind each test and the outcome it produced, so the account accumulates knowledge instead of just accumulating spend.
The result: systematic experimentation that would be impossible to track by hand.
This is the piece that separates a clever prompt from a real system. Character limits and variation counts are useful on day one. A memory of what your audience responded to last month is what keeps the output improving on day ninety. The framework remembers so the marketer does not have to.
5. What GTM and Marketing Teams Should Take From It
Read the four workflows together and one theme runs through all of them: tasks that used to require engineers are now handled by marketers.
Writing an ad-scoring agent, building a Figma plugin, standing up an MCP server, wiring a memory system. In the old model, each of those is a ticket in an engineering backlog. In Anthropic's setup, the marketing team builds them directly and ships more strategy while spending less time on manual execution.
None of these workflows is exotic. A copy generator that respects platform rules, a variation engine, a live analytics connection, and a memory of past tests. Any team running paid or outbound at volume has the raw material to build the same four. The gap is usually not capability, it is knowing that a marketer can now build the thing at all.
Before you can improve a channel, you need an honest read on where it stands. You can benchmark your current GTM setup and see where the biggest gaps sit, for free:
GTM Report Tool
We have lived this shift on our side. Running Claude Code across 70+ B2B clients, we multiplied our own output more than 4x by handing repetitive build-and-analyze work to the model instead of to headcount.
The four workflows above are worth copying, but treat them as a starting point rather than a ceiling. Anthropic applied Claude Code to ad creative and campaign analytics. The same pattern, an agent that reads your data, acts within your constraints, and remembers what worked, maps onto list building, personalization, reporting, and most of the manual work sitting in a GTM team's week. This is the surface. What you build under it is up to you.
---
