F*CK, HIS SETUP WILL NEVER GET OLD
BORIS CHERNY BUILT CLAUDE CODE AT ANTHROPIC, AND THIS IS HIS BIBLE FOR WORKING WITH CLAUDE
15 Claudes at once. 0 secret prompts. 1 rule that makes results 2 to 3x better
his own words: "my setup might be surprisingly vanilla"
here is what he runs, in plain words:
→ 5 sessions in his terminal + 5 to 10 more in the browser. some he starts from his phone every morning
→ plan mode first: Claude writes the plan, he approves it, then it builds
→ 1 shared Claude.md, a rules file for the whole team. every time Claude gets something wrong, they add a line so it does not happen again
→ /commit-push-pr, 1 shortcut command he uses "dozens of times every day"
→ 2 helper agents: 1 cleans up the code, 1 tests the whole app end to end
→ an automatic step that formats every file Claude touches
→ plugs into the boring tools: Slack, BigQuery and Sentry error logs
the 1 rule he says matters most:
give Claude a way to check its own work. he says that alone makes the result 2 to 3x better
everyone is hunting for the perfect prompt. the guy who built it is adding checks
checks, not tricks
save this, then give your Claude 1 way to check itself today
Holy sh*t. Swap an AI agent's hidden thoughts with another problem's and its math score drops from 84.6 to 44.8
Here's how the test works:
→ take an agent that thinks in hidden vectors
→ swap its thoughts with ones from a different problem
→ score the answer on MATH500
→ compare with its own thoughts: 84.6 vs 44.8
Noise twice as strong as the thought leaves 66.4
The meaning matters more than the noise
Stanford and Together AI ran it at two model sizes
At 14B it goes from 88.2 to 48.6
Steal the test for your own agent
Take the plan from task A, paste it into task B and see how much worse B gets
It tells you if your agent really uses its plan