ANTHROPIC JUST RELEASED INTERNAL DOCUMENT: HOW THEIR OWN TEAM USES AI FOR WORK
Anthropic (the company that created Claude, ChatGPT's main competitor) just published an internal document sharing how 10 teams in the company use Claude Code for daily work.
I read all 22 pages. There's one case study that made me stop and think:
THE MARKETING TEAM HAS ONLY 1 PERSON
Anthropic's Growth Marketing team has exactly 1 person. Doesn't know code. Responsible for Google Ads, Facebook Ads, email marketing, SEO, app store.
Previously, creating Google Ads content took 2 hours each time. Had to write each headline (30 character limit), each description (90 character limit), for hundreds of different campaigns.
Now this person uses Claude Code to:
- Read CSV file containing hundreds of old ads with performance metrics
- Automatically analyze which ads are underperforming
- Automatically generate hundreds of new variants, within character limits
=> From 2 hours to 15 minutes. Speed increased 10 times.
NOT STOPPING THERE
This person also created a Figma plugin to generate 100 ad image variants in half a second. Previously had to copy-paste manually, taking hours.
Then created an MCP server connected to Meta Ads API, to analyze Facebook Ads campaign performance right in Claude without switching between platforms.
Reminder: this person doesn't know code 😬.
NOT JUST MARKETING
The document also shares case studies from 9 other teams:
- Legal Team: Lawyer who doesn't know code built an app to support communication for family members with illness, in 1 hour
- Design Team: Designer implemented interfaces, fixed code themselves instead of waiting for programmers. Speed 2-3 times faster
- Finance Team: Finance staff wrote requests in plain English, AI ran data queries and output Excel
- Security Team: Debugged infrastructure issues from 15 minutes to 5 minutes
- Data Science Team: Wrote a 5,000-line TypeScript app despite "almost not knowing JavaScript"
LESSONS LEARNED
Common points among all 10 teams:
-
No one uses AI in a "ask 1 question, get 1 answer" way. They create workflows to automate entire processes
-
Claude.md file (file guiding AI on work context) is the most important thing. Teams with detailed Claude.md use AI much more effectively
-
AI gets it right on first try only about 1/3 of the time. But total time is still saved a lot because when right, it's much faster
-
Don't try to fix errors when AI is wrong. They save the current version first, let AI try, then check. If wrong, restore old version, let AI try again from start. Faster than editing AI-generated results.
=> The boundary between "knows code" and "doesn't know code" is blurring very quickly. The issue is no longer whether you know programming, but whether you can describe accurately what you need.
PS: Original PDF file and translation in comments below
Link: https://www-cdn.anthropic.com/58284b19e702b49db9302d5b6f135ad8871e7658.pdf
